From 1c21d1d3f4ea98d7a30f5ce16c2337d7d169da12 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 22 May 2026 09:37:53 +0000
Subject: [PATCH 01/13] Studio: surface every live provider model in the picker

Two regressions in the provider registry hide live models from the
chat picker:

1. The Anthropic denylist `-\d{8}$` strips every dated id. Per the
   models overview, the pre-4.6 generation ships only as dated ids
   (Opus 4.5, Sonnet 4.5, Haiku 4.5, Opus 4.1, the deprecated 4.0
   family). The denylist hides 6 of the 9 live models -- a user
   wanting Haiku 4.5 or Sonnet 4.5 can't pick them from the dropdown.

2. The OpenAI allowlist `^(gpt-5\.[345]|gpt-4\.5|o3)(?:[-.]|$)`
   silently drops every family OpenAI ships outside the hardcoded
   set. Existing chat models like gpt-3.5-turbo, gpt-4, gpt-4o,
   gpt-4.1, gpt-5, gpt-5.1, gpt-5.2, gpt-5.3, o1, o4-mini and
   anything OpenAI launches next get dropped on the floor.

Changes:

- Drop the Anthropic denylist entirely; show every live id.
- Update Anthropic default_models seed to use the canonical dated
  ids so the pre-load list matches what /v1/models returns.
- Replace the OpenAI allowlist with a non-chat denylist that drops
  only known non-chat model families: embeddings, TTS, image,
  moderation, whisper, audio, realtime, transcribe, search-preview,
  sora video, computer-use harness, legacy bases (babbage, davinci,
  ada, curie), fine-tunes (ft:*), and dated snapshots. Every chat
  family auto-surfaces the moment OpenAI lists it.
- Add `test_provider_registry_filters.py` with parametrized cases
  covering current + hypothetical-future chat ids that must survive,
  plus the full non-chat surface area that must be dropped.

Live verified against the real /v1/models on both providers:
- Anthropic: 9 -> 9 (was 9 -> 3).
- OpenAI: 129 -> 46 chat models (was 129 -> 12).
---
 studio/backend/core/inference/providers.py    |  45 +++++---
 .../tests/test_provider_registry_filters.py   | 108 ++++++++++++++++++
 2 files changed, 136 insertions(+), 17 deletions(-)
 create mode 100644 studio/backend/tests/test_provider_registry_filters.py

diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index fef9ba3e12..31112f2061 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -26,14 +26,24 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
         "supports_tool_calling": True,
         "auth_header": "Authorization",
         "auth_prefix": "Bearer ",
-        # Keep the model picker scoped to the current generation. The remote
-        # /v1/models listing returns dozens of historical snapshots, fine-tunes
-        # and non-chat models (embeddings, TTS, image, moderation) that we
-        # never want to surface in the chat UI. Filtering here so backend
-        # is the single source of truth.
-        "model_id_allowlist": re.compile(r"^(gpt-5\.[345]|gpt-4\.5|o3)(?:[-.]|$)"),
-        # Hide dated snapshots and the retired plain gpt-5.3 id.
-        "model_id_denylist": re.compile(r"^(gpt-5\.3)$|-\d{4}-\d{2}-\d{2}$"),
+        # The remote /v1/models listing returns the full account catalog,
+        # including non-chat models (embeddings, TTS, image, moderation,
+        # whisper, dall-e) and fine-tunes. Previously we used a hardcoded
+        # family allowlist (`gpt-5\.[345]|gpt-4\.5|o3`) which silently
+        # dropped every new family OpenAI shipped, including the gpt-5.5
+        # generation today. Switch to a non-chat denylist instead so any
+        # new chat family auto-appears the moment OpenAI lists it. The
+        # patterns below cover every non-chat model id OpenAI has ever
+        # published; chat ids never start with these prefixes.
+        "model_id_denylist": re.compile(
+            r"(?:^|-)(?:embedding|tts|whisper|moderation|image|search|audio|"
+            r"realtime|transcribe|babbage|davinci|ada|curie|sora)\b"
+            r"|^dall-e\b"
+            r"|^computer-use\b"
+            r"|^text-(?:embedding|moderation|davinci|curie|babbage|ada)\b"
+            r"|^ft:"
+            r"|-\d{4}-\d{2}-\d{2}$"
+        ),
     },
     "anthropic": {
         "display_name": "Anthropic",
@@ -42,16 +52,17 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
             "claude-opus-4-7",
             "claude-opus-4-6",
             "claude-sonnet-4-6",
-            "claude-opus-4-5",
-            "claude-sonnet-4-5",
-            "claude-haiku-4-5",
+            "claude-opus-4-5-20251101",
+            "claude-sonnet-4-5-20250929",
+            "claude-haiku-4-5-20251001",
         ],
-        # Anthropic /v1/models returns dated snapshot ids alongside the
-        # canonical names (e.g. claude-3-5-sonnet-20241022). Hide the
-        # YYYYMMDD-suffixed variants from the picker — same intent as the
-        # OpenAI denylist, just a different date format (no dashes between
-        # year/month/day).
-        "model_id_denylist": re.compile(r"-\d{8}$"),
+        # Anthropic's /v1/models returns dated ids for every model in the
+        # pre-4.6 generation -- `claude-opus-4-5-20251101`,
+        # `claude-haiku-4-5-20251001`, `claude-sonnet-4-5-20250929`,
+        # `claude-opus-4-1-20250805`. Per the models overview, those
+        # dated ids ARE the canonical names for that generation, not
+        # snapshots to hide. Dropping the previous `-\d{8}$` denylist
+        # so every live model the API returns reaches the picker.
         "supports_streaming": True,
         "supports_vision": True,
         "supports_tool_calling": False,
diff --git a/studio/backend/tests/test_provider_registry_filters.py b/studio/backend/tests/test_provider_registry_filters.py
new file mode 100644
index 0000000000..d709c4ebba
--- /dev/null
+++ b/studio/backend/tests/test_provider_registry_filters.py
@@ -0,0 +1,108 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Provider registry model-id filter regression tests.
+
+The OpenAI ``model_id_allowlist`` previously hardcoded the gpt-5.3/4/5
+families plus gpt-4.5 / o3 -- silently dropping every future family
+OpenAI shipped. Anthropic's ``model_id_denylist`` previously stripped
+every dated id, hiding the canonical names of every pre-4.6 model
+(Opus 4.5, Sonnet 4.5, Haiku 4.5, Opus 4.1, the 4.0 family).
+
+These tests pin the new non-chat denylist (OpenAI) and the empty
+denylist (Anthropic) by walking realistic ``/v1/models`` listings
+through ``PROVIDER_REGISTRY`` and asserting the surviving set.
+"""
+
+from core.inference.providers import PROVIDER_REGISTRY
+
+
+def _apply(provider_type: str, candidate_ids: list[str]) -> list[str]:
+    """Mirror the filter logic in ``routes/providers.list_models``."""
+    info = PROVIDER_REGISTRY[provider_type]
+    out = list(candidate_ids)
+    allow = info.get("model_id_allowlist")
+    if allow is not None:
+        out = [m for m in out if allow.match(m)]
+    deny = info.get("model_id_denylist")
+    if deny is not None:
+        out = [m for m in out if not deny.search(m)]
+    return out
+
+
+# ── OpenAI: non-chat denylist drops only non-chat ids ──────────────
+
+
+def test_openai_keeps_every_known_chat_family():
+    live = [
+        # Current generation (must survive).
+        "gpt-5.5", "gpt-5.5-pro",
+        "gpt-5.4", "gpt-5.4-pro", "gpt-5.4-mini", "gpt-5.4-nano",
+        "gpt-5.3-codex", "gpt-5.3-chat-latest",
+        "o3", "o3-pro", "o3-mini", "o3-deep-research",
+        # Hypothetical future families that the old allowlist would have
+        # silently dropped -- they MUST surface under the new denylist.
+        "gpt-5.6", "gpt-5.6-mini", "gpt-6", "gpt-6-pro",
+        "o4", "o4-pro", "o5",
+    ]
+    surviving = _apply("openai", live)
+    assert surviving == live, surviving
+
+
+def test_openai_drops_non_chat_ids():
+    noise = [
+        # Embeddings / TTS / image / moderation / whisper / audio etc.
+        "text-embedding-3-small", "text-embedding-3-large",
+        "text-embedding-ada-002",
+        "text-moderation-latest", "text-moderation-stable",
+        "tts-1", "tts-1-hd", "gpt-4o-tts",
+        "whisper-1",
+        "dall-e-2", "dall-e-3",
+        "gpt-image-1", "gpt-image-2", "gpt-image-1-mini",
+        "chatgpt-image-latest",
+        "gpt-audio-1.5", "gpt-realtime-2", "gpt-4o-realtime-preview",
+        "gpt-4o-transcribe", "gpt-4o-search-preview",
+        "gpt-4o-mini-search-preview", "gpt-4o-mini-transcribe",
+        "gpt-4o-mini-tts",
+        "omni-moderation-latest",
+        # Video generation.
+        "sora-2", "sora-2-pro",
+        # Computer-use is an agentic harness, not a chat id.
+        "computer-use-preview",
+        # Legacy bases.
+        "babbage-002", "davinci-002", "text-davinci-003",
+        "text-curie-001", "text-ada-001",
+        # Fine-tunes.
+        "ft:gpt-4o-mini:acme:abc:xyz",
+        # Dated snapshots are still hidden.
+        "gpt-4o-2024-08-06", "gpt-4o-mini-2024-07-18",
+        "gpt-5.5-2026-04-23",
+    ]
+    surviving = _apply("openai", noise)
+    assert surviving == [], surviving
+
+
+# ── Anthropic: empty denylist; dated ids ARE canonical ───────────────
+
+
+def test_anthropic_surfaces_every_live_model_including_dated_ids():
+    # The full set of ids /v1/models returns today.
+    live = [
+        "claude-opus-4-7",
+        "claude-sonnet-4-6",
+        "claude-opus-4-6",
+        "claude-opus-4-5-20251101",
+        "claude-sonnet-4-5-20250929",
+        "claude-haiku-4-5-20251001",
+        "claude-opus-4-1-20250805",
+        "claude-opus-4-20250514",
+        "claude-sonnet-4-20250514",
+    ]
+    surviving = _apply("anthropic", live)
+    assert surviving == live, surviving
+
+
+def test_anthropic_default_models_match_filter():
+    info = PROVIDER_REGISTRY["anthropic"]
+    surviving = _apply("anthropic", list(info["default_models"]))
+    assert surviving == list(info["default_models"]), surviving

From 4434e1eeb7026e1ac130576ae961cf703763b3d2 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Fri, 22 May 2026 09:38:31 +0000
Subject: [PATCH 02/13] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 .../tests/test_provider_registry_filters.py   | 65 ++++++++++++++-----
 1 file changed, 47 insertions(+), 18 deletions(-)

diff --git a/studio/backend/tests/test_provider_registry_filters.py b/studio/backend/tests/test_provider_registry_filters.py
index d709c4ebba..66c90027e2 100644
--- a/studio/backend/tests/test_provider_registry_filters.py
+++ b/studio/backend/tests/test_provider_registry_filters.py
@@ -36,14 +36,27 @@ def _apply(provider_type: str, candidate_ids: list[str]) -> list[str]:
 def test_openai_keeps_every_known_chat_family():
     live = [
         # Current generation (must survive).
-        "gpt-5.5", "gpt-5.5-pro",
-        "gpt-5.4", "gpt-5.4-pro", "gpt-5.4-mini", "gpt-5.4-nano",
-        "gpt-5.3-codex", "gpt-5.3-chat-latest",
-        "o3", "o3-pro", "o3-mini", "o3-deep-research",
+        "gpt-5.5",
+        "gpt-5.5-pro",
+        "gpt-5.4",
+        "gpt-5.4-pro",
+        "gpt-5.4-mini",
+        "gpt-5.4-nano",
+        "gpt-5.3-codex",
+        "gpt-5.3-chat-latest",
+        "o3",
+        "o3-pro",
+        "o3-mini",
+        "o3-deep-research",
         # Hypothetical future families that the old allowlist would have
         # silently dropped -- they MUST surface under the new denylist.
-        "gpt-5.6", "gpt-5.6-mini", "gpt-6", "gpt-6-pro",
-        "o4", "o4-pro", "o5",
+        "gpt-5.6",
+        "gpt-5.6-mini",
+        "gpt-6",
+        "gpt-6-pro",
+        "o4",
+        "o4-pro",
+        "o5",
     ]
     surviving = _apply("openai", live)
     assert surviving == live, surviving
@@ -52,30 +65,46 @@ def test_openai_keeps_every_known_chat_family():
 def test_openai_drops_non_chat_ids():
     noise = [
         # Embeddings / TTS / image / moderation / whisper / audio etc.
-        "text-embedding-3-small", "text-embedding-3-large",
+        "text-embedding-3-small",
+        "text-embedding-3-large",
         "text-embedding-ada-002",
-        "text-moderation-latest", "text-moderation-stable",
-        "tts-1", "tts-1-hd", "gpt-4o-tts",
+        "text-moderation-latest",
+        "text-moderation-stable",
+        "tts-1",
+        "tts-1-hd",
+        "gpt-4o-tts",
         "whisper-1",
-        "dall-e-2", "dall-e-3",
-        "gpt-image-1", "gpt-image-2", "gpt-image-1-mini",
+        "dall-e-2",
+        "dall-e-3",
+        "gpt-image-1",
+        "gpt-image-2",
+        "gpt-image-1-mini",
         "chatgpt-image-latest",
-        "gpt-audio-1.5", "gpt-realtime-2", "gpt-4o-realtime-preview",
-        "gpt-4o-transcribe", "gpt-4o-search-preview",
-        "gpt-4o-mini-search-preview", "gpt-4o-mini-transcribe",
+        "gpt-audio-1.5",
+        "gpt-realtime-2",
+        "gpt-4o-realtime-preview",
+        "gpt-4o-transcribe",
+        "gpt-4o-search-preview",
+        "gpt-4o-mini-search-preview",
+        "gpt-4o-mini-transcribe",
         "gpt-4o-mini-tts",
         "omni-moderation-latest",
         # Video generation.
-        "sora-2", "sora-2-pro",
+        "sora-2",
+        "sora-2-pro",
         # Computer-use is an agentic harness, not a chat id.
         "computer-use-preview",
         # Legacy bases.
-        "babbage-002", "davinci-002", "text-davinci-003",
-        "text-curie-001", "text-ada-001",
+        "babbage-002",
+        "davinci-002",
+        "text-davinci-003",
+        "text-curie-001",
+        "text-ada-001",
         # Fine-tunes.
         "ft:gpt-4o-mini:acme:abc:xyz",
         # Dated snapshots are still hidden.
-        "gpt-4o-2024-08-06", "gpt-4o-mini-2024-07-18",
+        "gpt-4o-2024-08-06",
+        "gpt-4o-mini-2024-07-18",
         "gpt-5.5-2026-04-23",
     ]
     surviving = _apply("openai", noise)

From 9a9c6b09496e854655fb3f386523687a5817d9b6 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 22 May 2026 11:06:17 +0000
Subject: [PATCH 03/13] Address review: tighten denylist + document mid-id
 matching intent

Two tweaks to the OpenAI model_id_denylist after review:

1. `(?:^|-)search\b` was a false positive: it dropped
   `gpt-4o-search-preview` and `gpt-4o-mini-search-preview`, both of
   which are chat-with-retrieval models that respond to standard
   /v1/chat/completions and should appear in the picker. Replaced
   with a `-search-api(?:-\d{4}-\d{2}-\d{2})?$` suffix match that
   only catches the standalone search-API endpoint
   (`gpt-5-search-api`, `gpt-5-search-api-2025-10-14`).

2. The legacy completion bases (babbage / davinci / ada / curie) had
   `(?:^|-)` anchoring, which would mid-id-match a hypothetical
   future chat id like `gpt-7-davinci-edition`. Tightened to
   `^(?:babbage|davinci|ada|curie)\b` so only the legacy ids
   (`babbage-002`, `davinci-002`, `text-davinci-003`, ...) are
   dropped while leaving any future chat id containing the word
   alone.

Expanded the regex docstring to spell out the intent per pattern
group (feature-suffix matches are intentional mid-id; legacy bases
are ^-anchored; standalone families are prefix-only) so a future
maintainer doesn't undo it.

Tests: added test_openai_search_preview_is_kept_search_api_is_dropped
to pin the chat-vs-API distinction, and
test_openai_legacy_completion_names_only_match_at_id_start to lock
in the mid-id safety on the legacy-base set.
---
 studio/backend/core/inference/providers.py    | 47 +++++++++++--
 .../tests/test_provider_registry_filters.py   | 66 ++++++++++++++++++-
 2 files changed, 105 insertions(+), 8 deletions(-)

diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index 31112f2061..bebebdc396 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -32,16 +32,51 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
         # family allowlist (`gpt-5\.[345]|gpt-4\.5|o3`) which silently
         # dropped every new family OpenAI shipped, including the gpt-5.5
         # generation today. Switch to a non-chat denylist instead so any
-        # new chat family auto-appears the moment OpenAI lists it. The
-        # patterns below cover every non-chat model id OpenAI has ever
-        # published; chat ids never start with these prefixes.
+        # new chat family auto-appears the moment OpenAI lists it.
+        #
+        # Pattern strategy:
+        # - Unambiguous *feature* indicators (tts/whisper/transcribe/audio/
+        #   realtime/image/embedding/moderation/sora) match anywhere in
+        #   the id with `(?:^|-)` because OpenAI uses them as the
+        #   primary qualifier on every variant -- e.g. canonical
+        #   `tts-1` AND family variants `gpt-4o-tts`, `gpt-4o-mini-tts`.
+        #   These words don't appear in legitimate chat model ids
+        #   mid-string, so the mid-id match is intentional and safe.
+        # - `search` is INTENTIONALLY NOT in the mid-id set because
+        #   `gpt-4o-search-preview` and `gpt-4o-mini-search-preview` are
+        #   chat-with-retrieval models that absolutely belong in the
+        #   picker. The standalone search API is caught separately by
+        #   `-search-api` (suffix) which never appears on a chat id.
+        # - Legacy completion bases (babbage / davinci / ada / curie)
+        #   are ^-anchored: they ONLY ever begin a legacy id like
+        #   `babbage-002`, `davinci-002`, `text-davinci-003`. A future
+        #   hypothetical `gpt-7-davinci-edition` chat model would NOT
+        #   be dropped, which is the right default for a denylist.
+        # - Prefix-only patterns (`^dall-e`, `^computer-use`, `^ft:`,
+        #   `^gpt-image`) cover canonical non-chat families without
+        #   risk of mid-id false positives.
+        #
+        # Verified against the live /v1/models listing 2026-05-22.
         "model_id_denylist": re.compile(
-            r"(?:^|-)(?:embedding|tts|whisper|moderation|image|search|audio|"
-            r"realtime|transcribe|babbage|davinci|ada|curie|sora)\b"
+            # Feature suffixes that mark a non-chat variant on any base.
+            r"(?:^|-)(?:embedding|tts|whisper|moderation|image|audio|"
+            r"realtime|transcribe|sora)\b"
+            # Legacy completion bases -- ^-anchored to avoid false
+            # positives on hypothetical future chat ids containing
+            # those words mid-string.
+            r"|^(?:babbage|davinci|ada|curie)\b"
+            r"|^text-(?:embedding|moderation|davinci|curie|babbage|ada)\b"
+            # Standalone search API (separate endpoint shape).
+            # `gpt-4o-search-preview` (chat-with-search) is intentionally
+            # NOT matched here.
+            r"|-search-api(?:-\d{4}-\d{2}-\d{2})?$"
+            # Canonical non-chat prefixes.
             r"|^dall-e\b"
             r"|^computer-use\b"
-            r"|^text-(?:embedding|moderation|davinci|curie|babbage|ada)\b"
+            # Fine-tunes carry the user's tenant in the id.
             r"|^ft:"
+            # Dated snapshot suffixes hide behind the canonical id
+            # which is also in the listing.
             r"|-\d{4}-\d{2}-\d{2}$"
         ),
     },
diff --git a/studio/backend/tests/test_provider_registry_filters.py b/studio/backend/tests/test_provider_registry_filters.py
index 66c90027e2..8c20d12173 100644
--- a/studio/backend/tests/test_provider_registry_filters.py
+++ b/studio/backend/tests/test_provider_registry_filters.py
@@ -48,6 +48,14 @@ def test_openai_keeps_every_known_chat_family():
         "o3-pro",
         "o3-mini",
         "o3-deep-research",
+        # `gpt-4o-search-preview` and the mini variant are chat-with-
+        # retrieval models that respond to standard /v1/chat/completions
+        # and absolutely belong in the picker. The previous regex
+        # dropped them as a false positive of `(?:^|-)search`; the
+        # tightened denylist intentionally only catches the standalone
+        # search-api suffix.
+        "gpt-4o-search-preview",
+        "gpt-4o-mini-search-preview",
         # Hypothetical future families that the old allowlist would have
         # silently dropped -- they MUST surface under the new denylist.
         "gpt-5.6",
@@ -84,11 +92,13 @@ def test_openai_drops_non_chat_ids():
         "gpt-realtime-2",
         "gpt-4o-realtime-preview",
         "gpt-4o-transcribe",
-        "gpt-4o-search-preview",
-        "gpt-4o-mini-search-preview",
         "gpt-4o-mini-transcribe",
         "gpt-4o-mini-tts",
         "omni-moderation-latest",
+        # Standalone search API endpoint (distinct from the
+        # `*-search-preview` chat models above).
+        "gpt-5-search-api",
+        "gpt-5-search-api-2025-10-14",
         # Video generation.
         "sora-2",
         "sora-2-pro",
@@ -111,6 +121,58 @@ def test_openai_drops_non_chat_ids():
     assert surviving == [], surviving
 
 
+def test_openai_search_preview_is_kept_search_api_is_dropped():
+    # Pin the search-vs-search-api distinction so a future regex tweak
+    # doesn't silently regress to dropping chat-with-search models.
+    kept = _apply(
+        "openai",
+        [
+            "gpt-4o-search-preview",
+            "gpt-4o-mini-search-preview",
+        ],
+    )
+    assert kept == [
+        "gpt-4o-search-preview",
+        "gpt-4o-mini-search-preview",
+    ], kept
+
+    dropped = _apply(
+        "openai",
+        [
+            "gpt-5-search-api",
+            "gpt-5-search-api-2025-10-14",
+        ],
+    )
+    assert dropped == [], dropped
+
+
+def test_openai_legacy_completion_names_only_match_at_id_start():
+    # Hypothetical future chat ids that happen to contain a legacy
+    # completion-family name mid-string must NOT be dropped. The
+    # `^(?:babbage|davinci|ada|curie)\b` anchor is what makes this safe.
+    kept = _apply(
+        "openai",
+        [
+            "gpt-7-davinci-edition",
+            "gpt-7-ada-chat",
+            "gpt-7-curie-pro",
+            "gpt-7-babbage-mini",
+        ],
+    )
+    assert kept == [
+        "gpt-7-davinci-edition",
+        "gpt-7-ada-chat",
+        "gpt-7-curie-pro",
+        "gpt-7-babbage-mini",
+    ], kept
+    # ...but the actual legacy-base ids stay dropped.
+    dropped = _apply(
+        "openai",
+        ["babbage-002", "davinci-002", "text-davinci-003"],
+    )
+    assert dropped == [], dropped
+
+
 # ── Anthropic: empty denylist; dated ids ARE canonical ───────────────
 
 

From c51526d244ab352c4ec38bea0abd8e3af92f0ab8 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 22 May 2026 11:38:30 +0000
Subject: [PATCH 04/13] Address review: keep gpt-audio family in the picker

Codex P2 caught that `(?:^|-)audio` mid-id matching dropped
`gpt-audio`, `gpt-audio-mini`, `gpt-audio-1.5` -- but those are
chat-completion-capable models per OpenAI's audio guide (text in /
text or audio out via /v1/chat/completions and /v1/responses). The
denylist was silently hiding supported chat options.

Removed `audio` from the mid-id feature suffix set. The narrower
non-chat audio endpoints stay dropped:
  - `tts-*` / `gpt-*-tts`: text-to-speech only
  - `whisper-*`: STT only
  - `gpt-*-transcribe`: /v1/audio/transcriptions only
  - `gpt-realtime*`: /v1/realtime (WebSocket) endpoint, not chat

Updated the regex comment to spell out the chat/non-chat split per
family. Existing test added gpt-audio* to the keep list and
moved gpt-audio-1.5 out of the drop list; new test
test_openai_audio_family_is_chat_capable_and_kept pins the
contract so a future regex tweak can't silently undo it.
---
 studio/backend/core/inference/providers.py    | 19 +++++++--
 .../tests/test_provider_registry_filters.py   | 40 ++++++++++++++++++-
 2 files changed, 53 insertions(+), 6 deletions(-)

diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index bebebdc396..850566757f 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -35,13 +35,23 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
         # new chat family auto-appears the moment OpenAI lists it.
         #
         # Pattern strategy:
-        # - Unambiguous *feature* indicators (tts/whisper/transcribe/audio/
-        #   realtime/image/embedding/moderation/sora) match anywhere in
-        #   the id with `(?:^|-)` because OpenAI uses them as the
+        # - Unambiguous *feature* indicators (tts/whisper/transcribe/
+        #   realtime/image/embedding/moderation/sora) match anywhere
+        #   in the id with `(?:^|-)` because OpenAI uses them as the
         #   primary qualifier on every variant -- e.g. canonical
         #   `tts-1` AND family variants `gpt-4o-tts`, `gpt-4o-mini-tts`.
         #   These words don't appear in legitimate chat model ids
         #   mid-string, so the mid-id match is intentional and safe.
+        # - `audio` is INTENTIONALLY NOT in the mid-id set: the
+        #   `gpt-audio` / `gpt-audio-mini` / `gpt-audio-1.5` family is
+        #   chat-completion-capable (text in / text or audio out via
+        #   /v1/chat/completions and /v1/responses), so dropping them
+        #   would hide supported chat models from the picker. The
+        #   audio-only endpoint families (`tts-*`, `whisper-*`,
+        #   `*-transcribe`) ARE caught above.
+        # - `realtime` keeps the mid-id match because the
+        #   `gpt-realtime*` family is a separate /v1/realtime
+        #   (WebSocket) endpoint, not chat-completions.
         # - `search` is INTENTIONALLY NOT in the mid-id set because
         #   `gpt-4o-search-preview` and `gpt-4o-mini-search-preview` are
         #   chat-with-retrieval models that absolutely belong in the
@@ -59,7 +69,8 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
         # Verified against the live /v1/models listing 2026-05-22.
         "model_id_denylist": re.compile(
             # Feature suffixes that mark a non-chat variant on any base.
-            r"(?:^|-)(?:embedding|tts|whisper|moderation|image|audio|"
+            # Note: `audio` is omitted on purpose -- see docstring above.
+            r"(?:^|-)(?:embedding|tts|whisper|moderation|image|"
             r"realtime|transcribe|sora)\b"
             # Legacy completion bases -- ^-anchored to avoid false
             # positives on hypothetical future chat ids containing
diff --git a/studio/backend/tests/test_provider_registry_filters.py b/studio/backend/tests/test_provider_registry_filters.py
index 8c20d12173..c055739c30 100644
--- a/studio/backend/tests/test_provider_registry_filters.py
+++ b/studio/backend/tests/test_provider_registry_filters.py
@@ -56,6 +56,13 @@ def test_openai_keeps_every_known_chat_family():
         # search-api suffix.
         "gpt-4o-search-preview",
         "gpt-4o-mini-search-preview",
+        # gpt-audio family: chat-completion-capable (text in / text or
+        # audio out) per OpenAI's audio guide. Earlier revisions of
+        # this denylist dropped them via `(?:^|-)audio`; the regex
+        # comment block now keeps them.
+        "gpt-audio",
+        "gpt-audio-mini",
+        "gpt-audio-1.5",
         # Hypothetical future families that the old allowlist would have
         # silently dropped -- they MUST surface under the new denylist.
         "gpt-5.6",
@@ -70,9 +77,39 @@ def test_openai_keeps_every_known_chat_family():
     assert surviving == live, surviving
 
 
+def test_openai_audio_family_is_chat_capable_and_kept():
+    # Pin the chat/non-chat split that lives in the regex comment:
+    # gpt-audio is kept (chat completions accept it), gpt-realtime is
+    # dropped (separate /v1/realtime endpoint), gpt-4o-transcribe is
+    # dropped (audio-only input via /v1/audio/transcriptions).
+    kept = _apply(
+        "openai",
+        [
+            "gpt-audio",
+            "gpt-audio-1.5",
+            "gpt-audio-mini",
+        ],
+    )
+    assert kept == ["gpt-audio", "gpt-audio-1.5", "gpt-audio-mini"], kept
+
+    dropped = _apply(
+        "openai",
+        [
+            "gpt-realtime",
+            "gpt-realtime-mini",
+            "gpt-4o-realtime-preview",
+            "gpt-4o-transcribe",
+            "gpt-4o-mini-transcribe",
+        ],
+    )
+    assert dropped == [], dropped
+
+
 def test_openai_drops_non_chat_ids():
     noise = [
-        # Embeddings / TTS / image / moderation / whisper / audio etc.
+        # Embeddings / TTS / image / moderation / whisper / etc.
+        # gpt-audio* is intentionally OMITTED from this list -- it is
+        # chat-capable. See test_openai_audio_family_is_chat_capable_and_kept.
         "text-embedding-3-small",
         "text-embedding-3-large",
         "text-embedding-ada-002",
@@ -88,7 +125,6 @@ def test_openai_drops_non_chat_ids():
         "gpt-image-2",
         "gpt-image-1-mini",
         "chatgpt-image-latest",
-        "gpt-audio-1.5",
         "gpt-realtime-2",
         "gpt-4o-realtime-preview",
         "gpt-4o-transcribe",

From 6ea20c5aa7c4241b1c65651939893a6791c4dfd9 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 22 May 2026 12:34:38 +0000
Subject: [PATCH 05/13] Address review: keep gpt-realtime family in the picker

Codex P1 follow-up to the audio-family fix: gpt-realtime /
gpt-realtime-mini / gpt-realtime-1.5 / gpt-realtime-2 are also
chat-completion-capable today per OpenAI's realtime guide. They
accept text in via /v1/chat/completions and /v1/responses, not
only the specialised /v1/realtime WebSocket transport. The
previous regex hid them from the picker AND the new regression
test pinned the regression -- both are fixed here.

Removed `realtime` from the mid-id feature suffix set, matching
the same reasoning that kept `audio` out. The audio-only endpoint
families (`tts-*`, `whisper-*`, `*-transcribe`) still get caught
above. Renamed the regression test to cover both families.
---
 studio/backend/core/inference/providers.py    | 34 +++++++++---------
 .../tests/test_provider_registry_filters.py   | 35 ++++++++++++-------
 2 files changed, 40 insertions(+), 29 deletions(-)

diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index 850566757f..018f3786ed 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -36,22 +36,21 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
         #
         # Pattern strategy:
         # - Unambiguous *feature* indicators (tts/whisper/transcribe/
-        #   realtime/image/embedding/moderation/sora) match anywhere
-        #   in the id with `(?:^|-)` because OpenAI uses them as the
-        #   primary qualifier on every variant -- e.g. canonical
-        #   `tts-1` AND family variants `gpt-4o-tts`, `gpt-4o-mini-tts`.
-        #   These words don't appear in legitimate chat model ids
-        #   mid-string, so the mid-id match is intentional and safe.
-        # - `audio` is INTENTIONALLY NOT in the mid-id set: the
-        #   `gpt-audio` / `gpt-audio-mini` / `gpt-audio-1.5` family is
-        #   chat-completion-capable (text in / text or audio out via
-        #   /v1/chat/completions and /v1/responses), so dropping them
-        #   would hide supported chat models from the picker. The
+        #   image/embedding/moderation/sora) match anywhere in the id
+        #   with `(?:^|-)` because OpenAI uses them as the primary
+        #   qualifier on every variant -- e.g. canonical `tts-1` AND
+        #   family variants `gpt-4o-tts`, `gpt-4o-mini-tts`. These
+        #   words don't appear in legitimate chat model ids mid-string,
+        #   so the mid-id match is intentional and safe.
+        # - `audio` and `realtime` are INTENTIONALLY NOT in the
+        #   mid-id set. Both the `gpt-audio*` and `gpt-realtime*`
+        #   families are chat/responses-capable today -- per OpenAI's
+        #   audio and realtime guides they accept text in via
+        #   /v1/chat/completions and /v1/responses, not only the
+        #   specialised /v1/realtime WebSocket transport. Dropping
+        #   them hid supported chat models from the picker. The
         #   audio-only endpoint families (`tts-*`, `whisper-*`,
-        #   `*-transcribe`) ARE caught above.
-        # - `realtime` keeps the mid-id match because the
-        #   `gpt-realtime*` family is a separate /v1/realtime
-        #   (WebSocket) endpoint, not chat-completions.
+        #   `*-transcribe`) are still caught above.
         # - `search` is INTENTIONALLY NOT in the mid-id set because
         #   `gpt-4o-search-preview` and `gpt-4o-mini-search-preview` are
         #   chat-with-retrieval models that absolutely belong in the
@@ -69,9 +68,10 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
         # Verified against the live /v1/models listing 2026-05-22.
         "model_id_denylist": re.compile(
             # Feature suffixes that mark a non-chat variant on any base.
-            # Note: `audio` is omitted on purpose -- see docstring above.
+            # Note: `audio` and `realtime` are omitted on purpose --
+            # see docstring above.
             r"(?:^|-)(?:embedding|tts|whisper|moderation|image|"
-            r"realtime|transcribe|sora)\b"
+            r"transcribe|sora)\b"
             # Legacy completion bases -- ^-anchored to avoid false
             # positives on hypothetical future chat ids containing
             # those words mid-string.
diff --git a/studio/backend/tests/test_provider_registry_filters.py b/studio/backend/tests/test_provider_registry_filters.py
index c055739c30..f8451e67ff 100644
--- a/studio/backend/tests/test_provider_registry_filters.py
+++ b/studio/backend/tests/test_provider_registry_filters.py
@@ -77,27 +77,39 @@ def test_openai_keeps_every_known_chat_family():
     assert surviving == live, surviving
 
 
-def test_openai_audio_family_is_chat_capable_and_kept():
+def test_openai_audio_and_realtime_families_are_chat_capable_and_kept():
     # Pin the chat/non-chat split that lives in the regex comment:
-    # gpt-audio is kept (chat completions accept it), gpt-realtime is
-    # dropped (separate /v1/realtime endpoint), gpt-4o-transcribe is
-    # dropped (audio-only input via /v1/audio/transcriptions).
+    # gpt-audio AND gpt-realtime are kept (both accept text in via
+    # /v1/chat/completions and /v1/responses, not only their
+    # specialised transports); gpt-4o-transcribe etc. are dropped
+    # (audio-only input via /v1/audio/transcriptions).
     kept = _apply(
         "openai",
         [
             "gpt-audio",
             "gpt-audio-1.5",
             "gpt-audio-mini",
+            "gpt-realtime",
+            "gpt-realtime-mini",
+            "gpt-realtime-1.5",
+            "gpt-realtime-2",
+            "gpt-4o-realtime-preview",
         ],
     )
-    assert kept == ["gpt-audio", "gpt-audio-1.5", "gpt-audio-mini"], kept
+    assert kept == [
+        "gpt-audio",
+        "gpt-audio-1.5",
+        "gpt-audio-mini",
+        "gpt-realtime",
+        "gpt-realtime-mini",
+        "gpt-realtime-1.5",
+        "gpt-realtime-2",
+        "gpt-4o-realtime-preview",
+    ], kept
 
     dropped = _apply(
         "openai",
         [
-            "gpt-realtime",
-            "gpt-realtime-mini",
-            "gpt-4o-realtime-preview",
             "gpt-4o-transcribe",
             "gpt-4o-mini-transcribe",
         ],
@@ -108,8 +120,9 @@ def test_openai_audio_family_is_chat_capable_and_kept():
 def test_openai_drops_non_chat_ids():
     noise = [
         # Embeddings / TTS / image / moderation / whisper / etc.
-        # gpt-audio* is intentionally OMITTED from this list -- it is
-        # chat-capable. See test_openai_audio_family_is_chat_capable_and_kept.
+        # gpt-audio* and gpt-realtime* are intentionally OMITTED from
+        # this list -- both are chat-capable. See
+        # test_openai_audio_and_realtime_families_are_chat_capable_and_kept.
         "text-embedding-3-small",
         "text-embedding-3-large",
         "text-embedding-ada-002",
@@ -125,8 +138,6 @@ def test_openai_drops_non_chat_ids():
         "gpt-image-2",
         "gpt-image-1-mini",
         "chatgpt-image-latest",
-        "gpt-realtime-2",
-        "gpt-4o-realtime-preview",
         "gpt-4o-transcribe",
         "gpt-4o-mini-transcribe",
         "gpt-4o-mini-tts",

From 4104691b98b3ca729a6098af3e4084d9e31e06ac Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 22 May 2026 12:48:06 +0000
Subject: [PATCH 06/13] Address review: scope OpenAI sampling drop to reasoning
 families only

Codex P1: the denylist-based picker now admits non-reasoning chat
families (gpt-4o*, gpt-4.1*, gpt-3.5-turbo*, gpt-audio*, gpt-realtime*)
that ACCEPT temperature and top_p on /v1/responses. But
_stream_openai_responses was unconditionally `del temperature, top_p`
on the assumption that the old allowlist only let reasoning-class
ids through. Users selecting newly-surfaced gpt-4o etc. saw their
slider settings silently ignored, breaking reproducibility.

Added an _OPENAI_REASONING_FAMILY regex covering gpt-5.x, o1/o3/o4,
and gpt-4.5 (the only families that reject the sampling knobs).
For non-reasoning models, forward temperature and top_p verbatim
when the caller passes them; for reasoning models, keep dropping
them with a comment pointing at reasoning.effort as the only
accepted knob.

Tests added:
- test_responses_forwards_sampling_for_non_reasoning_chat_families:
  8 models in (gpt-4o family, gpt-4.1, gpt-4, gpt-3.5-turbo,
  gpt-audio, gpt-realtime) all surface temperature=0.42 / top_p=0.85
  on the outbound body.
- test_responses_still_drops_sampling_for_reasoning_families: 7
  reasoning models (gpt-5.5 / gpt-5.4 / gpt-5 / o1 / o3-mini /
  o4-mini / gpt-4.5-preview) still get both fields stripped.

Existing gpt-5.5 test docstring updated to point at the sibling
non-reasoning test instead of claiming the registry allowlist
only admits reasoning families.
---
 .../core/inference/external_provider.py       | 51 ++++++++--
 .../test_openai_responses_translation.py      | 98 ++++++++++++++++++-
 2 files changed, 136 insertions(+), 13 deletions(-)

diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index 25e1725337..781ef2c612 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -68,6 +68,23 @@ _ANTHROPIC_4_7_SAMPLING_REMOVED = re.compile(
 )
 _OPENAI_REASONING_SUMMARY_UNSUPPORTED = re.compile(r"^o3(?:[-.]|$)")
 
+# Reasoning-class OpenAI families that reject `temperature` and `top_p`
+# with "Unsupported parameter" 400s on /v1/responses (and on
+# /v1/chat/completions for the same families). Sampling knobs MUST be
+# dropped for these. Other chat-completion families (gpt-4o*,
+# gpt-4.1*, gpt-3.5-turbo*, gpt-4*, gpt-audio*, gpt-realtime*)
+# accept the standard sampling shape and we forward both fields
+# verbatim so the user's slider settings actually take effect.
+#
+# Source: OpenAI's reasoning guide
+# (https://developers.openai.com/api/docs/guides/reasoning) and the
+# /v1/responses reference. Covers gpt-5.x family, o1/o3/o4 family,
+# and gpt-4.5 family. The `(?:[-.]|$)` anchor keeps gpt-50 / o30
+# hypotheticals out.
+_OPENAI_REASONING_FAMILY = re.compile(
+    r"^(?:gpt-5(?:\.\d+)?|o[134]|gpt-4\.5)(?:[-.]|$)"
+)
+
 
 class _AnthropicThinkingSpec(NamedTuple):
     prefixes: tuple[str, ...]
@@ -2650,21 +2667,35 @@ class ExternalProviderClient:
                 if translated_parts:
                     input_items.append({"role": role, "content": translated_parts})
 
-        # NOTE: gpt-5.x / o3 / gpt-4.5 are reasoning-class models. They reject
-        # temperature and top_p with `Unsupported parameter` 400s on
-        # /v1/responses (and on /v1/chat/completions for the same families).
-        # The PROVIDER_REGISTRY['openai'] model_id_allowlist already scopes
-        # the picker to those families, so we never need to send sampling
-        # knobs here. ``reasoning.effort`` defaults to "medium" server-side
-        # if omitted — surface it in a future commit if a knob is wanted.
-        del temperature, top_p  # explicit drop — params are accepted for
-        # API symmetry with the other stream methods but not forwarded.
-
+        # Sampling knob handling. The old allowlist-scoped picker only
+        # admitted reasoning-class families (gpt-5.x / o3 / gpt-4.5),
+        # all of which reject `temperature` and `top_p` with
+        # "Unsupported parameter" 400s on /v1/responses -- so the
+        # earlier code unconditionally dropped both.
+        #
+        # The denylist-based picker (PR 5684) now also admits
+        # non-reasoning chat models (gpt-4o*, gpt-4.1*, gpt-4*,
+        # gpt-3.5-turbo*, gpt-audio*, gpt-realtime*) that accept the
+        # standard sampling shape. Forward temperature/top_p for those;
+        # keep dropping them for the reasoning families.
         body: dict[str, Any] = {
             "model": model,
             "input": input_items,
             "stream": True,
         }
+        is_reasoning_family = bool(
+            _OPENAI_REASONING_FAMILY.match(model.strip().lower())
+        )
+        if not is_reasoning_family:
+            if temperature is not None:
+                body["temperature"] = temperature
+            if top_p is not None:
+                body["top_p"] = top_p
+        else:
+            # Reasoning-class: explicit drop. ``reasoning.effort`` is
+            # the only knob the API accepts and defaults to "medium"
+            # server-side if omitted.
+            del temperature, top_p
         # `summary: "auto"` is what makes /v1/responses emit reasoning
         # summary events — without it OpenAI returns no thinking text on
         # most reasoning models, the SSE handler has no <think>…</think>
diff --git a/studio/backend/tests/test_openai_responses_translation.py b/studio/backend/tests/test_openai_responses_translation.py
index 22ccba7058..51bb3c2343 100644
--- a/studio/backend/tests/test_openai_responses_translation.py
+++ b/studio/backend/tests/test_openai_responses_translation.py
@@ -101,9 +101,11 @@ def test_responses_request_body_uses_input_and_instructions(monkeypatch):
     assert body["input"] == [{"role": "user", "content": "Hi"}]
     assert body["max_output_tokens"] == 512
     assert body["stream"] is True
-    # Responses API on reasoning-class models (gpt-5.x / o3 / gpt-4.5 — the
-    # only OpenAI ids the registry allowlist exposes) rejects these as
-    # `Unsupported parameter`. Make sure we never silently forward them.
+    # gpt-5.5 is reasoning-class: Responses API rejects temperature
+    # and top_p with `Unsupported parameter` 400s, so the helper must
+    # drop them. Non-reasoning models (gpt-4o*, gpt-4.1*, gpt-3.5-turbo*,
+    # gpt-audio*, gpt-realtime*) are covered in a sibling test that
+    # asserts the inverse forwarding.
     assert "temperature" not in body
     assert "top_p" not in body
     assert "presence_penalty" not in body
@@ -112,6 +114,96 @@ def test_responses_request_body_uses_input_and_instructions(monkeypatch):
     assert "messages" not in body
 
 
+def test_responses_forwards_sampling_for_non_reasoning_chat_families(monkeypatch):
+    # Codex P1 follow-up to the picker filter change (PR 5684): the
+    # OpenAI denylist now admits non-reasoning chat families
+    # (gpt-4o*, gpt-4.1*, gpt-3.5-turbo*, gpt-audio*, gpt-realtime*).
+    # The Responses API ACCEPTS temperature and top_p on those, so
+    # the helper must forward the user's slider settings instead of
+    # silently dropping them.
+    for model in (
+        "gpt-4o",
+        "gpt-4o-mini",
+        "gpt-4o-2026-01-01",
+        "gpt-4.1",
+        "gpt-4",
+        "gpt-3.5-turbo",
+        "gpt-audio",
+        "gpt-realtime",
+    ):
+        captured: dict = {}
+
+        def handler(request: httpx.Request, _captured = captured) -> httpx.Response:
+            _captured["body"] = json.loads(request.content.decode("utf-8"))
+            return httpx.Response(
+                200,
+                content = _responses_sse(
+                    [{"type": "response.completed", "response": {}}]
+                ),
+                headers = {"content-type": "text/event-stream"},
+            )
+
+        _mock_http_client(monkeypatch, handler)
+
+        async def run():
+            client = _make_client()
+            async for _ in client._stream_openai_responses(
+                messages = [{"role": "user", "content": "Hi"}],
+                model = model,
+                temperature = 0.42,
+                top_p = 0.85,
+                max_tokens = 32,
+                enable_thinking = None,
+                reasoning_effort = None,
+            ):
+                pass
+            await client.close()
+
+        _drive(run())
+        body = captured["body"]
+        assert body["temperature"] == 0.42, (model, body)
+        assert body["top_p"] == 0.85, (model, body)
+
+
+def test_responses_still_drops_sampling_for_reasoning_families(monkeypatch):
+    # Sanity: parametrise across the reasoning-class families so a
+    # future regex tweak that accidentally weakens the drop surfaces
+    # here.
+    for model in ("gpt-5.5", "gpt-5.4", "gpt-5", "o1", "o3-mini", "o4-mini", "gpt-4.5-preview"):
+        captured: dict = {}
+
+        def handler(request: httpx.Request, _captured = captured) -> httpx.Response:
+            _captured["body"] = json.loads(request.content.decode("utf-8"))
+            return httpx.Response(
+                200,
+                content = _responses_sse(
+                    [{"type": "response.completed", "response": {}}]
+                ),
+                headers = {"content-type": "text/event-stream"},
+            )
+
+        _mock_http_client(monkeypatch, handler)
+
+        async def run():
+            client = _make_client()
+            async for _ in client._stream_openai_responses(
+                messages = [{"role": "user", "content": "Hi"}],
+                model = model,
+                temperature = 0.42,
+                top_p = 0.85,
+                max_tokens = 32,
+                enable_thinking = None,
+                reasoning_effort = None,
+            ):
+                pass
+            await client.close()
+
+        _drive(run())
+        body = captured["body"]
+        assert "temperature" not in body, (model, body)
+        assert "top_p" not in body, (model, body)
+
+
 def test_responses_translates_image_parts(monkeypatch):
     captured: dict = {}
 

From 27732404a36f5e1c81ead2210f37ebfd44f36d9c Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Fri, 22 May 2026 12:48:18 +0000
Subject: [PATCH 07/13] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/external_provider.py     |  4 +---
 .../backend/tests/test_openai_responses_translation.py | 10 +++++++++-
 2 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index 781ef2c612..b63c1f818c 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -81,9 +81,7 @@ _OPENAI_REASONING_SUMMARY_UNSUPPORTED = re.compile(r"^o3(?:[-.]|$)")
 # /v1/responses reference. Covers gpt-5.x family, o1/o3/o4 family,
 # and gpt-4.5 family. The `(?:[-.]|$)` anchor keeps gpt-50 / o30
 # hypotheticals out.
-_OPENAI_REASONING_FAMILY = re.compile(
-    r"^(?:gpt-5(?:\.\d+)?|o[134]|gpt-4\.5)(?:[-.]|$)"
-)
+_OPENAI_REASONING_FAMILY = re.compile(r"^(?:gpt-5(?:\.\d+)?|o[134]|gpt-4\.5)(?:[-.]|$)")
 
 
 class _AnthropicThinkingSpec(NamedTuple):
diff --git a/studio/backend/tests/test_openai_responses_translation.py b/studio/backend/tests/test_openai_responses_translation.py
index 51bb3c2343..390a6926f1 100644
--- a/studio/backend/tests/test_openai_responses_translation.py
+++ b/studio/backend/tests/test_openai_responses_translation.py
@@ -169,7 +169,15 @@ def test_responses_still_drops_sampling_for_reasoning_families(monkeypatch):
     # Sanity: parametrise across the reasoning-class families so a
     # future regex tweak that accidentally weakens the drop surfaces
     # here.
-    for model in ("gpt-5.5", "gpt-5.4", "gpt-5", "o1", "o3-mini", "o4-mini", "gpt-4.5-preview"):
+    for model in (
+        "gpt-5.5",
+        "gpt-5.4",
+        "gpt-5",
+        "o1",
+        "o3-mini",
+        "o4-mini",
+        "gpt-4.5-preview",
+    ):
         captured: dict = {}
 
         def handler(request: httpx.Request, _captured = captured) -> httpx.Response:

From 3661243e350a573132cf7be7b05e4b15c42bdc43 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 22 May 2026 17:33:24 +0000
Subject: [PATCH 08/13] Address review: deny realtime-translate and *-instruct
 in OpenAI filter

Codex flagged two leaks in the new denylist:

  * `gpt-realtime-translate*` / `gpt-audio-translate*` are audio
    translation variants. They share the gpt-realtime/gpt-audio
    name root we now intentionally keep, but the chat picker only
    knows the chat/responses transport and they would 4xx on
    selection.

  * `gpt-3.5-turbo-instruct*` (and any future legacy id ending in
    -instruct) speaks /v1/completions, not chat/responses. The
    previous denylist let them through because nothing in the
    pattern matched.

Extend the feature-suffix alternation to cover `translate` and
`instruct`. New regression tests pin both exclusions plus the
fact that bare gpt-realtime / gpt-realtime-mini / gpt-audio
still survive.
---
 studio/backend/core/inference/providers.py    | 11 +++-
 .../tests/test_provider_registry_filters.py   | 50 +++++++++++++++++++
 2 files changed, 59 insertions(+), 2 deletions(-)

diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index 018f3786ed..166f4abf63 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -69,9 +69,16 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
         "model_id_denylist": re.compile(
             # Feature suffixes that mark a non-chat variant on any base.
             # Note: `audio` and `realtime` are omitted on purpose --
-            # see docstring above.
+            # see docstring above. `translate` is here so the audio
+            # translation variants `gpt-realtime-translate*` /
+            # `gpt-audio-translate*` are dropped even though we keep
+            # the parent `gpt-realtime` / `gpt-audio` chat families.
+            # `instruct` catches the legacy completion-only suffix
+            # used by `gpt-3.5-turbo-instruct*`; that family is wired
+            # to /v1/completions and would 4xx on our chat/responses
+            # transport, so it should not surface in the chat picker.
             r"(?:^|-)(?:embedding|tts|whisper|moderation|image|"
-            r"transcribe|sora)\b"
+            r"transcribe|translate|instruct|sora)\b"
             # Legacy completion bases -- ^-anchored to avoid false
             # positives on hypothetical future chat ids containing
             # those words mid-string.
diff --git a/studio/backend/tests/test_provider_registry_filters.py b/studio/backend/tests/test_provider_registry_filters.py
index f8451e67ff..d366764a5f 100644
--- a/studio/backend/tests/test_provider_registry_filters.py
+++ b/studio/backend/tests/test_provider_registry_filters.py
@@ -168,6 +168,56 @@ def test_openai_drops_non_chat_ids():
     assert surviving == [], surviving
 
 
+def test_openai_realtime_translate_variants_are_dropped_but_parent_chat_family_survives():
+    """gpt-realtime / gpt-realtime-mini are chat-capable and stay,
+    but the audio-translation variants share the chat picker's
+    transport and would 4xx, so they must be filtered. Pin both
+    behaviours so a future regex tweak cannot collapse one into
+    the other."""
+    kept = _apply(
+        "openai",
+        [
+            "gpt-realtime",
+            "gpt-realtime-mini",
+            "gpt-audio",
+            "gpt-audio-mini",
+        ],
+    )
+    assert kept == [
+        "gpt-realtime",
+        "gpt-realtime-mini",
+        "gpt-audio",
+        "gpt-audio-mini",
+    ], kept
+
+    dropped = _apply(
+        "openai",
+        [
+            "gpt-realtime-translate",
+            "gpt-realtime-translate-mini",
+            "gpt-audio-translate",
+            "gpt-4o-realtime-translate-preview",
+        ],
+    )
+    assert dropped == [], dropped
+
+
+def test_openai_legacy_instruct_completion_ids_are_dropped():
+    """Legacy `*-instruct` completion-only ids (gpt-3.5-turbo-instruct
+    and friends) speak /v1/completions, not chat/responses. Our OpenAI
+    bridge only knows the chat/responses transport, so admitting them
+    into the picker would 4xx every selection."""
+    dropped = _apply(
+        "openai",
+        [
+            "gpt-3.5-turbo-instruct",
+            "gpt-3.5-turbo-instruct-0914",
+            "davinci-002-instruct",
+        ],
+    )
+    assert dropped == [], dropped
+
+
 def test_openai_search_preview_is_kept_search_api_is_dropped():
     # Pin the search-vs-search-api distinction so a future regex tweak
     # doesn't silently regress to dropping chat-with-search models.

From 1a7238c753030d8fb745090ea503add46b998703 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 23 May 2026 16:33:32 +0000
Subject: [PATCH 09/13] Studio: drop non-streaming gpt-realtime +
 gpt-audio-mini from picker

Codex P1 round 2 on this PR: by allowing the full gpt-realtime and
gpt-audio families through the denylist we let the chat picker surface
IDs that OpenAI's model cards mark Streaming: Not supported (gpt-
realtime, gpt-realtime-mini, gpt-audio-mini -- Realtime-API/WebSocket
only). Studio's path always sends stream:true, so selecting them 4xx'd
at request time instead of producing a reply.

Narrow the regex to drop ^gpt-realtime(?:$|-) and ^gpt-audio-mini\b;
gpt-audio (chat-streams over /v1/chat/completions) and gpt-4o-realtime-
preview (chat-streams via the /v1/responses adapter) still pass.

Updated test_provider_registry_filters.py to pin the new chat/non-chat
split: 9/9 filter tests pass.
---
 studio/backend/core/inference/providers.py    | 20 +++---
 .../tests/test_provider_registry_filters.py   | 69 +++++++------------
 2 files changed, 36 insertions(+), 53 deletions(-)

diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index 166f4abf63..7ff800e7e9 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -68,17 +68,19 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
         # Verified against the live /v1/models listing 2026-05-22.
         "model_id_denylist": re.compile(
             # Feature suffixes that mark a non-chat variant on any base.
-            # Note: `audio` and `realtime` are omitted on purpose --
-            # see docstring above. `translate` is here so the audio
-            # translation variants `gpt-realtime-translate*` /
-            # `gpt-audio-translate*` are dropped even though we keep
-            # the parent `gpt-realtime` / `gpt-audio` chat families.
-            # `instruct` catches the legacy completion-only suffix
-            # used by `gpt-3.5-turbo-instruct*`; that family is wired
-            # to /v1/completions and would 4xx on our chat/responses
-            # transport, so it should not surface in the chat picker.
+            # `audio`/`realtime` are NOT in this group so `gpt-audio`
+            # (chat-capable, streaming) stays; specific non-streaming
+            # IDs are caught further down.
             r"(?:^|-)(?:embedding|tts|whisper|moderation|image|"
             r"transcribe|translate|instruct|sora)\b"
+            # OpenAI realtime/audio variants that don't support chat
+            # streaming. Studio always sends `stream: true`, and OpenAI
+            # marks `gpt-realtime*` and `gpt-audio-mini` as Streaming:
+            # Not supported (Realtime API only); surfacing them in the
+            # picker would 4xx at request time. `gpt-audio` and
+            # `gpt-4o-realtime-preview` still stream and are kept.
+            r"|^gpt-realtime(?:$|-)"
+            r"|^gpt-audio-mini\b"
             # Legacy completion bases -- ^-anchored to avoid false
             # positives on hypothetical future chat ids containing
             # those words mid-string.
diff --git a/studio/backend/tests/test_provider_registry_filters.py b/studio/backend/tests/test_provider_registry_filters.py
index d366764a5f..3aa9b23b2e 100644
--- a/studio/backend/tests/test_provider_registry_filters.py
+++ b/studio/backend/tests/test_provider_registry_filters.py
@@ -56,12 +56,10 @@ def test_openai_keeps_every_known_chat_family():
         # search-api suffix.
         "gpt-4o-search-preview",
         "gpt-4o-mini-search-preview",
-        # gpt-audio family: chat-completion-capable (text in / text or
-        # audio out) per OpenAI's audio guide. Earlier revisions of
-        # this denylist dropped them via `(?:^|-)audio`; the regex
-        # comment block now keeps them.
+        # gpt-audio is chat-completion-capable with streaming. The
+        # `-mini` variant does NOT stream (Realtime API only) and is
+        # asserted dropped in the realtime/audio split test below.
         "gpt-audio",
-        "gpt-audio-mini",
         "gpt-audio-1.5",
         # Hypothetical future families that the old allowlist would have
         # silently dropped -- they MUST surface under the new denylist.
@@ -77,39 +75,35 @@ def test_openai_keeps_every_known_chat_family():
     assert surviving == live, surviving
 
 
-def test_openai_audio_and_realtime_families_are_chat_capable_and_kept():
-    # Pin the chat/non-chat split that lives in the regex comment:
-    # gpt-audio AND gpt-realtime are kept (both accept text in via
-    # /v1/chat/completions and /v1/responses, not only their
-    # specialised transports); gpt-4o-transcribe etc. are dropped
-    # (audio-only input via /v1/audio/transcriptions).
+def test_openai_audio_and_realtime_families_split_on_streaming_support():
+    # Pin the chat/non-chat split: Studio always sends `stream: true`,
+    # so the picker keeps gpt-audio* (chat streams over /v1/chat/
+    # completions) and gpt-4o-realtime-preview (chat streams via the
+    # /v1/responses adapter), but drops the new gpt-realtime* family
+    # and gpt-audio-mini, which OpenAI's model cards mark Streaming:
+    # Not supported (Realtime API / WebSocket only).
     kept = _apply(
         "openai",
         [
             "gpt-audio",
             "gpt-audio-1.5",
-            "gpt-audio-mini",
-            "gpt-realtime",
-            "gpt-realtime-mini",
-            "gpt-realtime-1.5",
-            "gpt-realtime-2",
             "gpt-4o-realtime-preview",
         ],
     )
     assert kept == [
         "gpt-audio",
         "gpt-audio-1.5",
-        "gpt-audio-mini",
-        "gpt-realtime",
-        "gpt-realtime-mini",
-        "gpt-realtime-1.5",
-        "gpt-realtime-2",
         "gpt-4o-realtime-preview",
     ], kept
 
     dropped = _apply(
         "openai",
         [
+            "gpt-audio-mini",
+            "gpt-realtime",
+            "gpt-realtime-mini",
+            "gpt-realtime-1.5",
+            "gpt-realtime-2",
             "gpt-4o-transcribe",
             "gpt-4o-mini-transcribe",
         ],
@@ -120,9 +114,9 @@ def test_openai_audio_and_realtime_families_are_chat_capable_and_kept():
 def test_openai_drops_non_chat_ids():
     noise = [
         # Embeddings / TTS / image / moderation / whisper / etc.
-        # gpt-audio* and gpt-realtime* are intentionally OMITTED from
-        # this list -- both are chat-capable. See
-        # test_openai_audio_and_realtime_families_are_chat_capable_and_kept.
+        # gpt-audio (no -mini) and gpt-4o-realtime-preview are
+        # intentionally OMITTED from this list -- they DO stream chat.
+        # See test_openai_audio_and_realtime_families_split_on_streaming_support.
         "text-embedding-3-small",
         "text-embedding-3-large",
         "text-embedding-ada-002",
@@ -169,26 +163,13 @@ def test_openai_drops_non_chat_ids():
 
 
 def test_openai_realtime_translate_variants_are_dropped_but_parent_chat_family_survives():
-    """gpt-realtime / gpt-realtime-mini are chat-capable and stay,
-    but the audio-translation variants share the chat picker's
-    transport and would 4xx, so they must be filtered. Pin both
-    behaviours so a future regex tweak cannot collapse one into
-    the other."""
-    kept = _apply(
-        "openai",
-        [
-            "gpt-realtime",
-            "gpt-realtime-mini",
-            "gpt-audio",
-            "gpt-audio-mini",
-        ],
-    )
-    assert kept == [
-        "gpt-realtime",
-        "gpt-realtime-mini",
-        "gpt-audio",
-        "gpt-audio-mini",
-    ], kept
+    """gpt-audio is chat-capable (streams over /v1/chat/completions)
+    and stays, but the audio-translation variants share the chat
+    picker's transport and would 4xx, so they must be filtered. The
+    new gpt-realtime* family and gpt-audio-mini are Realtime-only and
+    are dropped by a sibling assertion."""
+    kept = _apply("openai", ["gpt-audio"])
+    assert kept == ["gpt-audio"], kept
 
     dropped = _apply(
         "openai",

From e457a493c266a957237ce46229399c7fcd60f4e2 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 23 May 2026 16:49:54 +0000
Subject: [PATCH 10/13] fix(openai): also drop gpt-4o-realtime-preview from
 picker

Codex follow-up P1 on PR #5684 (providers.py:83): the prior round kept
`gpt-4o-realtime-preview` because I'd assumed it streams via the
/v1/responses adapter, but OpenAI's own model card and the Realtime API
docs confirm it is Realtime-only (WebRTC / WebSocket transport).
Studio always sends `stream: true` over Chat Completions / Responses,
so surfacing the model from the picker would 4xx at request time.

Extends the denylist with `^gpt-4o(?:-mini)?-realtime\b`, covering
`gpt-4o-realtime-preview` and `gpt-4o-mini-realtime-preview` (and any
future dated/suffixed variants). The audio-chat families
(`gpt-audio`, `gpt-audio-1.5`, `gpt-4o-audio-preview`) still stream
over Chat Completions and are explicitly kept. `gpt-audio-mini`
remains denied (Realtime-only per OpenAI).

Updated `test_openai_audio_and_realtime_families_split_on_streaming_support`
to pin the new split: `gpt-4o-audio-preview` is the surviving audio
chat id, and `gpt-4o-realtime-preview` / `gpt-4o-mini-realtime-preview`
join the dropped Realtime cohort. All 9 filter tests pass locally.
---
 studio/backend/core/inference/providers.py    | 12 ++++---
 .../tests/test_provider_registry_filters.py   | 32 ++++++++++++-------
 2 files changed, 28 insertions(+), 16 deletions(-)

diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index 7ff800e7e9..9ec1db3e6d 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -75,11 +75,15 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
             r"transcribe|translate|instruct|sora)\b"
             # OpenAI realtime/audio variants that don't support chat
             # streaming. Studio always sends `stream: true`, and OpenAI
-            # marks `gpt-realtime*` and `gpt-audio-mini` as Streaming:
-            # Not supported (Realtime API only); surfacing them in the
-            # picker would 4xx at request time. `gpt-audio` and
-            # `gpt-4o-realtime-preview` still stream and are kept.
+            # marks the full Realtime family as Streaming: Not supported
+            # (Realtime API only, WebRTC/WebSocket transport): `gpt-realtime`,
+            # `gpt-realtime-mini`, `gpt-4o-realtime-preview*` and
+            # `gpt-4o-mini-realtime-preview*`, plus `gpt-audio-mini`.
+            # Selecting them from the picker would 4xx at request time.
+            # `gpt-audio` and `gpt-4o-audio-preview` are kept -- they
+            # stream over Chat Completions / Responses.
             r"|^gpt-realtime(?:$|-)"
+            r"|^gpt-4o(?:-mini)?-realtime\b"
             r"|^gpt-audio-mini\b"
             # Legacy completion bases -- ^-anchored to avoid false
             # positives on hypothetical future chat ids containing
diff --git a/studio/backend/tests/test_provider_registry_filters.py b/studio/backend/tests/test_provider_registry_filters.py
index 3aa9b23b2e..e061fffac9 100644
--- a/studio/backend/tests/test_provider_registry_filters.py
+++ b/studio/backend/tests/test_provider_registry_filters.py
@@ -77,23 +77,28 @@ def test_openai_keeps_every_known_chat_family():
 
 def test_openai_audio_and_realtime_families_split_on_streaming_support():
     # Pin the chat/non-chat split: Studio always sends `stream: true`,
-    # so the picker keeps gpt-audio* (chat streams over /v1/chat/
-    # completions) and gpt-4o-realtime-preview (chat streams via the
-    # /v1/responses adapter), but drops the new gpt-realtime* family
-    # and gpt-audio-mini, which OpenAI's model cards mark Streaming:
-    # Not supported (Realtime API / WebSocket only).
+    # so the picker keeps the audio chat families (gpt-audio* without
+    # -mini, plus gpt-4o-audio-preview*) which stream over
+    # /v1/chat/completions and /v1/responses, but drops the entire
+    # Realtime family (gpt-realtime*, gpt-4o-realtime-preview*,
+    # gpt-4o-mini-realtime-preview*) and gpt-audio-mini, which OpenAI's
+    # model cards mark Streaming: Not supported (Realtime API only,
+    # WebRTC/WebSocket transport).
     kept = _apply(
         "openai",
         [
             "gpt-audio",
             "gpt-audio-1.5",
-            "gpt-4o-realtime-preview",
+            "gpt-4o-audio-preview",
+            "gpt-4o-audio-preview-2024-12-17",
         ],
     )
     assert kept == [
         "gpt-audio",
         "gpt-audio-1.5",
-        "gpt-4o-realtime-preview",
+        "gpt-4o-audio-preview",
+        # Dated audio-preview snapshot is dropped by the date-suffix rule;
+        # the canonical id above survives.
     ], kept
 
     dropped = _apply(
@@ -104,6 +109,8 @@ def test_openai_audio_and_realtime_families_split_on_streaming_support():
             "gpt-realtime-mini",
             "gpt-realtime-1.5",
             "gpt-realtime-2",
+            "gpt-4o-realtime-preview",
+            "gpt-4o-mini-realtime-preview",
             "gpt-4o-transcribe",
             "gpt-4o-mini-transcribe",
         ],
@@ -114,9 +121,9 @@ def test_openai_audio_and_realtime_families_split_on_streaming_support():
 def test_openai_drops_non_chat_ids():
     noise = [
         # Embeddings / TTS / image / moderation / whisper / etc.
-        # gpt-audio (no -mini) and gpt-4o-realtime-preview are
-        # intentionally OMITTED from this list -- they DO stream chat.
-        # See test_openai_audio_and_realtime_families_split_on_streaming_support.
+        # gpt-audio (no -mini) and gpt-4o-audio-preview are intentionally
+        # OMITTED from this list -- they DO stream chat. See
+        # test_openai_audio_and_realtime_families_split_on_streaming_support.
         "text-embedding-3-small",
         "text-embedding-3-large",
         "text-embedding-ada-002",
@@ -166,8 +173,9 @@ def test_openai_realtime_translate_variants_are_dropped_but_parent_chat_family_s
     """gpt-audio is chat-capable (streams over /v1/chat/completions)
     and stays, but the audio-translation variants share the chat
     picker's transport and would 4xx, so they must be filtered. The
-    new gpt-realtime* family and gpt-audio-mini are Realtime-only and
-    are dropped by a sibling assertion."""
+    full Realtime family (gpt-realtime*, gpt-4o*-realtime-preview*)
+    plus gpt-audio-mini are Realtime-only and dropped by a sibling
+    assertion."""
     kept = _apply("openai", ["gpt-audio"])
     assert kept == ["gpt-audio"], kept
 

From d8e8cbb38a3d8cbb057680443cea340069f7de01 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 23 May 2026 17:31:20 +0000
Subject: [PATCH 11/13] ci: re-trigger after transient Upload Playwright
 artifacts failure


From 2efc67761d92873a7b222711fcaa65f7bb0c5335 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 23 May 2026 18:31:04 +0000
Subject: [PATCH 12/13] ci: re-trigger after transient HTTP error fetching
 ggml-org/llama.cpp release metadata


From f5ace8824cf2aacf59abc673a003a6d5e3a2972f Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 23 May 2026 18:33:27 +0000
Subject: [PATCH 13/13] fix(openai): drop legacy MMDD snapshot suffixes from
 picker

Codex P2 follow-up on PR #5684 (providers.py:104): the new OpenAI
denylist only stripped the modern `-YYYY-MM-DD` dated form, so
legacy compact snapshots like `gpt-3.5-turbo-0125`, `gpt-4-0613`,
`gpt-4-1106-preview`, and `gpt-4-0125-preview` slipped through. The
listing returns both the snapshot and the canonical id; the picker
should surface only the canonical so users do not accidentally
select a deprecated snapshot.

Extends the dated-suffix branch with `-\d{4}(?:-preview)?$` to
cover the compact `MMDD` form and the `-MMDD-preview` variant.
The new rule is anchored to the end of the id so canonical chat
ids whose minor version happens to be a year-like number
(`gpt-4.5`, `gpt-5.5`, `o3`) are unaffected.

Added `test_openai_legacy_compact_snapshot_suffixes_are_dropped`
to pin both sides (8 snapshot ids dropped, 7 canonical ids kept).
10 / 10 filter tests pass locally.
---
 studio/backend/core/inference/providers.py    |  8 +++-
 .../tests/test_provider_registry_filters.py   | 47 +++++++++++++++++++
 2 files changed, 54 insertions(+), 1 deletion(-)

diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index 9ec1db3e6d..82111e9c90 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -100,8 +100,14 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
             # Fine-tunes carry the user's tenant in the id.
             r"|^ft:"
             # Dated snapshot suffixes hide behind the canonical id
-            # which is also in the listing.
+            # which is also in the listing. Covers both the modern
+            # `-YYYY-MM-DD` dated form and the legacy compact
+            # 4-digit `MMDD` form (e.g. `gpt-3.5-turbo-0125`,
+            # `gpt-4-0613`, `gpt-4-1106-preview`) so the picker only
+            # surfaces the canonical id even when the listing still
+            # returns the snapshot copy.
             r"|-\d{4}-\d{2}-\d{2}$"
+            r"|-\d{4}(?:-preview)?$"
         ),
     },
     "anthropic": {
diff --git a/studio/backend/tests/test_provider_registry_filters.py b/studio/backend/tests/test_provider_registry_filters.py
index e061fffac9..bd766261dd 100644
--- a/studio/backend/tests/test_provider_registry_filters.py
+++ b/studio/backend/tests/test_provider_registry_filters.py
@@ -207,6 +207,53 @@ def test_openai_legacy_instruct_completion_ids_are_dropped():
     assert dropped == [], dropped
 
 
+def test_openai_legacy_compact_snapshot_suffixes_are_dropped():
+    """Legacy `-MMDD` snapshot suffixes (gpt-3.5-turbo-0125,
+    gpt-4-0613, gpt-4-1106-preview, etc.) hide behind the canonical
+    id which the listing also returns; surface only the canonical so
+    users do not pick a deprecated snapshot by accident. The
+    `-\\d{4}(?:-preview)?$` rule must not catch canonical ids whose
+    minor version happens to be a year-like number (e.g. gpt-4.5,
+    o3) -- those are tested as KEEP below."""
+    dropped = _apply(
+        "openai",
+        [
+            "gpt-3.5-turbo-0125",
+            "gpt-3.5-turbo-0301",
+            "gpt-3.5-turbo-16k-0613",
+            "gpt-4-0613",
+            "gpt-4-0314",
+            "gpt-4-32k-0613",
+            "gpt-4-1106-preview",
+            "gpt-4-0125-preview",
+        ],
+    )
+    assert dropped == [], dropped
+
+    # Canonical chat ids that share a digit-heavy tail must survive.
+    kept = _apply(
+        "openai",
+        [
+            "gpt-3.5-turbo",
+            "gpt-4o",
+            "gpt-4.5",
+            "gpt-5.5",
+            "gpt-5.5-mini",
+            "gpt-5.5-pro",
+            "o3",
+        ],
+    )
+    assert set(kept) >= {
+        "gpt-3.5-turbo",
+        "gpt-4o",
+        "gpt-4.5",
+        "gpt-5.5",
+        "gpt-5.5-mini",
+        "gpt-5.5-pro",
+        "o3",
+    }, kept
+
+
 def test_openai_search_preview_is_kept_search_api_is_dropped():
     # Pin the search-vs-search-api distinction so a future regex tweak
     # doesn't silently regress to dropping chat-with-search models.