diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py index 25e1725337..b63c1f818c 100644 --- a/studio/backend/core/inference/external_provider.py +++ b/studio/backend/core/inference/external_provider.py @@ -68,6 +68,21 @@ _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 +2665,35 @@ 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 … diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py index fef9ba3e12..82111e9c90 100644 --- a/studio/backend/core/inference/providers.py +++ b/studio/backend/core/inference/providers.py @@ -26,14 +26,89 @@ 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. + # + # Pattern strategy: + # - Unambiguous *feature* indicators (tts/whisper/transcribe/ + # 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 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 + # 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( + # Feature suffixes that mark a non-chat variant on any base. + # `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 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(?:$|-)" + r"|^gpt-4o(?:-mini)?-realtime\b" + r"|^gpt-audio-mini\b" + # Legacy completion bases -- ^-anchored to avoid false + # positives on hypothetical future chat ids containing + # those words mid-string. + r"|^(?:babbage|davinci|ada|curie)\b" + r"|^text-(?:embedding|moderation|davinci|curie|babbage|ada)\b" + # Standalone search API (separate endpoint shape). + # `gpt-4o-search-preview` (chat-with-search) is intentionally + # NOT matched here. + r"|-search-api(?:-\d{4}-\d{2}-\d{2})?$" + # Canonical non-chat prefixes. + r"|^dall-e\b" + r"|^computer-use\b" + # 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. 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": { "display_name": "Anthropic", @@ -42,16 +117,17 @@ 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_openai_responses_translation.py b/studio/backend/tests/test_openai_responses_translation.py index 22ccba7058..390a6926f1 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 +114,104 @@ 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 = {} 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..bd766261dd --- /dev/null +++ b/studio/backend/tests/test_provider_registry_filters.py @@ -0,0 +1,332 @@ +# 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", + # `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", + # 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-1.5", + # 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_audio_and_realtime_families_split_on_streaming_support(): + # Pin the chat/non-chat split: Studio always sends `stream: true`, + # 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-audio-preview", + "gpt-4o-audio-preview-2024-12-17", + ], + ) + assert kept == [ + "gpt-audio", + "gpt-audio-1.5", + "gpt-4o-audio-preview", + # Dated audio-preview snapshot is dropped by the date-suffix rule; + # the canonical id above survives. + ], kept + + dropped = _apply( + "openai", + [ + "gpt-audio-mini", + "gpt-realtime", + "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", + ], + ) + assert dropped == [], dropped + + +def test_openai_drops_non_chat_ids(): + noise = [ + # Embeddings / TTS / image / moderation / whisper / etc. + # 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", + "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-4o-transcribe", + "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", + # 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 + + +def test_openai_realtime_translate_variants_are_dropped_but_parent_chat_family_survives(): + """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 + 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 + + 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_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. + 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 ─────────────── + + +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