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",