PR #25369: fix(tool_choice): respect GPT-5 and Azure AI models without explicit capability flag

• Repository: BerriAI/litellm
• Author: Hybirdss
• State: open | merged: False
• Link: https://github.com/BerriAI/litellm/pull/25369

Description (problem / solution / changelog)

Relevant issues

Fixes #25332.

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Type

🐛 Bug Fix

Changes

OpenAIGPT5Config.get_supported_openai_params and AzureAIStudioConfig.get_supported_openai_params both called supports_tool_choice() to decide whether to strip tool_choice from the supported params list. That helper returns False when the model cost map entry is missing the supports_tool_choice key, which is opt-in semantics — a missing flag means "not supported".

For GPT-5, this breaks any model that isn't in the OpenAI model registry yet:

• github_copilot/gpt-5, chatgpt/gpt-5.*, perplexity/openai/gpt-5.* — the provider-prefixed entries in model_prices_and_context_window.json are missing supports_tool_choice: true
• Any new GPT-5 model added before its cost map entry is updated
• Custom Azure deployment names that will never be in the cost map

The OpenAI API supports tool_choice for all non-chat GPT-5 models. Only gpt-5-chat* variants disable it, and those already have supports_tool_choice: false set explicitly.

The same file (gpt_5_transformation.py:129) already uses _is_explicitly_disabled_factory for reasoning effort — opt-out semantics, where a missing key means "allowed". This PR uses the same helper for tool_choice in the two transformation layers with the bug.

AzureOpenAIGPT5Config.get_supported_openai_params had a local workaround (if "tool_choice" not in params: params.append("tool_choice")) with a comment explaining the base config was dropping it for unknown Azure deployment names. With the base fix in place, the workaround is dead code and is removed in this PR.

Files

• litellm/llms/openai/chat/gpt_5_transformation.py — swap supports_tool_choice() for _is_explicitly_disabled_factory()
• litellm/llms/azure_ai/chat/transformation.py — same swap, add a clarifying comment
• litellm/llms/azure/chat/gpt_5_transformation.py — remove the now-redundant tool_choice workaround
• tests/test_litellm/llms/openai/test_gpt5_transformation.py — regression test for unknown GPT-5 model
• tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py — regression test for unknown Azure AI deployment
• tests/test_litellm/llms/azure/chat/test_azure_gpt5_transformation.py — regression test for unknown Azure GPT-5 deployment via the direct class path (replaces what the removed workaround was protecting)

Untouched

• gpt-5-chat* still has tool_choice removed (it has supports_tool_choice: false set explicitly, so _is_explicitly_disabled_factory correctly returns True)
• Fireworks AI and Bedrock use add-only opt-in patterns on purpose and are left alone
• supports_tool_choice() / _supports_factory utility functions are unchanged

Changed files

• litellm/llms/azure/chat/gpt_5_transformation.py (modified, +0/-5)
• litellm/llms/azure_ai/chat/transformation.py (modified, +10/-4)
• litellm/llms/openai/chat/gpt_5_transformation.py (modified, +7/-4)
• tests/test_litellm/llms/azure/chat/test_azure_gpt5_transformation.py (modified, +19/-0)
• tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py (modified, +22/-0)
• tests/test_litellm/llms/openai/test_gpt5_transformation.py (modified, +27/-0)

PR #25405: fix(model_prices): add supports_tool_choice for GPT-5 variants across providers

• Repository: BerriAI/litellm
• Author: avarga1
• State: open | merged: False
• Link: https://github.com/BerriAI/litellm/pull/25405

Description (problem / solution / changelog)

Changes

Fix — missing `supports_tool_choice` for GPT-5 variants (closes #25332)

23 GPT-5 model entries across `chatgpt/`, `github_copilot/`, `gmi/openai/`, `perplexity/openai/`, and `replicate/openai/` had `supports_tool_choice` absent from their entries in `model_prices_and_context_window.json`. Since `get_supported_openai_params` gates `tool_choice` on `supports_tool_choice(model)`, this caused:

litellm.UnsupportedParamsError: openai does not support parameters: ['tool_choice'], for model=gpt-5.4-mini

All affected entries already had `supports_function_calling: true`, confirming tool use is supported — `supports_tool_choice` was simply never added.

Fix: added `"supports_tool_choice": true` to all 23 affected entries. Databricks GPT-5 entries (`databricks-gpt-5`, etc.) are excluded as their `supports_function_calling` is also absent and their capabilities are unconfirmed.

Test plan

• `litellm.utils.supports_tool_choice(model="gpt-5.4-mini")` returns `True`
• `litellm.utils.supports_tool_choice(model="chatgpt/gpt-5.4")` returns `True`
• `litellm.utils.supports_tool_choice(model="github_copilot/gpt-5")` returns `True`
• `litellm.utils.supports_tool_choice(model="perplexity/openai/gpt-5.2")` returns `True`
• `litellm.utils.supports_tool_choice(model="replicate/openai/gpt-5-nano")` returns `True`

Changed files

• model_prices_and_context_window.json (modified, +47/-24)
• tests/test_litellm/test_utils.py (modified, +17/-16)