fix(bedrock): map reasoning_effort to reasoning.effort for OpenAI GPT-5.x on Converse

[6matt]

TLDR

Problem this solves:

• OpenAI GPT-5.x on Bedrock Converse silently loses reasoning_effort
• Sent as Anthropic thinking, Converse 400s with unknown_parameter

How it solves it:

• Map effort to additionalModelRequestFields.reasoning.effort for GPT-5.x
• Set supports_reasoning on the six bedrock_converse gpt-5.6 entries

User Flow

Before: a developer calling a GPT-5.6 Bedrock Converse model with reasoning effort gets a hard 400 and no reasoning

1. They POST https://litellm-domain/v1/chat/completions with "model": "gpt-5.6-sol-bedrock" and "reasoning_effort": "high"
2. The response is 400 with bedrock does not support parameters: ['reasoning_effort']
3. If they set drop_params: true to get past the 400, the answer comes back but the model ran with no reasoning applied

After: the same request returns 200 and the model runs at the requested effort

1. They POST https://litellm-domain/v1/chat/completions with the same "model": "gpt-5.6-sol-bedrock" and "reasoning_effort": "high"
2. The response is 200 with the assistant answer
3. No drop_params workaround is needed and the effort reaches Bedrock

Relevant issues

Related to #34105 (same root cause: the Converse reasoning-effort handler assumes any non gpt-oss / non Nova-2 model is Anthropic and discards the value). This PR fixes the GPT-5.x case; Qwen3 and other providers in that issue are out of scope here

Linear ticket

Resolves LIT-5984

Pre-Submission checklist

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

• I have added meaningful tests
• The handful of test files covering my change pass locally, e.g. uv run pytest tests/test_litellm/<your_test_file>.py -v. Leave the suites (make test-unit-*, make test-unit) to CI: it finishes in ~15 minutes where a laptop takes an hour or more
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Real Bedrock Converse calls against us.openai.gpt-5.6-sol, no mocks, billed to a live AWS account. Both legs run the identical config and requests and differ only in the commit the proxy was booted from: Before at the merge base 1ff615c335, After at the PR tip 418012aac5. Each leg is one proxy with --num_workers 2 (Before on port 18342, After on port 27519)

Config (qa_config.yaml):

model_list:
  - model_name: gpt-5.6-sol-bedrock
    litellm_params:
      model: bedrock/converse/us.openai.gpt-5.6-sol
      aws_region_name: us-east-1
litellm_settings:
  drop_params: false

Proxy launched with AWS_BEARER_TOKEN_BEDROCK=... LITELLM_MASTER_KEY=sk-qa5984 litellm --config qa_config.yaml --port <port> --num_workers 2

Before (1ff615c)

/v1/chat/completions with reasoning_effort

1. Send the request:

curl -s http://localhost:18342/v1/chat/completions \
  -H "Authorization: Bearer sk-qa5984" -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.6-sol-bedrock","messages":[{"role":"user","content":"In one sentence, what is 17 * 23? Show the number."}],"reasoning_effort":"high"}'

2. Observed 400:

{"error":{"message":"litellm.UnsupportedParamsError: bedrock does not support parameters: ['reasoning_effort'], for model=converse/us.openai.gpt-5.6-sol. To drop these, set `litellm.drop_params=True` ...","type":"None","param":null,"code":"400"}}

/v1/responses with reasoning.effort

1. Send the request:

curl -s http://localhost:18342/v1/responses \
  -H "Authorization: Bearer sk-qa5984" -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.6-sol-bedrock","input":"In one sentence, what is 19 * 21? Show the number.","reasoning":{"effort":"high"}}'

2. Observed the same 400 UnsupportedParamsError naming reasoning_effort

/v1/messages with reasoning_effort

1. Send the request:

curl -s http://localhost:18342/v1/messages \
  -H "Authorization: Bearer sk-qa5984" -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.6-sol-bedrock","max_tokens":512,"messages":[{"role":"user","content":"In one sentence, what is 13 * 29? Show the number."}],"reasoning_effort":"high"}'

2. Observed the same 400 UnsupportedParamsError naming reasoning_effort

After (418012a)

/v1/chat/completions with reasoning_effort

1. Send the identical request against the tip proxy on port 27519
2. Observed 200:

{"model":"gpt-5.6-sol-bedrock","object":"chat.completion","choices":[{"finish_reason":"stop","index":0,"message":{"content":"17 × 23 = 391.","role":"assistant"}}],"usage":{"completion_tokens":12,"prompt_tokens":22,"total_tokens":34}}

/v1/responses with reasoning.effort

1. Send the identical request against the tip proxy
2. Observed 200 with "status": "completed" and output text 19 × 21 = 399.

/v1/messages with reasoning_effort

1. Send the identical request against the tip proxy
2. Observed 200 with content text 13 × 29 = 377.

Notes:

• Before 400 with drop_params: false proves refusal, never silent drop
• So After 200 means Bedrock accepted additionalModelRequestFields.reasoning.effort
• Direct boto3: thinking shape gets ValidationException, reasoning.effort shape accepted
• Harder prompt billed 36 completion tokens, 5 visible: hidden reasoning real
• Usage lacks separate reasoning token counts here; pre-existing, unchanged

Type

🐛 Bug Fix

Caveats (if any)

Low

• Scoped to OpenAI GPT-5.x; other Converse providers in [Bug]: Bedrock Converse silently drops reasoning_effort for non-Anthropic/Nova2/GPT-OSS models (e.g. Qwen3 on Bedrock) #34105 still drop effort
• thinking and output_config stay rejected for GPT-5.x, as before this PR
• Opaque inference-profile ARN aliases still misroute effort; pre-existing, every model family
• Verified live on us.openai.gpt-5.6-sol; terra and luna share the same code path

Changes

_handle_reasoning_effort_parameter gains a branch for openai.gpt-5 models that maps reasoning_effort to {"reasoning": {"effort": ...}} under additionalModelRequestFields. The six bedrock_converse gpt-5.6 entries (us. and global. for sol, terra, luna) gain supports_reasoning: true as correct capability metadata

get_supported_openai_params advertises only reasoning_effort for GPT-5.x (not the Anthropic-only thinking / output_config), and map_openai_params skips the thinking mapping for these models. Together this closes an order-dependent leak where a request combining thinking with reasoning_effort could still serialize a thinking block that Bedrock rejects, and it makes such a request fail fast with a clear client-side error instead. Parametrized regression tests cover the supported-param gate, the mapped shape, and that thinking never survives regardless of parameter order

---

Note

Medium Risk
Touches shared Bedrock Converse parameter routing for reasoning; wrong model-id matching could affect non–GPT-5 models, but scope is narrow and covered by regression tests.

Overview
Fixes OpenAI GPT-5.x on Bedrock Converse so reasoning_effort is honored instead of rejected or misrouted to Anthropic thinking (which Bedrock returns as an unknown parameter).

Bedrock Converse mapping adds a GPT-5 branch in _handle_reasoning_effort_parameter that sets additionalModelRequestFields.reasoning to {"effort": <value>}. get_supported_openai_params now advertises reasoning_effort for these models, and map_openai_params no longer applies the thinking mapper when the model id contains openai.gpt-5.

Model metadata sets supports_reasoning: true on the six bedrock_converse GPT-5.6 variants in the cost/context JSON files. A BedrockConverseGptReasoningEffortBlock TypedDict documents the effort shape.

Tests assert the mapped request shape, that thinking/output_config are not offered or forwarded (including when combined with reasoning_effort), and update cross-region profile expectations accordingly.

^{Reviewed by Cursor Bugbot for commit 418012a. Bugbot is set up for automated code reviews on this repo. Configure here.}

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

• 418012a passes /live-pr-risk

[greptile-apps]

Greptile Summary

This PR maps Bedrock Converse GPT-5.x reasoning effort to the provider’s reasoning.effort request shape

• Advertises reasoning support for six GPT-5.6 Bedrock Converse profiles
• Prevents Anthropic-only thinking fields from being forwarded for GPT-5.x
• Adds typed request metadata and focused parameter-mapping regression coverage

Confidence Score: 5/5

The PR appears safe to merge

No blocking failure remains

Important Files Changed

Filename	Overview
litellm/llms/bedrock/chat/converse_transformation.py	Adds GPT-5.x reasoning capability detection and maps reasoning effort without restoring Anthropic thinking fields
litellm/types/llms/bedrock.py	Adds a read-only typed dictionary for the GPT reasoning effort request block
model_prices_and_context_window.json	Marks six Bedrock Converse GPT-5.6 profiles as supporting reasoning
litellm/model_prices_and_context_window_backup.json	Keeps backup model capability metadata aligned with the primary catalog
tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py	Covers GPT-5.x support reporting, wire mapping, and parameter-order behavior
tests/test_litellm/llms/bedrock/test_cross_region_inference_profile_mapping.py	Updates cross-region profile expectations for GPT-5.6 reasoning support

_{Reviews (4): Last reviewed commit: "fix(bedrock): type GPT-5 reasoning field..." | Re-trigger Greptile}

[6matt]

@greptileai please review the new commit (9cc276a) and update the confidence score if applicable

Addressing the earlier findings:

P1 (parameter order restores thinking): fixed. get_supported_openai_params now advertises only reasoning_effort for GPT-5.x, not the Anthropic-only thinking/output_config, so _check_valid_arg drops or rejects a stray thinking before map_openai_params runs. map_openai_params also skips the thinking branch for these models, so neither parameter order can leak a thinking block into additionalModelRequestFields. A regression test feeds thinking and reasoning_effort in both orders and asserts only reasoning reaches the request

P2 (redundant reasoning comment): removed

P2 (hardcoded GPT-5 classification): kept as an intentional call. The supports_reasoning flag is the config-driven gate for whether a model reasons at all, and it is set on these entries. What no config field encodes is the reasoning wire shape, which differs per family: gpt-oss passes reasoning_effort raw, Nova 2 uses reasoningConfig, Anthropic uses thinking, and GPT-5.x uses reasoning.effort. Selecting that shape by model name matches the existing gpt-oss and Nova 2 branches in this same file. A reasoning_format capability field would be the config-driven alternative but is a broader change beyond this bug

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing 6matt:litellm_bedrock_converse_gpt5_reasoning_effort (418012a) with litellm_internal_staging (ba8d8b6)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (434add7) during the generation of this report, so ba8d8b6 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[mateo-berri]

bugbot run

[mateo-berri]

@greptileai

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 418012a. Configure here.}

[greptile-apps]

Aliases use Anthropic reasoning

When a Bedrock Converse deployment uses an alias without openai.gpt-5 while its resolved base_model identifies GPT-5, parameter detection accepts reasoning_effort but this branch classifies only the alias and falls through to the Anthropic thinking mapping, causing Bedrock to reject the request with a 400.

[mateo-berri]

Pre-existing: opaque profile ARNs map to thinking identically at merge base, for gpt-oss and Nova 2 too. Plumbing base_model through map_openai_params is out of scope.

[greptile-apps]

Agreed. Since opaque profile ARNs already have this behavior at merge base for GPT-OSS and Nova 2, and threading base_model through map_openai_params would expand the scope, I’m withdrawing this as a PR-blocking finding.

[mateo-berri]

@greptileai

[mateo-berri]

LGTM. Thanks for the contribution!