fix(anthropic): handle per-level reasoning_effort flags without supports_reasoning

[tin-berri]

Problem

gpt-5-search-api declares supports_minimal_reasoning_effort: true but has no supports_reasoning flag. The resolver was treating this as "no reasoning support" and falling back to the chain floor, causing minimal requests to downgrade to low.

Solution

When a model has any per-level flag set (not None), treat supports_reasoning as implicitly True and proceed through the flag resolution path. This allows models like gpt-5-search-api to correctly advertise and accept their declared effort levels.

Behavior change

20 azure deployments (those with max or xhigh per-level flags set) will now forward those efforts instead of degrading to high. This is intended (avoiding unnecessary degradation) but is a latency/cost change; operators should monitor reasoning effort changes in logs if they see bill increases.

Testing

• ✅ New test verifies per-level flag enables resolution without explicit supports_reasoning
• ✅ All existing reasoning effort capability tests pass

Fixes the regression identified in PR #38492 follow-up review.

---

Note

Medium Risk
Changes router-advertised reasoning effort levels for map entries with per-level flags but no supports_reasoning, which can alter forwarded effort (including higher tiers on some Azure deployments) and affect latency/cost.

Overview
Fixes incorrect downgrades when a catalog entry sets per-level supports_*_reasoning_effort flags but omits supports_reasoning (e.g. gpt-5-search-api with supports_minimal_reasoning_effort: true).

resolve_supported_reasoning_efforts now treats any non-None per-level flag as implicit evidence that reasoning is supported, so resolution continues through the flag/opt-in/opt-out path instead of behaving like a mapped non-reasoning model (()) or unknown. supports_reasoning: false is checked first and still returns no efforts, even if per-level flags are present.

Tests cover implicit enablement from a lone per-level flag and that explicit supports_reasoning: false overrides inherited per-level metadata.

^{Reviewed by Cursor Bugbot for commit 4603619. Bugbot is set up for automated code reviews on this repo. Configure here.}

[greptile-apps]

Greptile Summary

The PR corrects reasoning-effort capability resolution when per-level flags are present without a top-level supports_reasoning declaration.

• Treats declared per-level effort flags as implicit reasoning support.
• Preserves explicit supports_reasoning: false as the highest-precedence opt-out.
• Adds regression tests for implicit support and explicit opt-out precedence.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
litellm/router_utils/reasoning_effort_capability.py	Correctly infers reasoning support from per-level metadata while returning no efforts immediately for an explicit opt-out.
tests/test_litellm/router_utils/test_reasoning_effort_capability.py	Adds focused coverage for per-level implicit support and explicit-false precedence without weakening existing assertions.

_{Reviews (3): Last reviewed commit: "fix(anthropic): explicit supports_reason..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[tin-berri]

@greptileai can you reply here with score & why

[tin-berri]

bugbot run

[greptile-apps]

Score: 5/5. The change correctly distinguishes explicit supports_reasoning=False, unset reasoning support with no per-level metadata, and implicit support when any per-level flag is present. It preserves existing mapped/off-map unknown semantics, honors inherited bare-model flags, and leaves declared effort lists and Azure none gating intact. The added tests cover both the new implicit-support path and the explicit-false override; I found no blocking issues.

[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 4603619. Configure here.}

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing fix/reasoning-effort-per-level-flags (4603619) with litellm_internal_staging (98c5233)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (3300fc3) during the generation of this report, so 98c5233 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[tin-berri]

@greptileai review pls

[tin-berri]

@greptileai based on the response / comment, can you give me a score now!