fix(bedrock): drop toolSpec.strict for four more Claude Converse ids

[pdecat]

Bedrock Converse routes Haiku 4.5, Sonnet 4.5/4.6 and Opus 4.6 through the same Anthropic-compatible validator as Opus 4.7/4.8 and Sonnet 4/5 (#31582), but their cost-map entries were never flagged, so toolSpec.strict is still forwarded for them.

Forwarding strict: true makes the model compile the tool schema into a constrained-decoding grammar, and that grammar rejects ordinary JSON Schema keywords. A tool carrying maxItems on an array property then fails the whole request:

BedrockException - {"message":"The model returned the following errors: tools.1.custom: For 'array' type, property 'maxItems' is not supported"}

Neither field fails on its own, only the pair. Measured through a proxy against Sonnet 4.6, two tools sent so the failing one is tools.1:

strict	maxItems	result
false	absent	200
false	present	200
true	absent	200
true	present	400

So it surfaces as a sudden breakage for tool-calling clients that set strict on their tools, on any upgrade crossing v1.90.0, where toolSpec.strict forwarding to Converse was introduced. Reproduced against Bedrock for Haiku 4.5, Sonnet 4.6 and Opus 4.6.

Change

bedrock_converse_supports_strict_tools: false on all 30 Bedrock entries for those four ids (bare key, every region prefix, the us-gov keys), in model_prices_and_context_window.json and its backup, matching how Opus 4.7/4.8, Sonnet 4 and Sonnet 5 are already flagged. No code changes: the gate is data, so the fix reaches deployments through the cost map rather than another hardcoded model pattern.

test_bedrock_converse_strict_tools_opus_47_48.py asserted the pre-fix behaviour for Sonnet 4.5/4.6 and Opus 4.6, listing them as models that accept strict, so those cases move to the "strict dropped" side and the cost-map assertions cover the new keys. 66 passed locally, plus tests/test_litellm/test_model_prices_schema.py and ci_cd/check_files_match.py.

Sonnet 4.5 has no separate repro of its own and is flagged as the remaining member of the same cohort. Happy to drop it if you would rather ship only reproduced ids.

Fixes #34388. Overlaps #34609, which flags two of these ids; this one covers all four and updates the test that contradicts them, so close whichever fits better.

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing pdecat:fix/bedrock-strict-tools-unflagged-claude-ids (dde0e60) with litellm_internal_staging (ec3f818)}

[greptile-apps]

Greptile Summary

The PR marks additional Claude Haiku 4.5, Sonnet 4.5/4.6, and Opus 4.6 Bedrock entries as not supporting Converse strict tools.

• Keeps the primary and backup model cost maps synchronized.
• Updates transformation, helper, and cost-map tests to verify that toolSpec.strict is removed for the affected model identifiers.

Confidence Score: 5/5

The PR appears safe to merge, with the new model metadata consistently applied and exercised by focused regression coverage.

The Bedrock Converse helper consumes these capability flags on the intended transformation path, affected prefixed identifiers resolve to flagged entries, non-Anthropic-prefixed forms are already guarded as unsupported, and both cost maps remain synchronized.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds the strict-tools capability override to the intended Bedrock model and regional entries without leaving a reachable affected identifier uncovered.
litellm/model_prices_and_context_window_backup.json	Mirrors the primary cost-map updates used by packaged local model metadata.
tests/test_litellm/litellm_core_utils/prompt_templates/test_bedrock_converse_strict_tools_opus_47_48.py	Moves the newly unsupported models to strict-removal coverage and verifies representative transformations and cost-map entries without weakening existing supported-model checks.

_{Reviews (1): Last reviewed commit: "fix(bedrock): drop toolSpec.strict for f..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!