fix(bedrock_mantle): correct grok-4.3 and gemma-4 token limits

[xyos]

Title

fix(bedrock_mantle): correct grok-4.3 and gemma-4 token limits

Relevant issues

None filed — found while auditing bedrock_mantle metadata against the live endpoint. Same class of issue as #36299 (GPT-5.6 context window), different models.

Pre-Submission checklist

• I have Added testing in the tests/litellm/ directory
• I have added a screenshot/output of my test passing (below)
• Keep scope isolated

Type

🐛 Bug Fix

Changes

Four bedrock_mantle entries advertise token limits that don't match what the service enforces:

model	field	current	actual
xai.grok-4.3	input	131072	1048576
xai.grok-4.3	output	16384	1048576
google.gemma-4-31b	input	256000	262144
google.gemma-4-26b-a4b	input	256000	262144
google.gemma-4-e2b	input	128000	131072

xai.grok-4.3 is the significant one — the input limit is understated 8x. As with #36299, the non-Bedrock counterpart already carries the larger value (xai/grok-4.3 reads 1000000), which is what suggested these were stale rather than deliberate.

Understated limits cause premature context-window errors and early auto-compaction, and clamp max_tokens far below what the model accepts.

Verification (live endpoint, bedrock-mantle us-east-2)

Input — the service reports its own cap:

xai.grok-4.3           -> prompt tokens (1300031) exceed customer model maximum (1048576)
google.gemma-4-31b     -> prompt tokens (1300034) exceed customer model maximum (262144)
google.gemma-4-26b-a4b -> prompt tokens (1300034) exceed customer model maximum (262144)
google.gemma-4-e2b     -> prompt tokens (1300030) exceed customer model maximum (131072)

A real request with 394,937 prompt tokens — 3x the currently advertised max — completed successfully on xai.grok-4.3 and correctly returned a fact planted at 85% depth, so the window is usable rather than merely accepted:

prompt_tokens : 394,937
finish_reason : stop
content       : copper-meridian-19
needle found  : True

Output — same caps apply, verified at the boundary:

xai.grok-4.3  max_tokens=1048576 -> accepted
xai.grok-4.3  max_tokens=1048577 -> 'max_tokens' (1048577) exceeds model maximum (1048576)
gemma-4-31b   max_tokens=300000  -> 'max_tokens' (300000) exceeds model maximum (262144)
gemma-4-e2b   max_tokens=140000  -> 'max_tokens' (140000) exceeds model maximum (131072)

The gemma-4 entries already had max_output_tokens == max_input_tokens, so those move together.

Left alone deliberately

• openai.gpt-oss-* keep 131072 — they aren't served on the Responses or /openai/v1 chat routes (model ... isn't supported on this route), so I couldn't probe their caps the same way and didn't want to guess
• openai.gpt-5.x entries — handled separately in fix(bedrock_mantle): correct GPT-5.6 context window to 1M tokens #36299

Tests

Added test_context_windows_match_bedrock_limits and updated the existing test_gemma_4_bedrock_mantle_model_metadata parametrization, which pinned the old values.

$ pytest tests/test_litellm/llms/bedrock_mantle/ -q
167 passed in 4.40s

$ pytest tests/test_litellm/test_model_prices_schema.py \
         tests/test_litellm/litellm_core_utils/test_get_model_cost_map.py \
         tests/test_litellm/test_gpt_5_4_model_metadata.py -q
88 passed in 1.46s

Reverting only the JSON changes fails 7 tests, confirming the coverage is meaningful:

$ git stash -- model_prices_and_context_window.json ...backup.json
$ pytest tests/test_litellm/llms/bedrock_mantle/ -q
7 failed, 160 passed

ruff and black clean on the touched file.

Note on ci_cd/check_files_match.py

Both JSON files are updated identically. That check already fails on a clean litellm_internal_staging checkout — the backup has unrelated drift (zai/glm-5.1, openrouter/z-ai/glm-5.1, darkbloom/*). I kept this diff to just these four entries rather than absorbing that; happy to include the full sync if you'd prefer.

[greptile-apps]

Greptile Summary

The PR corrects Bedrock Mantle context and output token metadata for Grok 4.3 and three Gemma 4 models, updating both runtime metadata copies and their local tests.

• Raises Grok 4.3 input, output, and legacy token limits to 1,048,576.
• Aligns Gemma 4 limits with their binary boundary values.
• Adds assertions covering the revised input and output limits.

Confidence Score: 4/5

The PR appears safe to merge after the non-blocking source-comment convention issue is cleaned up.

The changed metadata is synchronized across both runtime sources and is covered by hermetic assertions; the only accepted concern is the newly added test docstring.

Files Needing Attention: tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_transformation.py

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Updates the canonical Bedrock Mantle token-limit metadata consistently with the stated endpoint boundaries.
litellm/model_prices_and_context_window_backup.json	Mirrors the four canonical metadata corrections in the bundled runtime fallback.
tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_transformation.py	Updates pinned Gemma limits and adds coverage for all revised entries, but introduces a docstring contrary to the repository convention.

_{Reviews (1): Last reviewed commit: "fix(bedrock_mantle): correct grok-4.3 an..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[xyos]

Removed the docstring — you're right, most tests in this file don't have one. Assertions and the JSON fix are unchanged.

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing xyos:fix/bedrock-mantle-grok-gemma-context-windows (c8e4b28) with litellm_internal_staging (f6b9518)}