fix(anthropic): allow effort="max" on Claude Opus 4.7

[shang309073819]

Title

fix(anthropic): allow effort="max" on Claude Opus 4.7

Relevant issues

Fixes #25957

Pre-Submission Checklist (Click to expand)

• I have Read the Contributing Guide
• I have added a test for this PR
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible

Type

🐛 Bug Fix

Changes

AnthropicConfig._apply_output_config previously rejected effort="max" for any model that wasn't Claude Opus 4.6:

# litellm/llms/anthropic/chat/transformation.py (before)
# ``max`` is Claude Opus 4.6 only (not Sonnet 4.6, not Opus 4.5/4.7).
# Keep this hardcoded so the error message is specific and stable.
if effort == "max" and not self._is_opus_4_6_model(model):
    raise ValueError(
        f"effort='max' is only supported by Claude Opus 4.6. "
        f"Got model: {model}"
    )

This guard was added in #22234 (back when 4.6 was the only model accepting max) and was intentionally kept in the day-0 4.7 PR (#25867). However, Anthropic's Messages API accepts output_config.effort="max" on claude-opus-4-7-* as well — calling the API directly with model=claude-opus-4-7-*, output_config={"effort":"max"}, and thinking={"type":"adaptive"} returns a normal 200 completion.

So LiteLLM was the only thing in the chain rejecting a perfectly valid configuration, breaking flows like Claude Code via ANTHROPIC_BASE_URL and any direct SDK call against Opus 4.7 with reasoning_effort="max".

This PR allows effort="max" on both Opus 4.6 and Opus 4.7. _is_opus_4_7_model already exists on AnthropicConfig (added in #25867), so the diff is minimal:

if effort == "max" and not (
    self._is_opus_4_6_model(model) or self._is_opus_4_7_model(model)
):
    raise ValueError(
        f"effort='max' is only supported by Claude Opus 4.6 and 4.7. "
        f"Got model: {model}"
    )

I kept the substring-style check (rather than switching to a model-map supports_max_reasoning_effort lookup) so date-variant model names (e.g. claude-opus-4-7-20260408) keep working without requiring every dated entry to be present in the cost map. Switching to a data-driven lookup would be a reasonable follow-up but is out of scope for this fix.

Tests

• Updated test_max_effort_rejected_for_opus_45 and test_max_effort_rejected_for_sonnet_46 to match the new error message.
• Added test_max_effort_accepted_for_opus_47 (regression for [Bug]: effort="max" hardcoded as Opus 4.6 only — Opus 4.7 rejected even though Anthropic API accepts it #25957).
• Existing test_max_effort_accepted_for_opus_46 still passes unchanged.

$ pytest tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py -k "effort or output_config or opus"
......................                                                   [100%]
22 passed, 115 deselected in 0.32s

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 17, 2026 11:28pm

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

This PR fixes #25957 by allowing effort=\"max\" on Claude Opus 4.7 (and also correctly enables it for Sonnet 4.6 per Anthropic's documentation). It fully addresses the prior review concern by replacing the hardcoded _is_opus_4_6_model substring guard with the data-driven _supports_effort_level(model, \"max\") helper, mirroring the existing xhigh pattern — enabling future models via a pure model-map change. The ProviderSpecificModelInfo TypedDict and _get_model_info_helper are updated to carry the new supports_max_reasoning_effort field.

Confidence Score: 5/5

Safe to merge; all remaining findings are P2 style nits and the core logic is correct.

The hardcoded model check is fully replaced with the data-driven approach required by the team's custom rule; supports_max_reasoning_effort is wired through all layers (TypedDict → helper → model map); tests are strengthened not weakened. The only open finding is a trivial indentation inconsistency in two JSON entries.

model_prices_and_context_window.json and its backup — minor indentation inconsistency on the vertex_ai/claude-sonnet-4-6 and vertex_ai/claude-sonnet-4-6@default entries.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/chat/transformation.py	Switches the effort=max guard from a hardcoded _is_opus_4_6_model substring check to the data-driven _supports_effort_level(model, max) helper, matching the team's model-map rule and the existing xhigh pattern.
model_prices_and_context_window.json	Adds supports_max_reasoning_effort: true to Opus 4.6/4.7 and Sonnet 4.6 entries (including regional/vertex variants); vertex_ai/claude-sonnet-4-6 and its @default variant have a minor indentation inconsistency on the new key.
litellm/model_prices_and_context_window_backup.json	Mirror of model_prices_and_context_window.json changes; same indentation inconsistency on vertex_ai/claude-sonnet-4-6 entries.
litellm/types/utils.py	Adds supports_max_reasoning_effort: Optional[bool] to ProviderSpecificModelInfo TypedDict, completing the type coverage for the new model-map key.
litellm/utils.py	Threads supports_max_reasoning_effort through _get_model_info_helper so the new TypedDict field is populated from the model cost map.
tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py	Updates error-message assertions to match the new generic wording, inverts test_max_effort_rejected_for_sonnet_46 to test_max_effort_accepted_for_sonnet_46 per Anthropic docs, and adds test_max_effort_accepted_for_opus_47 as a regression test for #25957.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["_apply_output_config(model, effort)"] --> B{effort == max?}
    B -- No --> E[Pass through to request]
    B -- Yes --> C["_supports_effort_level(model, 'max')"]
    C --> D["_get_model_info_helper looks up\nsupports_max_reasoning_effort\nin model cost map JSON"]
    D -- true --> E
    D -- "false or not found" --> F["raise ValueError: effort=max not supported"]
    E --> G[Set output_config on request data]

    subgraph "Entries with supports_max_reasoning_effort: true"
        M1["claude-opus-4-6 and dated variants"]
        M2["claude-opus-4-7 and dated variants"]
        M3["claude-sonnet-4-6 and dated variants"]
        M4["All regional and vertex_ai variants"]
    end

Loading

_{Reviews (7): Last reviewed commit: "test(anthropic): force LITELLM_LOCAL_MOD..." | Re-trigger Greptile}

[greptile-apps]

Hardcoded model-flag violates the "use model map" rule

The team rule requires that model-capability flags live in model_prices_and_context_window.json and be queried via _supports_effort_level / get_model_info, not hardcoded substring checks. The xhigh effort level already follows this pattern (one line below); max should too.

The Opus 4.7 entries (claude-opus-4-7, claude-opus-4-7-20260416) already carry supports_xhigh_reasoning_effort: true but are missing supports_max_reasoning_effort: true. The complete fix is a two-part change:

1. Add "supports_max_reasoning_effort": true to the Opus 4.7 entries in model_prices_and_context_window.json (and its bedrock/azure_ai/regional variants).
2. Replace the expanded substring check with the existing data-driven helper:

if effort == "max" and not self._supports_effort_level(model, "max"):
    raise ValueError(
        f"effort='max' is only supported by Claude Opus 4.6 and 4.7. "
        f"Got model: {model}"
    )

This makes enabling max for any future model a pure JSON change—exactly the pattern the project enforces.

Rule Used: What: Do not hardcode model-specific flags in the ... (source)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[shang309073819]

Note on the failing lint check:

After the chore: refresh uv.lock commit, the uv lock --check step now passes, but the next step (black --check --exclude '/enterprise/' .) fails on 331 pre-existing files in litellm/ that haven't been formatted with the pinned black==24.10.0 (they're stuck on an older style). None of these files are touched by this PR — only litellm/llms/anthropic/chat/transformation.py is, and within it only lines 1534-1545 (the effort == "max" guard).

Concretely, black --diff against my changed file shows reformat hunks at lines 1015, 1122, 1198, 1224, 1573, 1869 — all in code that predates this PR and is unchanged here.

The reason previous PRs against litellm_internal_staging haven't surfaced this is that the lint workflow was already failing earlier, on uv lock --check, and exiting before reaching the black step (e.g. run #24567699793 exits at the lockfile step). Now that I've refreshed the lockfile, the workflow proceeds and exposes the pre-existing black drift.

Happy to either:

1. Leave it as-is and let the maintainer merge with the lint job overridden (since the failure isn't from this PR), or
2. Send a separate chore(format): apply black 24.10.0 across litellm/ PR (~331 files, mechanical, no logic changes) before this one.

Let me know which you'd prefer. The actual bugfix in this PR is small and self-contained, and unit-test, core-utils, all proxy jobs, Vertex, etc. are passing.

[congqiao]

Claude Sonnet 4.6 also supports max. https://platform.claude.com/docs/en/build-with-claude/effort#effort-levels

[shang309073819]

Thanks both — addressed in 532c47a.

@greptile-apps (P2): swapped the hardcoded _is_opus_4_6_model OR _is_opus_4_7_model substring guard for the same data-driven _supports_effort_level(model, "max") pattern the xhigh branch one line below already uses. Adding support for a new model is now a pure model-map change.

@congqiao: good catch — confirmed against Anthropic's effort docs¹ that max is supported on Opus 4.6, Opus 4.7, and Sonnet 4.6. Added supports_max_reasoning_effort: true to:

• the 10 Opus 4.7 entries (anthropic / bedrock_converse {global,us,eu,au} / azure_ai / vertex_ai)
• the 9 Sonnet 4.6 entries (same providers)

Plumbed supports_max_reasoning_effort through ModelInfoBase / ModelInfo so get_model_info() surfaces it (mirrors the existing xhigh flag).

Tests:

• flipped test_max_effort_rejected_for_sonnet_46 → test_max_effort_accepted_for_sonnet_46
• updated opus-4.5 rejection test to match the new error message (effort='max' is not supported by this model)
• repointed the opus-4.6 acceptance test at a SKU that exists in the map (claude-opus-4-6-20260205)

All 137 tests in test_anthropic_chat_transformation.py pass locally.

Footnotes

1. https://platform.claude.com/docs/en/build-with-claude/effort — "The effort parameter is supported by ... Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6." ↩

[shang309073819]

Rebased onto latest litellm_internal_staging to clear the uv.lock conflict from PRs #25872 / #25873 (yj_bump_apr16_2 + extras bump). The conflict was purely metadata in uv.lock (literally one line — litellm == 1.83.9 arrived from upstream while we had 1.83.7 pinned earlier). No code-level conflicts in any of the changed files.

Rebase summary:

• Force-pushed fix/effort-max-opus-4-7 (now at fa541915) onto current upstream tip f69b9d65.
• All 4 commits preserved (fix → uv.lock refresh → data-driven refactor → LITELLM_LOCAL_MODEL_COST_MAP test isolation).
• All 137 tests in tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py still pass locally (LITELLM_LOCAL_MODEL_COST_MAP=True).
• The 27 effort/max-related tests pass independently.

Re: the lint check — same situation as before (the black 24.10.0 drift on 331 pre-existing files surfaces because the uv.lock step now passes). Happy to file a separate chore(format) PR if you'd prefer that ordering. The actual fix in this PR is unchanged and self-contained.

mergeable should now flip to MERGEABLE. Ready for review whenever you have a moment 🙏

[shang309073819]

Rebased onto latest litellm_internal_staging (2f22a1293e — bump litellm-proxy-extras to 0.4.67) to clear the merge conflict from the 1.83.9 → 1.83.10 bump (PR #26043 / yj_bump_apr18).

Rebase summary:

• Force-pushed fix/effort-max-opus-4-7 (now at cb38100ea3) onto current upstream tip.
• The standalone chore: refresh uv.lock for litellm 1.83.9 / extras 0.4.66 commit was dropped during rebase (auto-skipped) since upstream already carries an equivalent or newer uv.lock for 1.83.10 / extras 0.4.67. The 3 substantive commits are preserved:
  • de196e4c41 — fix: allow effort='max' on Claude Opus 4.7
  • 364a556062 — refactor: make effort='max' data-driven via model map
  • cb38100ea3 — test: force LITELLM_LOCAL_MODEL_COST_MAP for new max-effort tests
• All 137 tests in tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py still pass locally with LITELLM_LOCAL_MODEL_COST_MAP=True.

Re: the previous All Other Providers / Run tests failure on the prior push — that was caused by a leftover Git merge-conflict marker (<<<<<<< worktree-rustling-wishing-kite) at tests/test_litellm/llms/bedrock/test_bedrock_common_utils.py:37 that landed on litellm_internal_staging independently of this PR. It is already cleaned up on the current upstream tip, so this run should pass.

Re: the lint check — same pre-existing situation as before: uv lock --check is now satisfied (we inherit upstream's lock), and black --check will continue to flag the same ~331 pre-existing files in litellm/ that this PR does not touch. Happy to send a separate chore(format): apply black 24.10.0 across litellm/ PR if a maintainer prefers that route.

PR is otherwise green: unit-test, core-utils, all proxy-*, vertex-ai, mcp-secrets-misc, responses-caching-types, Greptile, Veria, Codecov, GitGuardian, CLA — all passing.

@krrishdholakia @ishaan-jaff — could one of you take a look when you have a moment? The bug is small (just lets effort="max" through for Opus 4.7, which the Anthropic API accepts) and the fix is fully data-driven now. Thanks!

[codgician]

Any updates on this PR?

[github-actions]

This pull request has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.