docs(blog): fix Opus 4.6/4.7/4.8 thinking mode and effort guidance

[mateo-berri]

Summary

The day-0 blog posts for Claude Opus 4.6, 4.7, and 4.8 carried a copy-pasted "Adaptive Thinking" note that told readers to pass the native thinking parameter directly to use explicit thinking budgets with type: "enabled". That guidance is wrong for Opus 4.7 and 4.8.

Per Anthropic's adaptive thinking and effort docs, adaptive thinking is the only supported mode on Opus 4.7 and 4.8, and a manual thinking: {type: "enabled", budget_tokens: N} is rejected with a 400 error. LiteLLM forwards the native thinking parameter unchanged on both /chat/completions and the /v1/messages passthrough, so following the old instructions just produces that 400. On Opus 4.6 the explicit budget still works but is deprecated and no longer recommended. Each note now states the actual behavior and points readers at output_config.effort, paired with adaptive thinking, as the way to control thinking depth.

Effort ladder correction

While verifying the above, the 4.7 post claimed max effort was "Claude Opus 4.6 only and is not available on 4.7." Anthropic's effort docs list max as available on Opus 4.6, 4.7, and 4.8, and LiteLLM's REASONING_EFFORT_TO_OUTPUT_CONFIG_EFFORT maps max to max with no per-model gating. So the 4.7 effort section now lists all five levels (low, medium, high, xhigh, max), the 4.8 "max (previously Opus 4.6 only)" parenthetical is corrected to "also available on Opus 4.6 and 4.7," and the reasoning_effort enumeration in each Adaptive Thinking note now includes max so it matches that model's own effort section. The 4.6 note gains max but not xhigh, since 4.6 never supported xhigh.

Verification

Cross-checked against Anthropic's official docs (the adaptive thinking and effort pages) and the LiteLLM source. _map_reasoning_effort short-circuits every effort value to thinking: {type: "adaptive"} for adaptive-thinking models, the thinking parameter is passed through verbatim, and the effort mapping applies no per-model max restriction. These are prose-only changes to three blog markdown files, with no code or behavior changes.

Type

Documentation

---

Note

Low Risk
Markdown-only documentation corrections with no code or configuration changes.

Overview
Corrects Adaptive Thinking and Effort Levels guidance in the Opus 4.6, 4.7, and 4.8 day-0 blog posts so it matches Anthropic’s current API behavior and LiteLLM’s passthrough.

The shared note no longer tells readers to use native thinking: {type: "enabled", budget_tokens: ...} as the primary path. Opus 4.7 and 4.8 are documented as adaptive-only (explicit budgets → 400); depth should be tuned with output_config.effort alongside adaptive thinking. Opus 4.6 keeps explicit budgets as still accepted but deprecated, with the same effort-based recommendation.

Effort ladder fixes: reasoning_effort lists now include max where appropriate (4.6 adds max, not xhigh; 4.7/4.8 include max). 4.7 drops the wrong claim that max is 4.6-only and documents five levels including max, with updated guidance on xhigh vs max. 4.8 fixes the parenthetical that max was “previously Opus 4.6 only.”

Prose-only; no runtime or gateway behavior changes.

^{Reviewed by Cursor Bugbot for commit c1d5fb6. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	May 29, 2026 5:46pm

[mateo-berri]

bugbot run

[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 c1d5fb6. Configure here.}