> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ai-stats.phaseo.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Anthropic: Claude Opus 4.6 to 4.7

> Model migration notes for upgrading Anthropic Claude Opus 4.6 traffic to Claude Opus 4.7.

# Anthropic: Claude Opus 4.6 to 4.7

Use this guide when moving from `anthropic/claude-opus-4.6` to `anthropic/claude-opus-4.7`.

Claude Opus 4.7 introduces request-shape changes that can hard-fail old 4.6 patterns if left unchanged.

## What changed

* New effort level: `xhigh` (Messages API)
* Extended thinking budgets were removed for Opus 4.7 (`thinking: { type: "enabled", budget_tokens: ... }` now fails)
* Sampling params are no longer supported on this model when set to non-default values (`temperature`, `top_p`, `top_k`)
* Thinking content is omitted by default unless you explicitly opt in
* Updated tokenizer behavior can increase token usage vs Opus 4.6

## AI Stats gateway compatibility behavior

To reduce migration breakage, AI Stats accepts legacy Opus 4.6 fields for Opus 4.7 and normalizes them before sending upstream:

* `temperature`, `top_p`, and `top_k` are accepted but stripped
* legacy thinking budgets are not forwarded
* Opus 4.7 thinking is always sent as `thinking: { "type": "adaptive", "display": "summarized" }`
* `reasoning.effort` is still honored and mapped to `output_config.effort` (`xhigh` supported)

## What to change in your integration

### 1. Update model ID

Move from:

* `anthropic/claude-opus-4.6`

to:

* `anthropic/claude-opus-4.7`

### 2. Move to adaptive thinking with summarized display

For Opus 4.7, use:

* `thinking: { "type": "adaptive", "display": "summarized" }`

Avoid legacy extended-thinking budgets on this model.

### 3. Remove sampling controls

Do not send:

* `temperature`
* `top_p`
* `top_k`

Prefer prompt and instruction tuning instead.

### 4. Rebaseline effort controls

If you use reasoning controls, retest with:

* `output_config.effort = "high"` as baseline
* `output_config.effort = "xhigh"` for hardest coding/agentic paths

If you call AI Stats through OpenAI-style payloads, `reasoning.effort` maps to this same `output_config.effort` value for Opus 4.7.

### 5. Recheck token ceilings

Because tokenization changed, audit:

* `max_tokens` / `max_output_tokens`
* timeout budgets
* cost controls and alerts

## What to test

* strict schema output pass rate with your production prompts
* long-running tool workflows (tool call count, completion rate, latency)
* response quality deltas at `high` vs `xhigh`
* token-cost deltas on your highest-volume prompt classes

## Safe rollout

1. Deploy request-shape changes before shifting traffic.
2. Shadow Opus 4.7 on your production eval suite.
3. Canary with rollback to Opus 4.6.
4. Promote once quality, latency, and cost deltas are within target.

## Sources

* [What's new in Claude Opus 4.7](https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7)
* [Anthropic migration guide](https://platform.claude.com/docs/en/about-claude/models/migration-guide)
* [Claude Opus 4.7 System Card (PDF)](https://cdn.sanity.io/files/4zrzovbb/website/037f06850df7fbe871e206dad004c3db5fd50340.pdf)