Opus 4.7 Playbook

Ship-ready migration kit for Anthropic's new Opus 4.7 model — breaking changes, new effort levels, tokenizer re-budgeting, task budgets, and drop-in config templates. Works with direct SDK code and with Claude Code environments.

Opus 4.7 went GA on 2026-04-16. Five API params return HTTP 400 on 4.7 that were fine on 4.6. One silent default change will make your streaming UI look hung. A new tokenizer quietly uses up to 35% more tokens per request. The playbook is everything you need to ship the migration in an afternoon.

TL;DR

• temperature, top_p, top_k, budget_tokens, and assistant prefill all return 400 now. Remove them.
• Replace thinking: {type: "enabled", budget_tokens: N} with thinking: {type: "adaptive"}. Depth is controlled by output_config.effort — new levels low / medium / high / xhigh / max.
• If your UI streams reasoning, set thinking.display: "summarized" — the default just became "omitted".
• Raise max_tokens by at least 35% to absorb the new tokenizer. Minimum 64,000 at xhigh / max.
• xhigh is the new default for coding and agentic work. max often overthinks structured outputs.
• Task budgets (beta) let the model pace itself across a full agentic loop. Messages API only. Opt-in.

Full breaking-change detail: docs/01-whats-new.md. Full checklist: docs/02-migration-checklist.md.

The 30-second migration

# Before (4.6)
client.messages.create(
    model="claude-opus-4-6",
    max_tokens=32_000,
    temperature=0.7,                                 # remove
    thinking={"type": "enabled", "budget_tokens": 8_000},   # replace
    messages=[...],
)

# After (4.7)
client.messages.create(
    model="claude-opus-4-7",
    max_tokens=64_000,                               # +35% headroom
    thinking={"type": "adaptive"},                   # no budget_tokens
    output_config={"effort": "xhigh"},               # depth control
    messages=[...],
)

That is the minimum diff. It is also sufficient for most codebases. See examples/minimal_call.py for the runnable version and examples/agentic_loop.py for the tool-calling pattern.

Quick install (Claude Code users)

If you run Claude Code and want the operating notes, CARL rules, and global pointer in your environment:

git clone https://github.com/7alexhale5-rgb/opus-4-7-playbook
cd opus-4-7-playbook
bash migration/install.sh     # interactive — nothing installs without a prompt

What it offers (each step is opt-in):

1. Drops templates/operating-notes.md into ~/.claude/references/
2. Installs the templates/carl-opus-4-7 CARL domain (if you use CARL)
3. Appends the templates/CLAUDE-md-global-snippet.md pointer to your global CLAUDE.md

For per-project config, copy templates/CLAUDE-md-project-snippet.md into each project's CLAUDE.md.

If you don't run Claude Code, the docs are still the fastest migration reference on the internet — start with docs/02-migration-checklist.md and work top-to-bottom.

What's in this repo

Path	What
docs/	Migration checklist, effort-level rubric, adaptive thinking, tokenizer re-budgeting, task budgets, vision at 3.75 MP
templates/	Copy-paste config for Claude Code: operating notes, CARL domain, global + project CLAUDE.md snippets, agent-pinning reference
examples/	Runnable Python — minimal call, agentic loop, task-budget wrapper, full-resolution vision
migration/	Shell scripts — pin agent model IDs, scan a repo for deprecated params, interactive installer
social/	Instagram carousel (Markdown + social/export/ PNGs), Twitter thread, LinkedIn post
tests/	CI-enforced sanitization, code-syntax, and link checks

Who this is for

• Engineers with an Anthropic SDK app that needs to move to 4.7 this week
• Claude Code users who want their subagents and skills aligned to the new model's strengths
• Agents / builders who want a spec-complete reference instead of reading five docs pages to find one API shape

Looking for the full env?

This repo is the focused migration kit. For the complete Claude Code setup it's extracted from — agents, CARL policy engine, 20 lifecycle hooks, 78-skill catalog, MCP server list, drop-in installer — see the umbrella repo:

gravity-stack — top 1% AI-native development environment.

Both repos are MIT. Both pass sanitization in CI. Use one, the other, or both.

Sources

Every factual claim in the docs cites Anthropic's primary sources:

• Migration guide
• Model overview
• Adaptive thinking
• Effort
• Task budgets
• Release post

Contributing

PRs welcome. Two hard rules:

1. bash tests/run_all.sh must pass locally before you open a PR.
2. No personal / proprietary references — the sanitization test blocks them in CI.

License

MIT. See LICENSE.

---

Built by Alex Hale / PrettyFly. If this kit saves you a day, say hi.