Opus 4.7 — What Changed and How to Use It
Claude Opus 4.7 shipped on 2026-04-16. What's different, what's better, and Anthropic's official guidance on prompting it differently to unlock the upgrades.
Claude Opus 4.7 shipped on April 16, 2026. Same pricing as 4.6: $5 per million input tokens, $25 per million output. What changed is everything around the model, including how you should prompt it.
Vision: 4.7 processes images up to 2,576 pixels on the long edge, more than three times what prior Claude models could handle. A full Stripe dashboard screenshot or a dense architecture diagram no longer needs cropping. Claude sees the detail on the first pass.
Coding: 4.7 solves 3x more tasks than 4.6 on real software engineering benchmarks. Code review recall is up 10%. Financial document reasoning makes 21% fewer errors.
The change that matters most isn't in the benchmarks: 4.7 verifies its own outputs before reporting back. Fewer "are you sure that works?" cycles.
That shifts how you prompt more than what you can ask for. Anthropic's official guidance: treat 4.7 as a delegated engineer, not a pair programmer.
Don't guide it line by line. Give it full context upfront: intent, constraints, acceptance criteria, which files to touch. Batch your questions. Let it run. Every back-and-forth turn adds reasoning overhead. 4.7 rewards you for specifying more upfront and getting out of the way.
To support that shift, 4.7 ships a new default effort level: xhigh, between high and max. Claude Code users who hadn't set an effort level were auto-upgraded. Use max only for genuinely hard problems (prone to overthinking) and high for concurrent sessions where cost matters. Even medium and low on 4.7 outperform the same levels on 4.6.
Extended Thinking with a fixed budget is gone, replaced by adaptive thinking: the model decides when to think hard. You steer with words, not dials. More thinking: "Think carefully and step-by-step; this problem is harder than it looks." Less: "Prioritize responding quickly. When in doubt, respond directly."
Three quiet behavior changes will break your 4.6 prompts if you don't know about them:
Less verbose by default. If you need a specific length or voice, say so. Positive examples ("write like this") work better than negative instructions.
Calls tools less. If your workflow depends on aggressive tool use, spell out when and why explicitly.
Spawns fewer subagents. If you want parallel fan-out, tell it directly: "Spawn multiple subagents when fanning out across items or reading multiple files."
Claude Code got sharper too. New slash command /ultrareview runs a deeper bug-detection pass; Pro and Max users get three free sessions. Auto Mode expanded to Claude Code Max in research preview (Shift+Tab to toggle). Task budgets are in public beta on the API: actual cost caps on long agent runs.
One cost note: the new tokenizer can count the same input as 1.0 to 1.35x more tokens, with code and structured data at the high end. Per-token pricing held, but measure your real workload before assuming the bill is a wash. Boris Cherny, who leads Claude Code, took a few days to re-learn 4.7; your tuned 4.6 prompts are probably leaving capability on the table.
What changed: vision 3x bigger, 3x more coding tasks solved, self-verification, new xhigh default, adaptive thinking, /ultrareview, Auto Mode, task budgets. Less verbose, fewer tools, fewer subagents by default. Same price per token, but watch volume with the new tokenizer.
How to use it: treat it like a delegated engineer. Specify upfront, let it run, steer adaptive thinking with words, and spell out when to use tools or spawn subagents. The deeper playbook lives in Tracks 2 through 5.