You spent weeks tuning your setup. Your CLAUDE.md — a rules file that tells the AI how to behave in your project — is perfect. Your prompt patterns are dialed in. Your AI coding assistant finally writes code the way you want it. Life is good.

Then on April 16, Anthropic swapped the brain.

The overnight lobotomy

Claude Opus 4.7 went live on April 16, 2026, across Claude.ai, the API (a way for programs to talk to Anthropic's servers), AWS Bedrock, Google Vertex, and GitHub Copilot. If you use Claude Code — Anthropic's terminal-based coding agent, which scored a 91% CSAT in JetBrains' AI Pulse survey — and your config says model: opus, your tool silently started running a different model. No notification. No opt-in dialog. No changelog in your face.

GitHub made it even clearer. Their April 16 changelog says Opus 4.7 will replace Opus 4.5 and 4.6 in the Copilot model picker "over coming weeks." Not "alongside." Replace.

And on April 23 — two days from now — Claude Code auto-switches the default for Enterprise and API users too.

Every dependency has a lockfile. Except the most important one.

Think about how software normally works. When you use a package from npm (JavaScript's app store for code libraries) or a Docker image (a frozen snapshot of a server environment), you pin the exact version. There's a lockfile — a receipt that says "I'm using version 2.3.1, and I chose this on purpose." There's a changelog — a list of what changed. There's git blame — a way to see who changed what and when.

AI models have none of this.

According to Anthropic's own documentation, Opus 4.7 introduced breaking API changes: Anthropic removed thinking budgets (400 errors if you try), killed sampling parameters, and swapped in a new tokenizer — the component that chops text into chunks the AI can read — that makes the same input cost up to 35% more tokens. Simon Willison independently measured system prompts costing 1.46× more and images costing 3.01× more.

But the behavioral changes are worse than the breaking ones, because they're invisible. Anthropic describes Opus 4.7 as "more literal instruction following" with a "more direct, opinionated tone." Fewer tool calls by default. Fewer subagents. Different response lengths. In other words: the model thinks differently — and you optimized your carefully crafted rules file for a brain that no longer exists.

Tian Pan, an ex-Uber engineer, published "The Silent Regression" on April 17, nailing the core problem: "Power users are the ones most harmed by behavioral drift...who invested most in understanding the system's quirks."

What "better" actually means

Opus 4.7 scores 87.6% on SWE-bench Verified — a standardized test for AI coding ability. Impressive on paper. But "better on average benchmarks" is not "better for your codebase." Your project isn't a benchmark. Your conventions aren't average.

Cursor offers a model picker but no version pin within a model family. Copilot pushes replacements. Claude Code does let power users pin via environment variables (ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-4-7), but subscription users on Pro and Max tiers — the majority — get moved silently. The Augment Code CTO checklist calls tools without version pinning an "immediate disqualification" for enterprise adoption.

What to do right now

Treat model updates like infrastructure changes. Version-control your rules files. If you're on Claude Code, pin the full model name instead of the opus alias. Keep a small set of "prompt tests" — known inputs where you expect specific outputs — and re-run them after any model change. And demand behavioral changelogs from your vendor, not just benchmark tables.

The punchline

Remember that perfectly tuned setup you had last week? You tuned it for a model that's already being deprecated. The most impactful component in your entire coding toolchain — the AI model — is the only one with zero change management. The first vendor to ship version pinning, behavioral diffs, and rollback as first-class features won't just win developer trust. They'll win the enterprise tier, where "it changed and we don't know why" isn't a feature — it's a compliance violation.