Why Opus 4.7 Felt Dumber, and What Anthropic Actually Fixed

For about a month, every serious builder I follow kept saying the same thing: Claude Opus 4.7, Anthropic's newest flagship, had gotten dumber. On April 16, 2026, developer Simon Willison, whose AI model benchmarks the community treats as a signal, ran his pelican-on-a-bicycle test and watched a quantized Qwen (a compressed open-weights model from Alibaba running on his laptop) draw a cleaner bicycle frame than Opus 4.7.^[4] A Reddit thread titled “Opus 4.7 is not an upgrade but a serious regression” climbed past 2,000 upvotes.^[8] A GitHub issue filed the same day described a month of quiet, daily-driver degradation across custom skills and long-running agents.^[7] On April 23, Anthropic's Claude Code lead Boris Cherny acknowledged it in two sentences on Threads: “Over the past month, some of you reported Claude Code's quality had slipped. We investigated, and published a post-mortem.”^[2]

Anthropic's postmortem is the interesting document.^[1] Not because it apologizes, though it does. Because of where the three bugs lived.

What Anthropic Actually Fixed

Three changes, all in the Claude Code harness, all compounding over roughly six weeks. None of them touched the weights.

The third one is the most instructive. A 25-word cap on pre-tool-call reasoning sounds harmless. It was the kind of instruction-level tweak you'd write in an afternoon and ship behind a feature flag. It measured as a 3% drop on an internal coding eval, borderline noise in any A/B test. It shipped with the biggest model release of the quarter. Users noticed.

Why the Evals Missed All Three

The same reason the previous quality postmortem missed things. Anthropic's evaluation suite tests the model. The bugs lived in the harness. A benchmark that grades Opus 4.7 on a fixed set of prompts sees the weights; it doesn't see what the Claude Code client prepends, caches, or truncates before those prompts reach the weights. When the system prompt changed on April 16, the model evals kept showing normal numbers. The users running agentic coding sessions through Claude Code saw a different product.

Anthropic said this plainly in their last quality postmortem, eight months ago, about an entirely different set of bugs.^[3] Different class of failure then. Same admission.

The Second Postmortem in Eight Months

The September 2025 postmortem covered three bugs of a different kind. A context-window routing error that ran from August 5 to September 18, silently sending some requests to the wrong model. Output corruption on a batch of Google TPU servers. A miscompilation in XLA (the compiler Google uses for TPU code) that hurt sampling quality. All infrastructure. Every one was found because users complained loudly enough that Anthropic went looking. The April 2026 postmortem just did it again, with a different class of bug: harness and system prompts instead of hardware and inference. What's common is the loop. Community reports a regression, Anthropic's internal evals say everything is fine, eventually someone at Anthropic goes digging and finds something real.

I don't think this is negligence. A frontier model is a moving target. Anthropic ships model updates, harness updates, infrastructure updates, prompt changes, and tool-use changes on overlapping cadences. Every one of those layers can regress independently, and the eval suite that covers all of them doesn't exist yet at any lab. Credit Anthropic for writing the failure down, twice, in useful detail. Less comfortable: both times, it took a month of public complaint to get there.

Community-sourced regression reports beat internal evals twice now.

If You're Building on Claude

Here's what matters if you ship on Claude. If you're on Claude Code, upgrade to v2.1.116 or later and all three regressions are gone. If you call the Anthropic API directly, you were never affected. None of the three bugs touched the API path. The model card does document real tradeoffs in Opus 4.7 vs 4.6 on some benchmarks (long-context retrieval regressed, software-engineering scores improved), and Willison's pelican suggests the weights have taste regressions on certain creative work too.^[6] Those are weight-level tradeoffs Anthropic is upfront about in the model card. The Claude Code cliff was separate.

The broader lesson for anyone depending on a hosted frontier model: treat the model plus its harness as the product, because that's what ships. Keep a canary set of prompts you run against your actual deployment weekly. The lab's eval suite doesn't watch its own harness. You'll catch changes first. It's roughly the same lesson the Claude outage pattern has been teaching since last fall: the product you rely on is a stack, and the stack moves under you.

The eval suite watches the model. For eight months, the users have been catching what it misses.