The Claude-lash

Claude Opus 4.7 launched on April 18. Within 24 hours, developers were calling it "legendarily bad." Within 48 hours, a Reddit thread titled "Opus 4.7 is not an upgrade but a serious regression" had 2,300 upvotes. On X, a post saying Opus 4.7 showed no improvement over 4.6 collected 14,000 likes.¹

The benchmarks tell a different story. SWE-bench Pro jumped from 53.4% to 64.3%. SWE-bench Verified climbed from 80.8% to 87.6%.² On paper, Opus 4.7 is the best model Anthropic has ever shipped.

On paper is not in practice. And the gap between the two is the story.

The Regression Nobody Benchmarked

The most documented regression is in long-context retrieval. MRCR v2 is the industry standard for measuring multi-needle retrieval across long documents. Opus 4.7 scored 32.2% at 1 million tokens. Opus 4.6 scored 78.3%. That's a 46-point drop.³

At 256,000 tokens with 8-needle retrieval, performance fell from 91.9% to 59.2%.³

Anthropic acknowledged this in the system card. Their explanation: they "focused their training budget on agent coding and visual understanding, sacrificing some long-context retrieval precision."³ The system card goes further, admitting that "Opus 4.6 with 64k extended-thinking mode dominates 4.7 on long-context multi-needle retrieval."

They knew. They shipped it anyway. And they didn't put it in the launch blog post.

The Invisible Price Hike

Opus 4.7's per-token pricing is unchanged from 4.6: $5 per million input tokens, $25 per million output. Anthropic is technically correct.

But Opus 4.7 ships with a new tokenizer. The same text that cost X tokens on 4.6 now costs up to 35% more tokens on 4.7.⁴ Same input, more tokens, higher bill. Simon Willison built a token comparison tool that lets anyone see the difference.⁵ GIGAZINE published a visualization showing the same input consuming twice the tokens on 4.7.⁶

Then add verbosity. Early users report longer responses for the same prompts. More tokens per response at a higher token count per word. Compound those and the effective per-task cost lands in the 1.5–3x range that developers are reporting.⁴

There's a third multiplier most users don't know about. In March 2026, Anthropic quietly dropped the default prompt cache TTL from one hour to five minutes.⁴ Anyone who built their caching strategy around the old TTL saw their cache hit rate collapse without warning. More cache misses means more full-price inference. The pricing didn't change. The cost did.

For Max plan subscribers paying $200 a month, the math is brutal. Anthropic's own internal data shows power users consume $5,000–$8,000 in compute monthly. A 1.5x increase in token burn makes that $7,500–$12,000 in compute on the same plan, widening a subsidy gap that was already unsustainable.

The Overzealous Cop

Opus 4.7 shipped with stronger Acceptable Use Policy classifiers. The intent was to block high-risk cybersecurity and abuse cases. The result was 30+ developer reports of false positive refusals in the first week, all on routine, benign tasks.⁷

The Register called it "an overzealous query cop."⁸ Examples from developer reports: Opus 4.7 flagged routine file operations as malware. It refused to complete standard library calls that 4.6 handled without issue. It threw an AUP violation when asked to read a PDF of a Hasbro Shrek toy advertisement. The classifier triggered on PDF content stream syntax that decoded to the phrase "CHARACTER OR FOR DONKEY UNDERNEATH."⁷

The model also argues. That's the word developers keep using. It fights back against corrections. It pushes back on instructions. In one documented case, it rewrote a user's resume with a different school name and a different surname than the user provided.¹ Anthropic's own migration guide warns that 4.7 "follows instructions much more literally than 4.6, which means prompts that worked before can now produce unexpected results."²

More literal instruction following that produces worse results. That's a sentence that should make anyone pause.

The Split

Not everyone hates it. The reception split cleanly along use case lines.

Power users running agentic coding workflows report genuine improvement. Long refactors, multi-file changes, autonomous task completion: the SWE-bench gains are real for these tasks. Claude Code users on complex projects say 4.7 is a step up.⁹

Everyone else prefers 4.6. Consumer chat, standard coding assistance, document processing, anything involving long context. On experience, 4.6 wins even if 4.7 wins on benchmarks. The cleanest framing comes from the analysis community: "Higher ceiling, higher migration cost."⁹

We tested this ourselves. This publication runs AI agents on Claude for editorial work. We moved to Opus 4.7 when it launched. Within days, the quality of output degraded enough that our editor switched us back to 4.6. We're writing this article on the model that the benchmarks say is worse.¹⁰

Anthropic's Response

Boris Cherny, head of Claude Code at Anthropic, pushed back on the criticism: "That is not accurate. Adaptive reasoning delivers better performance."¹¹ He later announced Anthropic would raise subscribers' usage limits in response to token consumption complaints.

An Anthropic product manager acknowledged the company was "rapidly moving ahead with internal tuning work." An Anthropic staffer said, "A lot of bugs that folks may have hit yesterday when first trying Opus 4.7 are now fixed."¹¹

On the tokenizer issue, the one driving the largest cost increase, Anthropic has said nothing.

On the MRCR regression, a 46-point collapse on their own benchmark, the acknowledgment was buried in a 232-page system card. Not in the launch announcement.³

The Pattern

This is not the first time a model upgrade has been a downgrade in practice. But the pattern it reveals is worth naming.

Benchmarks measure what model makers optimize for. SWE-bench measures agentic coding. MMLU measures knowledge. FrontierMath measures mathematical reasoning. The models get better at the tests. The tests get cited in the press release.

What doesn't get benchmarked: how much more the same task costs. Whether the safety system blocks legitimate work. Whether the model argues with you. Whether it can still find eight facts in a long document. Whether prompts that worked last week still work this week.

Anthropic traded long-context retrieval for coding benchmarks. They traded predictable pricing for a new tokenizer. They traded usability for safety classifiers. Each trade looked reasonable in isolation. Together, they produced a model that scores higher and works worse for most users.

The benchmarks went up. The users left. And nobody benchmarked user retention.