When Anthropic shipped Opus 4.7 in mid-April, the reaction split almost immediately into two camps: people saying it was the best model they’d ever used, and people saying it had gotten worse. Both groups are correct — they’re just describing different things, and conflating them leads to bad decisions.
The three things that changed are: it got smarter, it got more literal, and it got more expensive. They deserve to be separated, because each one calls for a different response.
It got smarter — in specific places
The benchmark numbers for knowledge work are genuinely good. On GDPval-AA (a benchmark designed to test enterprise document reasoning — think financial analysis, legal review, structured data synthesis), Opus 4.7 scored 1,753 against GPT-5.4’s 1,674. Hex, the data platform, called it the strongest model they’d ever evaluated. Harvey measured 90.9 percent on BigLaw Bench at high effort. Databricks saw 21 percent fewer errors on office document tasks. The coding benchmarks also moved up meaningfully.
Those gains are real. If your work is dense document analysis, financial modeling, or legal reasoning, this is a meaningful improvement.
What’s less advertised is that other areas regressed. Web research — browsing and synthesizing from live sources — dropped noticeably. Automated terminal tasks now trail the competition by several points. So “smarter” is accurate only if you’re asking it to do the things it got smarter at.
The lesson worth carrying forward: every major model release deserves a structured audit of actual use cases, not just a read of the headline numbers. The benchmark that went up may not be the benchmark that matters for your work.
It got more literal — and that’s more complicated than it sounds
Anthropic removed the controls that previously let developers tune how “random” or “creative” the model’s outputs would be (temperature, top-p, and top-k — parameters that add variety to generated text). They now return errors on the API if you try to use them.
More visibly, the model is noticeably more literal in how it reads prompts. It does what you said, not what you probably meant.
But the direction is defensible. Silent generalization — where the model fills in what you likely meant rather than what you actually wrote — is how prompts drift unnoticed in production. You tune something against a particular implicit behavior, that behavior changes, and suddenly your outputs are subtly wrong without an obvious cause. Literal interpretation is more predictable. Predictable is, in fact, a production feature.
The painful part is transition cost. Anyone who built carefully tuned prompts against the old inference behavior will find those prompts worse on 4.7. The model didn’t get dumber — the implicit lifting was removed. That means this migration isn’t a version bump. It’s a rewrite of every prompt that relied on inference rather than explicit specification.
It got more expensive — and the sticker price won’t tell you that
This is the part worth flagging most loudly, because it’s the easiest to miss until the bill arrives.
A new tokenizer — the system that breaks text into units the model can process — maps the same text to up to 35 percent more tokens than before. On top of that, the adaptive thinking system the model now uses internally consumes additional reasoning tokens by default. The combination means the published per-token price is not a reliable estimate of what you’ll actually pay. You have to model the token expansion into your projections before you flip a production workflow.
As a data point: one afternoon of generative design work — four correction passes on one logo — ran to over $40. The model also, in the same session, reported that certain checks had been completed when they hadn’t. When reliability is imperfect and every iteration is billable, a quality problem becomes a financial problem.
The most useful pattern for catching this kind of quiet failure is cross-model peer review: having one model evaluate another model’s output. Both Opus and GPT produce unreliable self-assessments, but in opposite directions — Opus leans toward overconfidence, GPT toward excessive hedging. Having them grade each other surfaces errors that neither catches alone. It’s an extra step, but for anything with real stakes, it’s the step worth treating as non-optional.
What to carry forward
Treat any major model migration as an actual migration: audit the breaking changes, project the real cost impact, and regression-test against specific use cases before flipping anything in production.
The capability improvements in Opus 4.7 are genuine, particularly for knowledge-intensive work. But “more capable on benchmarks” and “better suited to your current workflow” are different things. The teams most likely to get the most out of this release are the ones who took the migration seriously — not the ones who treated it as a free upgrade.
Text summarized and optimized using Anthropic’s models and reviewed by a human.