Claude Opus 5 Tops the Benchmarks While Early Testers Call It “Hard to Love”

Anthropic released Claude Opus 5 on Friday afternoon. By that evening it was top of the Artificial Analysis intelligence leaderboard. The people who had been testing it for a week before launch spent that same evening explaining why they don’t enjoy using it.

That gap is the story. Opus 5 costs $5 per million input tokens and $25 per million output, the same as Opus 4.8. Anthropic says it lands within 0.5% of Fable 5 on CursorBench at half the cost per task. Those numbers hold up. So do the complaints, and they come from the same reviewers who were praising Anthropic’s last model a month ago.

Claire Vo built ChatPRD and hosts the How I AI podcast. She opened her day-zero review by saying she hates working with it, then said she ranked it above every other model in a blind test, above Fable 5 and above GPT-5.6. Her review is titled “Claude Opus 5 review: this model is brilliant (but annoying)”. It is a seven-model, six-task blind scoring run. Her notes single out the model’s verbosity and what she calls its neurotic personality, including a merge conflict it refused to touch.

What are the testers complaining about?

Dan Shipper and Katie Parrott at Every ran Opus 5 through coding, writing, knowledge work and their internal agent for a week before launch. Their verdict went up as “Vibe Check: Claude Opus 5 Is Brilliant in Flashes, Frustrating in Practice”.

“It’s a poor man’s Fable. It has many of Fable’s personality quirks without Fable’s genius.”

Dan Shipper, Every

The complaints are about behaviour, not capability. It talks back, it stops early, and it ignores scaffolding that worked fine on Opus 4.8. Anthropic’s own prompting guide admits one piece of this: Opus 5’s default responses run longer than previous Opus models. For a release sold on efficiency, that is an odd thing to ship.

So why did it still finish first?

Because on the tests, it does finish first. Artificial Analysis put it at the top of their index and called it the new leader in agentic knowledge work, with 1861 Elo on GDPval-AA v2, more than 100 points clear of both Fable 5 and GPT-5.6 Sol.

Two caveats before you weigh that. Artificial Analysis discloses that Anthropic paid them to evaluate the model ahead of release. And their index runs used Opus 4.8 as a fallback.

Independent results point the same way. ARC Prize recorded 30.2% on ARC-AGI-3 and said the gain comes from stronger logical reasoning. Cursor put it at 66.7 on CursorBench against Fable 5’s 66.5 at default effort. Box CEO Aaron Levie reported gains on his company’s internal eval, including 30% on a life sciences task and 17% on due diligence.

Where does it lose?

Two places, and both sit inside Anthropic’s own numbers.

The first is factual accuracy. Artificial Analysis found accuracy improved 7 points over Opus 4.8, but the model now answers more often when it isn’t sure. That pushed its hallucination rate up 14 points, to 50%. It is the most under-reported finding of the launch.

The second is agentic coding, where the marketing got ahead of the table. On FrontierCode, Anthropic’s launch chart scored Opus 5 at 53.4% and Fable 5 at 53.5%, then shaded Opus 5 as the winner. Readers on Hacker News caught it within the hour.

Anthropic swapped the image out quietly between 17:02 and 19:46 UTC on launch day and never posted a correction. On the other agentic coding row, DeepSWE, Opus 5 scores 68.8% against Fable 5’s 69.7% and GPT-5.6 Sol’s 72.7%. Opus 5 loses both coding rows in Anthropic’s own table, and one of them was briefly presented as a win.

Turn the effort dial down

The most useful thing to come out of launch day is that everyone arrived at the same fix independently: give it less.

Shipper’s team found that deleting their existing skills and starting from scratch made results dramatically better, and that medium or low effort beat high. His read was that the more thinking time you give it, the more likely the annoying behaviour shows up.

Anthropic’s own system card has the same finding, as a chart. Opus 5’s FrontierCode score peaks at 53.4% on medium effort, then falls to 48.0% on high and 43.6% on xhigh. Every other model in the chart just climbs.

Anthropic explains why in the same document, and the explanation is the tester complaint in Anthropic’s own words:

“This reflects a tendency for Opus 5 at these effort levels to make more changes than the task requires, e.g. refactoring or making other edits to improve the codebase.”

Claude Opus 5 system card, p.151

Anthropic adds that a short instruction telling the model to stay in scope recovered most of the lost score, and that it published the numbers without that fix applied. So the behaviour is steerable. It just isn’t the default.

Anthropic’s engineers were saying a version of this on launch day too. Adam Wolff, who works on Claude Code, wrote that Opus 5 at medium effort is his go-to for shipping a PR, and that he still reaches for Fable on bigger projects. Thariq on the same team posted that they removed roughly 80% of the Claude Code system prompt for these newer models.

If you’ve built an elaborate prompt library around Opus 4.8, that is the migration cost nobody priced in.

Key background

Anthropic has now shipped three frontier-tier models in about seven weeks. Fable 5 arrived in June as a Mythos-class model, was pulled during a 19-day US export ban, and came back on July 1 behind a safety classifier that routes anything resembling cybersecurity work to Opus 4.8 instead. Users called that nerfed. Then Sonnet 5 landed at a fifth of Fable’s price and took most of the everyday work.

Opus 5 now claims near-Fable performance at half of Fable’s price, without Fable’s classifier in front of it. The question being asked loudest in the Claude community since Friday is what Fable 5 is still for.

One number complicates the cheaper story. Artificial Analysis puts Opus 5’s cost per task at $2.03 at max effort. That is below Fable 5’s $2.75, but above Opus 4.8’s $1.80 and Sonnet 5’s $1.53. Cheaper than the flagship, not cheaper than what you were probably already using.

Opus 5 is live now on all paid plans and in the API as claude-opus-5, and Anthropic’s full write-up is on its announcement page. The context window is 1 million tokens, though Pro accounts are being told the 1M version isn’t available to them. If you’re moving over from Opus 4.8, two things changed under you: thinking is on by default, and you can’t switch it off at xhigh or max effort.