Claude Opus 5 Arrives: Near-Fable 5 Performance at Half the Price

Pricing and Performance

The part of this release that actually bites is that performance moved and the price did not.

Pricing Held at Opus 4.8 Levels

Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens. Tokens are the small units text is split into, and AI pricing is calculated per token. The price is exactly what the previous Opus 4.8 charged, with performance the only thing that moved.

The other axis is the gap to the tier above. Anthropic describes Opus 5 as coming close to the frontier intelligence of Claude Fable 5 at half the price. Fable 5 costs $10 input and $50 output, so the halving is literal.

The framing appears in the opening line of the announcement itself.

There is also a Fast mode for when speed matters, running at around 2.5 times the default. That one costs twice the base price.

For how pricing has moved across generations, see the Claude Opus generation cost comparison.

Where It Lands on Benchmarks

Anthropic cites both coding and knowledge work. On Frontier-Bench v0.1, a software development evaluation, Opus 5 surpasses every other model and more than doubles Opus 4.8's score, while the cost per task goes down rather than up.

On CursorBench 3.2 the gap to Fable 5's peak score comes in under 0.5%, at half the cost per task. The pattern holds on the knowledge-work side. On ARC-AGI 3, which tests solving novel problems, Opus 5 scores three times as high as the next-best model, and on the computer-use benchmark OSWorld 2.0 it beats Fable 5's best result at just over a third of the cost.

Tuning Intelligence and Cost With Effort

Alongside price, the headline feature is that how much the model thinks is now your call.

What the Effort Setting Changes

Effort specifies how much work the model puts into thinking. Turn it up to prioritise intelligence, turn it down to conserve tokens for faster and cheaper results. Note that in the Japanese-language build of Claude Code this setting is labelled "工数" (workload), which our editorial team confirmed on the actual app; wording may differ by version.

Per the official documentation, the default on the Claude API and Claude Code is high. Without an explicit setting, it runs on the higher side.

The sub-0.5% gap to Fable 5's peak score on CursorBench 3.2 is the figure at max effort, meaning the thinking budget pushed to its ceiling.

Capacity for Long Sessions

The context window, meaning how much can be read in at once, is 1M tokens, with a 128k-token output ceiling. That holds up when you hand over long documents or an entire codebase and keep investigating and editing inside it.

The point up to which its knowledge is considered reliable is May 2026. Anything more recent has to come from search or supplied documents. For Claude as a whole, see our guide to Claude, Anthropic's generative AI.

Safety and the Cyber Question

A large share of the announcement is spent not on performance but on safety.

The Most Aligned Model to Date

Anthropic runs an automated audit of model behaviour before deployment. The result: Opus 5 was found to be the company's most aligned model so far, where alignment means how little the model deviates from intended behaviour.

Its overall misaligned-behaviour score in that audit is 2.3, the lowest among Anthropic's recent models. Lower means fewer unintended behaviours.

It is also described as the best so far at avoiding reckless actions with irreversible side effects. The more work is handed to an AI to run unattended over long stretches, the more that property earns its keep.

Cyber Capability Is Deliberately Held Back

On offensive cyber work, a line is drawn. The model's ability to find vulnerabilities has risen close to Mythos 5, but its ability to turn those vulnerabilities into material threats, effectively writing working exploits, remains substantially behind.

That is deliberate: as with Opus 4.8, Anthropic says it intentionally avoided training Opus 5 on cyber tasks. So that legitimate security work is not blocked, enterprises and researchers already in the Cyber Verification Program have immediate access to a version with fewer security restrictions.

Conclusion

Claude Opus 5 is a solid, practical update: the price stayed put and the performance went up. With a measured gap to Fable 5 under 0.5% on CursorBench 3.2 at half the cost, putting day-to-day production work on Opus 5 is the straightforward choice.

The effort switch lets you push hard problems to max and economise on routine tasks. On safety it takes the best alignment result Anthropic has published while keeping offensive cyber capability held back, and that boundary is stated plainly.

Official announcements come out in English, which makes reading the primary source work. When you want an English web page turned into something easier to read, the following tool helps.

Free ToolURL to Markdown ConverterConvert any public web page URL to Markdown. Preserves headings, tables, lists, and links — perfect for LLM and RAG preprocessing, research notes, and archiving web articles.Try it now →