Anthropic launched Claude Opus 5 today, 24 July 2026. The pitch isn't to beat OpenAI or Google, but to get close to Fable 5, its own most powerful model, for half the price. Pricing remains identical to the Opus 4.8 it replaces: $5 input and $25 output per million tokens. It becomes the default model on Claude Max and the most capable model available on Claude Pro.
There's a phrase that sums up this launch better than any benchmark: this time, Anthropic isn't trying to beat its rivals, but its own top tier on price. That's a notable shift in messaging, and it says a lot about the phase the AI race has entered. Let's break it down.
What's out today
Claude Opus 5 is available immediately across all of Anthropic's platforms: Claude.ai, the developer platform, Claude Code and Claude Cowork. It becomes the default model on the Claude Max subscription and the most powerful model available to Claude Pro subscribers.
Pricing is the centrepiece of the announcement: $5 per million input tokens and $25 per million output tokens, exactly the price of Opus 4.8. By comparison, Fable 5, launched in June, is priced at $10 input and $50 output. Hence Anthropic's formula: approach Fable 5's cutting-edge intelligence for half the price.
Two technical additions accompany the release. A fast mode, roughly 2.5 times quicker than the default setting, but billed at double the rate. And, more importantly, an effort dial, which lets you trade off between output quality, response speed and token consumption. This is the mechanism we explained in our article on reasoning models: the more a model thinks, the better it gets, but the more it costs.
The announced figures
Anthropic is leading with coding and knowledge-work tests rather than the usual reasoning leaderboards, which betrays its target: the use cases that show up on enterprise invoices.
| Test | What it measures | Announced result |
|---|---|---|
| ARC-AGI-3 | Solving novel problems | 30.2%, roughly 3 times the next model |
| CursorBench 3.2 | Assisted coding | Within 0.5% of Fable 5's top score, at half the cost per task |
| Zapier AutomationBench | Carrying a business task to completion | Roughly 1.5 times the success rate of the closest competitor at equal cost |
| OSWorld 2.0 | Autonomous computer use | Exceeds Fable 5's peak with roughly a third of the budget |
The gains go beyond coding. Anthropic reports improvements across all its life-sciences evaluations, with a jump of more than ten points in organic chemistry, particularly on reading molecular structures from spectroscopy data.
A testimonial from an early-access customer gives a more concrete sense of what this changes day to day. The head of applied research at Harvey, a legal AI company, reports that Opus 5 matched the quality of Opus 4.8 pushed to its maximum reasoning setting, while cutting average token consumption by 26%.
As with any launch, these results are published by Anthropic itself and have not yet been audited by independent third parties. They indicate a trend, not a carved-in-stone truth. It's the reflex we consistently recommend in our article on benchmarks: look at who measured before you look at the score.
The effort dial, the innovation that really matters
Behind the benchmarks, the most structurally significant novelty is probably the effort dial. Until now, choosing a model meant choosing a quality level and accepting the corresponding cost. With an effort setting, the same company can decide task by task how much it's willing to spend on thinking.
This isn't a cosmetic detail. It's a direct response to the most common complaint from professional AI buyers in recent months: unpredictable bills. An agent looping over thousands of tasks can see its cost explode without anyone understanding why. Being able to cap effort turns an incurred expense into a managed one.
What Anthropic admits it isn't
One point deserves highlighting, because it's rare in a commercial announcement. Anthropic explicitly states that Opus 5 is not its most advanced model on dual-use capabilities, such as offensive cybersecurity and biology research. In those areas, Mythos 5 remains ahead.
Likewise, for the longest autonomous tasks, the company says Fable 5 remains the best choice. Opus 5 is positioned as a daily professional workhorse, not as the absolute champion of every category. This clarity in segmentation is welcome, at a time when every launch typically claims to crush everything else.
On behaviour, Anthropic says its automated audit places Opus 5 at the lowest level it has measured on its misalignment scale. The company also says it continues to work with government agencies in its pre-deployment testing phases, including for Opus 5. A detail that's far from trivial after the suspension episode of Fable 5 and Mythos 5 in June.
What this launch reveals
Opus 5 is the fourth model published by Anthropic in under two months, after Mythos 5, Fable 5 and Sonnet 5. That pace is itself information: the competition no longer leaves time to catch your breath.
But the real signal lies elsewhere. For two years, the AI race was about raw capability, with every lab chasing the top spot on a leaderboard. This launch tells a different story: the battle is shifting towards the economics of everyday use. What matters now isn't just which model is the smartest, but how much a completed task costs. That's exactly the logic we observed regarding Anthropic's strategy against OpenAI, and it's also what's pushing Google to cut prices and Chinese labs to flood the market with highly competitive open models.
For the user, this evolution is excellent news. When a new model arrives with better performance at the same price, or with top-tier quality for half the cost, it's a sign of a market where competition works. The interesting question now is no longer who will have the most powerful model in six months, but which one will make that power affordable enough that you stop thinking twice before launching a task.