Claude Opus 5 Lands at Half the Price of Anthropic's Best Model

Anthropic released Claude Opus 5 on July 24, and the pricing line says more about the company's strategy than the benchmark chart does. The model costs $5 per million input tokens and $25 per million output tokens: exactly what Opus 4.8 cost two months ago, and half of what Fable 5 costs today.

It shipped everywhere at once, across the Claude apps, the API, Claude Code and Claude Cowork.

The benchmark picture

Anthropic's own numbers put Opus 5 in the same conversation as its larger sibling rather than a tier below it. On Frontier-Bench v0.1, the company says Opus 5 more than doubles Opus 4.8's score at a lower cost per task. On CursorBench 3.2 it lands within half a percentage point of Fable 5's peak score at half the cost per task. On OSWorld 2.0 it beats Fable 5's best result at a little over a third of the cost. On ARC-AGI 3 it scores roughly three times the next-best model.

Reporting from CNBC and others puts the headline figures at 43.3% on Frontier-Bench and 30.2% on ARC-AGI-3. Anthropic still points customers to Fable 5 for the hardest jobs, particularly agents that run autonomously for days, and says Mythos 5 remains ahead on biology research and offensive cybersecurity.

An effort dial, and what it is really for

The most useful addition for anyone running volume workloads is a low/medium/high effort setting that lets you trade capability against cost on a per-request basis. Fast mode roughly doubles the price and runs about 2.5 times faster. Two features arrived in beta: swapping tools mid-conversation, and automatic fallback routing, which sends a request to a lighter model when a safety filter fires instead of returning an error.

One line in the release matters more to enterprise buyers than any benchmark: Opus 5 carries no data retention requirement. Anthropic also says the model shows its lowest recorded rates of deceptive behavior, and that cybersecurity classifiers intervene around 85% less often than they do for Fable 5, which means fewer false alarms for teams doing legitimate security work.

The strategy underneath the price tag

Holding the price flat while roughly doubling measured performance is a deliberate signal. Anthropic is betting that most paying customers, especially those running agents at scale, care about cost per successful task rather than about who tops a leaderboard this quarter. With frontier scores converging for several release cycles now, that looks like a reasonable read of how procurement conversations actually go.

If you are budgeting for agents

The reflex we see most often in the field is teams picking whatever sits at the top of the model list and wiring every workflow to it. A support-reply generator and a 40-page contract analyser do not deserve the same model, and paying frontier rates for the first one quietly eats the budget that should fund the second.

That is why the effort dial is worth more than the price comparison. Used properly, it lets a single architecture spend differently on different jobs. Model your costs per successful task rather than per token, and treat model choice as a per-workflow decision.

Sources: Anthropic, CNBC, TechCrunch