Anthropic Releases Claude Opus 5: Near Fable 5 Intelligence at Half the Price

Anthropic released Claude Opus 5 on Thursday, a model the company describes as 'thoughtful and proactive' and one that comes close to the capability of its top-end Fable 5 at half the price. The model is available immediately through the Claude API, Claude.ai, Claude Code and Claude Cowork.

What the benchmarks show

Anthropic published results across more than a dozen evaluation suites. On Frontier-Bench v0.1 the model scores twice as high as Opus 4.8. On CursorBench 3.2 in maximum reasoning mode it lands within half a percentage point of Fable 5 while costing half as much. On ARC-AGI 3, which measures performance on genuinely novel tasks, the score is three times higher than the next-best model.

In life sciences work Opus 5 improves on its predecessor by 10.2 percentage points on organic chemistry tasks and 7.7 percentage points on protein sequence prediction. On financial modelling the company reports on average 9 percentage points higher accuracy, 60 percent less time and one third fewer tool calls. On first-turn contract redlines the model scores nearly double Opus 4.8.

On FrontierCode 1.1, Claude Opus 5 approaches Fable-level performance at half the cost - Scott Wu, chief executive, Devin

A cheaper route to the frontier

The significant change is not the ceiling but the ratio of quality to price. Anthropic positions Opus 5 as a model for daily work rather than for occasional hardest-case problems. On OSWorld 2.0, which tests computer operation, the model outperforms every alternative at one third the cost of Fable 5.

Partner statements point the same way. Sualeh Asif, co-founder of Cursor, describes intelligence close to Fable 5 at Opus speed and cost. Wade Foster, chief executive of Zapier, says the model topped the company's internal AutomationBench leaderboard without spending more tokens than its competitors.

Safety and safeguards

Anthropic stresses that Opus 5 does not advance the frontier in risky, dual-use capabilities. In the company's internal behavioural audit the model scored 2.3, the lowest and therefore best result among its recent models, with the lowest rate of deceptive behaviour. On OSS-Fuzz tests it identifies vulnerabilities at near parity with Mythos 5, but is considerably less successful at building working exploits.

The cybersecurity filters have also changed: they are roughly 85 percent less restrictive than those applied to Fable 5. The model may look for vulnerabilities in source code, while binary scanning, penetration testing and exploit generation remain blocked. Requests flagged by the classifiers can be routed automatically to Opus 4.8.

New features for developers

Alongside the model Anthropic shipped two API changes in beta. The first allows developers to change which tools the model can use mid-conversation without invalidating the prompt cache, lowering the cost of long agentic sessions. The second lets requests flagged by safety classifiers route automatically to a different model instead of returning a refusal.

What it means for buyers

For teams paying AI vendors by the token, the arithmetic is the story. Holding the price at 5 and 25 dollars per million tokens while closing most of the gap to a more expensive model means a share of work currently routed to the top tier can move down a tier without a quality loss. At the volumes typical of companies deploying agents in customer service or document analysis, that difference is material.

The reported drop in tool calls and completion time matters just as much. In practice the cost of an AI agent is driven not by the price of a single request but by the number of steps needed to finish the job. A model that verifies its own work and reaches for tools less often cuts the bill by more than the rate card alone suggests.

The limits are worth noting. Anthropic itself concedes that Opus 5 remains behind Mythos 5 on offensive cybersecurity tasks, and the benchmark results come from the vendor and have not yet been independently verified. Companies planning a rollout should test the model on their own workloads before switching production processes.