San Francisco: Anthropic has unveiled Claude Opus 5, a new flagship AI model designed to deliver near frontier-level performance for coding and knowledge work at roughly half the cost of its top-tier Claude Fable 5 model.
Available immediately, Claude Opus 5 becomes the default model for Claude Max subscribers and the most capable model available on Claude Pro. The company said it matches the pricing of its predecessor, Opus 4.8, at USD 5 per million input tokens and USD 25 per million output tokens.
Anthropic said Opus 5 sets new benchmarks across software engineering and knowledge work evaluations, outperforming rival models on tests such as Frontier-Bench v0.1, CursorBench 3.2 and GDPval-AA.
The company added that Opus 5 delivers significantly better performance than Opus 4.8 while allowing users to adjust the model’s “effort” setting to balance intelligence, speed and cost.
On ARC-AGI 3, a benchmark designed to test reasoning on unfamiliar problems, Anthropic said Opus 5 scored three times higher than the next-best competing model. It also outperformed rivals on AutomationBench, which measures whether AI systems can complete end-to-end business tasks, and OSWorld 2.0, a benchmark for computer-use capabilities.
Anthropic said Opus 5 is designed to verify its own work more thoroughly and iterate until tasks are completed successfully.
Among the examples shared by the company, the model reportedly built its own computer vision pipeline to recreate a mechanical part as a 3D model after being denied direct access to the original drawing. In another case, it identified the root cause of a software bug that community-developed fixes had failed to address.
Early users also reported that the model successfully built a market data feed for a financial exchange, including creating its own testing framework when no live data was available for validation.
Beyond software engineering, Anthropic said Opus 5 delivers meaningful improvements for scientific research.
The model showed higher performance across internal life sciences evaluations covering structural biology, organic chemistry and bioinformatics. According to the company, the biggest gains came in tasks involving molecular structure analysis and predicting the effects of protein sequence variations.
Anthropic also highlighted improvements in visual reasoning, saying Opus 5 can generate more sophisticated scientific and engineering visualisations than previous Claude models.
Anthropic said Opus 5 is its most aligned AI model to date, exhibiting lower levels of deceptive behaviour and better adherence to the company’s constitutional AI principles than Opus 4.8, Sonnet 5 and Fable 5.
The company stressed that although Opus 5 has become more capable in cybersecurity through broader improvements in reasoning, it intentionally remains behind Mythos 5 in offensive cyber capabilities and exploit generation.
To balance usability and security, Anthropic has introduced updated cyber safeguards that allow vulnerability identification while continuing to block exploit generation, penetration testing and certain binary-based security tasks. The company said these safeguards are expected to intervene around 85 per cent less frequently than those applied to Claude Fable 5.
Alongside Opus 5, Anthropic introduced two beta features for developers: the ability to change available tools during an active conversation without invalidating prompt caches, and automatic API fallbacks that route safety-flagged requests to alternative Claude models instead of blocking them entirely.
Anthropic also launched a Fast mode for Opus 5, offering response speeds around 2.5 times faster than the standard version at double the base usage price.
With stronger coding performance, improved reasoning, enhanced scientific capabilities and refined safety controls, Claude Opus 5 represents Anthropic’s latest attempt to balance cutting-edge AI performance with everyday usability for developers and enterprise customers.