Anthropic has announced Claude Opus 5, a major upgrade to its Opus tier AI model. The new model is designed to power long-running agents and delivers significant improvements in coding and professional work tasks.
SynthePulse Insight · AI deep reading
Version 1 · 2 sources
Anthropic releases Claude Opus 5, claiming it offers near-Fable 5 capabilities at half the price, sets new records on multiple benchmarks, and strengthens cybersecurity protections.
Claude Opus 5 achieves near-frontier performance on multiple benchmarks. On Frontier-Bench v0.1, it surpasses all other models, performing more than twice as well as Opus 4.8 at a lower cost per task. On CursorBench 3.2, with maximum effort settings, its performance is within 0.5% of Fable 5's peak, but at half the cost per task. On ARC-AGI 3, its score is three times that of the next-best model. On Zapier AutomationBench, its pass rate is about 1.5 times that of the next-best model, and even its lowest effort setting outperforms other models. On OSWorld 2.0, it surpasses all models at just over one-third the cost of Fable 5's best result.
Pricing is the same as Opus 4.8 ($5/M input tokens, $25/M output tokens), half that of Fable 5, and slightly lower than OpenAI's GPT-5.6. Anthropic also introduced a 'fast mode' (research preview) that doubles speed but doubles price.
This release comes weeks after Anthropic's latest round of clashes with the U.S. government and days after OpenAI's safety incident. Previously, Fable 5 and Mythos 5 were taken offline for weeks due to government concerns and were re-released with enhanced cybersecurity protections. Anthropic says Opus 5 is the 'most aligned Opus model,' resistant to jailbreaking, and has stronger cybersecurity protections than Opus 4.8. Spokesperson Danielle Ghiglieri stated the company continues to work with the government on independent testing.
Anthropic reported several early test cases: In a Frontier-Bench task, Opus 5 was asked to reconstruct a 3D model from blueprints but could not directly view them; it autonomously wrote a computer vision pipeline to extract geometric information from pixels and successfully reconstructed the model, while other models failed all five attempts. When fixing a bug in a popular open-source package manager, Opus 5 found the root cause and fixed an edge case missed by the community patch, whereas competing models only fixed surface symptoms. An engineer at a trading firm used Opus 5 to build a market data feed for a new exchange in a single session; Opus 5 even built its own testing tools to verify the code.
Early customer feedback includes: In Devin, Opus 5 excelled at difficult debugging and root cause analysis tasks; on Zapier AutomationBench, it ran a complete customer churn prevention sequence end-to-end with a 100% pass rate; in genomics analysis, it chose the correct statistical test and cross-validated results like a meticulous scientist.
This article is primarily based on Anthropic's official press release and reporting by The Verge. Performance data comes from Anthropic's internal evaluations and has not been independently verified. Early test cases were provided by Anthropic and its early access customers and may be selective.
Claude Opus 5 offers near-frontier capabilities at a lower price and shows significant progress in autonomy and reliability, but its cybersecurity performance still lags behind Mythos 5, and its release timing is influenced by recent government regulation and safety incidents.
Primary report
Primary source
Also covered by 2 sources