Anthropic has released Claude Opus 5 – a cheaper model close to Fable 5 in performance

Anthropic has released the new model Claude Opus 5. The price remains the same as for Opus 4.8 (5 USD per million input tokens and 25 USD per million output tokens), and the company presents it as half the price of Fable 5. The context window remains unchanged at 1 million tokens. "Fast mode" is now also available in research preview, offering a 2.5-fold speedup at twice the price. Opus 5 becomes the default model in Claude Max and the most capable available model in Claude Pro. Users can choose between five "effort" levels (low, medium, high, xhigh, max), and Anthropic recommends low or medium settings for routine tasks in its prompting guide, while recommending xhigh for programming and agentic tasks; on two benchmarks, the max setting reportedly scores worse than xhigh despite costing more.

According to benchmarks from Anthropic, Opus 5 leads in agentic terminal coding (Frontier-Bench v0.1: 43.3 % versus 33.7 % for Fable 5, 34.4 % for GPT-5.6 Sol, and 21.1 % for Opus 4.8) and in knowledge work (GDPval-AA v2, Elo 1861 versus 1747 for Fable 5). On ARC-AGI-3, a benchmark focused on solving novel tasks, it achieved 30.2 %, nearly four times the second-best result (GPT-5.6 Sol, 7.8 %) – both Anthropic and The Decoder describe this result as striking, and it is uncertain whether a lead of the same size will also appear in everyday use. On the agentic coding benchmark DeepSWE v1.1, however, GPT-5.6 Sol leads ahead of Fable 5, with Opus 5 finishing third; according to Anthropic, Fable 5 and Mythos 5 perform better on healthcare and legal tasks, respectively. The Verge adds that Anthropic continues to recommend Fable 5 for the most demanding and long-term agentic projects, while positioning Opus 5 as a model for everyday use, knowledge work, and biological research.

In cybersecurity, Anthropic says that Opus 5 was not deliberately trained on offensive cyber tasks – it comes close to Mythos 5 in finding vulnerabilities but falls well behind it in exploiting them. The company claims that cybersecurity safety classifiers trigger about 85 % less often with Opus 5 than with Fable 5, whose frequent interventions drew criticism; blocked requests in Claude.ai, Claude Code, and Claude Cowork switch to Opus 4.8 by default. Anthropic describes the model as the "most balanced" model in the Opus series, the least susceptible to misuse. The Verge notes that the release comes weeks after a dispute between Anthropic and the US government over Fable 5 (temporarily withdrawn over concerns about cyber risks) and days after a security incident at OpenAI; a spokesperson for Anthropic said that the government is conducting its own independent testing of Opus 5.

According to Anthropic, in one Frontier-Bench test, the model was given only a drawing of a mechanical component without being able to view it directly, wrote its own computer vision pipeline to extract geometry from pixels, and built a functional 3D model in FreeCAD – reportedly, no other model managed the task even in five attempts. Anthropic also cites an example of fixing a bug in a popular open-source package manager, where Opus 5, unlike a competing model, eliminated the root cause, and a case of an engineer at a trading firm who used the model to create a data feed for a new exchange in a single session. In an informal test of drawing a pelican on a bicycle, Simon Willison noted that the bicycle was missing its wheels in the first attempt, while the second attempt turned out better. Details on scientific research results and additional benchmarks can be found in the source articles.