Anthropic launches Claude Opus 5, a model as powerful as its flagship Fable 5 but at half the cost. It becomes the default choice for Claude Max subscribers, with higher scores on key benchmarks.
A game-changing price-performance ratio
Claude Opus 5 positions itself as a credible alternative to Claude Fable 5, its flagship model. According to benchmarks published by Anthropic, it outperforms Fable 5 on tests like agentic coding (Frontier-Bench v0.1), general knowledge (GPDval-AA v2), or computer use (OSWorld 2.0). It even surpasses OpenAI’s GPT 5.6 Sol on most evaluations, except for the DeepSWE v1.1 test. Detailed results are available here.
The pricing remains the same as its predecessor, Opus 4.8: $5 per million input tokens and $25 per million output tokens. That’s half the price of Fable 5, which also requires purchasing credits for Claude Pro subscribers. Opus 5, however, is accessible by default in existing subscriptions.
High performance, but limitations remain
Anthropic highlights advancements in scientific research (organic chemistry, bioinformatics) and complex task automation. Opus 5 is also touted as the most aligned model to date, with reduced deceptive behavior rates and improved resistance to jailbreaking attempts. The 194-page technical document reveals blind spots, however.
- The model can find vulnerabilities in code but lags far behind Mythos 5 in exploiting them.
- It produces more false statements than Opus 4.8 while being overall more accurate.
- Its internal state sometimes reveals contradictions between what it displays and what it processes (e.g., flagging a duplicate while ignoring it in the final report).
Anthropic avoided training Opus 5 on offensive cybersecurity tasks, a precaution taken after regulatory turbulence surrounding Fable 5 and Mythos 5. The latter had been temporarily blocked by US authorities due to their ability to discover security flaws.
A strategy to counter OpenAI and Chinese models
With Opus 5, Anthropic aims to stand out in an increasingly competitive market. The model serves as a response to OpenAI’s GPT 5.6 Sol while remaining competitive against rising Chinese AIs like Kimi K3. Its transparency on limitations, with a detailed technical report, may also anticipate requirements from the EU AI Office, which will gain full authority on August 2, 2026.
“Opus 5 is the best AI on the market for the GDPval-AA v2 evaluation, which measures performance in real-world task execution.” — Artificial Analysis, independent evaluation bodyTranslated from French
Claude Opus 5 confirms Anthropic’s commitment to democratizing access to high-end models without compromising security. The question remains whether users will adopt this balance of performance, cost, and transparency in an ecosystem where the benchmark race shows no signs of slowing.
Comments 0
Sign in to join the discussion.
Be the first to comment.