Following in the footsteps of Fable 5 and Sonnet 5, Anthropic is introducing Opus 5, an LLM that achieves virtually the same results as Fable at half the price. Interestingly, according to Anthropic’s benchmarks, it outperforms Fable when it comes to coding and knowledge work.
Opus 5 is available starting today on all platforms. The price remains the same as that of Opus 4.8 and earlier versions. You pay $5 per million input tokens and $25 per million output tokens. This positions the model as a bridge between the more affordable Sonnet 5 and the more expensive Fable 5, which are significantly pricier at $10 and $50 per million tokens, respectively.
Anthropic views Opus 5 as the new standard model on Claude Max and the most powerful model on Claude Pro. The company calls it a model for everyday use that operates more efficiently than other models. Since Fable 5 is no longer available to Pro users and quickly exceeds the 5-hour limits on the more expensive subscription tiers, Opus sounds like a more sensible “daily driver” for AI tasks.
Another very important advantage for large organizations is the absence of the rather draconian data retention policy that Fable 5 enforces. Unlike with Fable, Anthropic chooses not to store prompts for a month for security reasons. As we’ll discuss later, Opus 5 is therefore nowhere near as cyber-savvy as Fable, which is based on Mythos.
Records in coding and knowledge work
In the field of software engineering, Opus 5 is thus said to be more powerful than even Fable 5, possibly due to a more mature reinforcement learning process. Fable, based on Mythos, which was already running internally at Anthropic as early as 2026, is actually starting to show its age in this regard. On Frontier-Bench v0.1, Opus 5 outperforms all other models and doubles the score of Opus 4.8 at a lower cost per task. On CursorBench 3.2, at maximum effort, it comes within 0.5 percent of Fable 5’s peak score, but at half the cost.
More important for Anthropic is its performance relative to GPT-5.6 Sol. The figures sometimes differ dramatically (7.8 percent for Sol on ARC-AGI-3 versus 30.2 percent for Opus 5; Opus 4.8 had achieved only 1.5 percent in this test). However, the general consensus is likely that the choice between Sol and Opus users comes down primarily to personal preference. DeepSWE, a renowned benchmark for agentic coding, shows 72.7 percent for Sol compared to 68.8 percent for Opus 5.
Deliberately weak on cybersecurity
As with Sonnet 5, Anthropic has made a conscious decision not to train Opus 5 on cybersecurity tasks. Nevertheless, the model has improved in this area. It detects vulnerabilities almost as well as Mythos 5, but lags far behind when it comes to writing exploits. This difference is clearly evident in the OSS-Fuzz test.
According to Anthropic, Opus 5 is also the most “aligned” model to date, with a score of 2.3 for anomalous behavior. The cyber classifiers are expected to intervene 85 percent less frequently than with Fable 5. Blocked requests default to Opus 4.8.
The model is also available in Fast mode, which runs at about 2.5 times the standard speed for double the price. In addition, Anthropic is introducing two beta features: tool switching mid-conversation and automatic fallbacks via the API. Developers can get started via claude-opus-5.