The new models score as well or better than Fable 5 for lower cost on the lowest setting. While the sticker price is the same ($10 per M input, $50 out), it is much more efficient and gets the work done with 25% fewer tokens, with 45% less for «highly agentic work.»
Anthropic also says it takes fewer shortcuts in its work, leading to higher-quality outputs compared to Fable 5, and that its safeguards have reduced false positives by 60% — meaning that it will be less restrictive.
The model also stands out in science, scoring more than double Fable 5’s result in the Terminal-Bench-Science benchmark, and while Anthropic aren’t making any specific claims, they do say that «AI models will soon make important contributions to scientific discovery,» and that the new models «offer an early glimpse.»
On cybersecurity, Fable 5.1 is good enough to discover vulnerabilities in code and software, but it is not capable of develop exploits for them. For that, you would need access to Mythos 5.1, which is, as before, the same model with looser safeguards that is only available to trusted partners.
Read more: Anthropic’s announcement and X launch post. The Verge, TechCrunch, Mashable, and Artificial Analysis. Discussion on Hacker News and r/Singularity.