— The capabilities of models like Fable 5 and Mythos 5 have the potential to do profound good for the world, Anthropic writes in its presentation, and adds — We’ve also seen it in life sciences research, where the models are positing novel hypotheses and speeding up the development of new therapeutics.
Expanding Project Glasswing
Fable 5 will be generally released through Claude, while Mythos 5 has fewer guardrails and gets distributed to a small set of security researchers and strategically important companies through Project Glasswing, which should be expanding its scope «soon,» including for biomedical users.
Both models are a leap forward in benchmarks, exceeding Opus 4.8 by 15—20% in some cases, and coming in miles ahead of GPT-5.5. Anthropic says the models get better with more complex tasks.
Second AI to monitor
The key to Anthropic’s safeguards is having a second AI, called a Classifier, constantly manage throughput and sending attempts at jailbreaking and distillation to a lesser model. Advanced queries on chemistry, biology and cybersecurity are also sent to Opus 4.8. This has resulted in a near zero rate of abuse.
Both models are priced at double the rate for Opus 4.8, but half the price of Mythos Preview at $50 per million output tokens and $10 per million input tokens.
Read more: Anthropic’s announcement, launch thread, with Ars Technica, The Register, The Verge, and Engadget. Discussion on r/Singularity.