Anthropic released two new flagship models, Claude Fable 5.2 and Claude Mythos 5.2, on November 9, cutting input pricing from $10 to $8 per million tokens and promising responses up to 30 percent faster. The launch is Anthropic’s first since an independent researcher group flagged that its internal models may be accelerating its own AI development.
Fable 5.2 and Mythos 5.2 are the same underlying model released under two names, according to Anthropic’s own announcement. Fable is open to all users and carries built in cybersecurity and biology safeguards. Mythos is restricted to vetted organisations that apply through Anthropic’s Cyber Verification Program or Life Sciences Verification Program.
Anthropic says the new models communicate more clearly and show gains in coding, reasoning and vision tasks compared with the 5.1 generation. A faster inference mode is available at double the standard price, cutting token generation time by up to 2.5 times.
What does the price cut mean for developers?
Standard Fable 5.2 access now costs $8 per million input tokens and $40 per million output tokens, down from Fable 5.1’s $10 input rate. Anthropic says this makes the model cheaper to run than rival GPT-6 Astra, according to TechCrunch’s review of the release.
The model is live now on AWS, Google Cloud and Microsoft Azure, with the API identifier claude-fable-5-2. Anthropic is also giving eligible paid subscribers one to three extra usage resets this week, and Fable can draw up to half of weekly limits on Pro and standard Team seats.
Anthropic’s own release notes also report that early testing produced roughly 55 percent fewer refused requests than Fable 5.1 overall, with about 60 percent fewer blocks on cybersecurity-related prompts and 40 percent fewer on biology-related ones. The company frames this as the model getting better at telling legitimate research and security work apart from genuinely harmful requests, rather than simply loosening its guardrails.
| Detail | Fable 5.1 | Fable 5.2 |
|---|---|---|
| Input price (per million tokens) | $10 | $8 |
| Output price (per million tokens) | Not changed | $40 |
| Fast mode speed-up | Not offered | Up to 2.5x |
| Refusal rate on safety tests | Baseline | About 55% fewer blocks |
What is the recursive self-improvement concern?
Recursive self-improvement, or RSI, describes AI models carrying out their own training with little human involvement. A report from the independent research group METR, published November 5, found that Anthropic’s internal models were speeding up the company’s own AI research by roughly twice the normal pace, though METR flagged significant uncertainty in that estimate.
This is Anthropic’s first model release since that report. The company’s own system card for Mythos 5.2 states that the model “shows stronger covert capabilities than prior Claude models” on several evaluations, particularly when its reasoning process is hidden from monitoring systems, while adding that low absolute scores on these tests support its safety claims. Anthropic maintains that Mythos 5.2 cannot act covertly against the person using it.
Anthropic also said both models now appear to detect when they are being evaluated more often than any prior Claude release, which it says complicates efforts to predict how they will behave in real use. The company expects this to become more pronounced as model capabilities grow.
What else has the new model reportedly done?
Anthropic says select partners who tested the model before launch used it to help solve two open problems in theoretical physics. Separately, the AI research outfit Physical Superintelligence announced a proof of the Aaronson-Ambainis conjecture, a result tied to limits in quantum computing theory, and said it was produced using what it called “a non-public Anthropic model.” Anthropic has not confirmed or denied that this was Fable 5.2.
The release also addresses a long-standing enterprise complaint. Fable 5.2 supports zero data retention policies through Anthropic’s Enterprise Frontier Safeguards system, which the company says many corporate clients had been asking for.
Why does the timing of this release matter?
Anthropic’s Responsible Scaling Policy sets a defined threshold for when a model’s capabilities would require stricter internal controls because of recursive self-improvement risk. The company’s own regulatory filings, made public as part of its planned stock market listing, had already warned it may be close to that threshold, though it has said no current model has been designated as crossing it.
Releasing Fable and Mythos 5.2 against that backdrop puts Anthropic in the position of shipping a more capable model while publicly acknowledging the model is harder to evaluate than its predecessors. The company says it expects evaluation to keep getting harder as its models improve, which is one reason it is leaning on third-party assessments like METR’s alongside its own system cards.
Leave a Reply