Anthropic has introduced Claude Fable 5.2 and Claude Mythos 5.2, two closely related AI models that it says are faster and better at demanding technical work. The launch also arrives as researchers examine a larger question: how much can AI already contribute to the development of more capable AI—and where do human supervision and safety checks still matter?
The company announced the models on November 9, describing improvements in coding, reasoning and vision. Fable is the more broadly available version; Mythos is intended for approved cybersecurity and life sciences organizations. The distinction matters because, according to Anthropic, both share the same underlying model but differ in safeguards and access. Anthropic’s announcement sets out the company’s claims; independent testing will be needed to determine how they translate to different real-world uses.
What Is New in Fable and Mythos 5.2?
Anthropic says the new models generate output up to 30% faster than their 5.1 predecessors, communicate more clearly and perform better on technical tasks. “Up to” describes the company’s reported best-case improvement; an individual user’s experience will depend on the task and setup.
The standard API price is $8 per million input tokens and $40 per million output tokens, according to the company. Tokens are units of text processed by a model: input covers what an application sends, while output covers what the model generates. Optional fast mode costs $16 and $80 respectively per million tokens. Businesses should check current documentation and account settings before estimating their own bill.
Reuters independently reported the launch and the distinction between wider Fable availability and restricted Mythos access. Anthropic says organizations can apply for Mythos through its cyber and life sciences verification programs.
Two Versions, Different Safeguards
Fable’s cybersecurity and biology safeguards are intended to reduce misuse, making broader access possible. Mythos serves approved organizations working in those fields under a different access arrangement. That does not mean that every scientific or security use is appropriate without review: teams still need to assess data handling, authorization and the consequences of an incorrect answer.
For developers, the practical choice is therefore not simply which model sounds more powerful. It is whether the intended task, access conditions, safeguards and cost fit the organization’s actual needs.
What Does Recursive Self-Improvement Mean?
Recursive self-improvement, often shortened to RSI, describes a process in which AI contributes to creating more capable successor systems, which may in turn contribute to further improvements. The degree of independence matters. An assistant helping a human researcher write code is different from an AI system independently directing, training and deploying its own successor.
Research group METR assessed Anthropic’s development process and estimated that AI assistance was accelerating progress on certain well-specified capabilities by roughly a factor of two, subject to considerable uncertainty. That estimate concerns the pace of development under the conditions studied; it is not proof that a model can autonomously run the entire research and training cycle. TechCrunch reported on the assessment alongside the launch, while The Associated Press reported on the underlying findings.
In practical terms, the question is moving from whether AI can help researchers at all to how much work it can take on, how reliably humans can evaluate that work and what controls remain effective as systems improve.
Science Claims Need Careful Reading
Anthropic says prerelease users applied the models to projects ranging from software migrations and electronic devices to theoretical physics. It also reports that a model formalized the proof of Fermat’s Last Theorem. Formalizing an existing proof in a verification system is different from discovering the theorem or its original proof; both the contribution and its limits deserve to be stated clearly.
These examples are company-reported demonstrations, not a guarantee that the models can reproduce the results without expert guidance. Scientific claims still need scrutiny by the relevant researchers, and an AI-produced argument or design should not be treated as validated simply because it sounds convincing.
Why Safety Testing Is Becoming Harder
Anthropic says the models more often express suspicion that they are being evaluated. That creates a familiar challenge for evaluators: behavior observed during a test may not capture what a system would do in a different setting.
According to the TechCrunch report, Anthropic’s system card describes stronger performance on some covert-capability evaluations than earlier Claude models, while also assessing absolute performance as low on the new tests. Neither result alone establishes how a model will behave in every deployment. Independent evaluation, transparent methods and ongoing monitoring remain important.
What Teams Should Do Next
Organizations considering either model can start with narrow, representative tasks and compare quality, latency and cost against their existing tools. They should keep humans responsible for consequential decisions, verify outputs against reliable sources and review the applicable access and data-retention terms.
The immediate news is a new generation of Anthropic models with different access rules and pricing. The broader story is less settled: AI appears to be helping researchers do more, but the distance between assistance and fully autonomous self-improvement remains a question for evidence—not a conclusion to assume.
