Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on 9 November. Both use the same underlying model, but Fable has additional cybersecurity and biology safeguards, while Mythos access goes through verification programs. The distinction matters more than the names: buyers should first establish which version they can actually deploy. Official announcement.
What developers get
Fable’s documentation specifies a one-million-token context window, up to 128,000 output tokens and the API identifier claude-fable-5-2. Anthropic positions it for demanding reasoning and prolonged agent sessions. A large window expands the material a request can contain; it does not guarantee that every detail will be used correctly. Fable specifications.
Mythos uses claude-mythos-5-2 and shares Fable’s specifications and pricing. Access requires a verified organization. The documentation lists a June 2026 knowledge cutoff, so recent events still need fresh sources rather than answers based solely on training. Mythos specifications.
Price the finished task
Standard API pricing is $8 per million input tokens and $40 per million output tokens. Fast mode costs $16 and $80 respectively. These are US-dollar API rates, not a flat-price consumer subscription. Pricing documentation.
For a simplified request with 100,000 input and 10,000 output tokens, those rates imply $1.20 in standard mode or $2.40 in fast mode, before caching, tools and other adjustments. This is our calculation, not a quoted typical bill. A workflow that repeatedly resends context or generates long intermediate responses can cost much more than its final answer suggests.
Speed claims need a workload test
The release notes advertise up to 2.5 times faster output in fast mode and record a price reduction from the previous generation’s $10/$50 rates. They also specify a 30-day retention requirement, with zero-retention use needing express authorization. 9 November release notes.
Faster token generation is only one part of an agent’s completion time. Searching, running tests and waiting for external services can dominate a job. A useful trial measures time to an accepted result, total spend and the amount of human correction. Teams with confidential material should resolve the retention requirement before uploading it.
A migration needs more than a new model name
Anthropic warns that changing earlier conversation content can invalidate retained thinking blocks. Older computer-use integrations also need attention: computer_20251124 is rejected outside Amazon Bedrock. Compatibility changes.
The migration guide says adaptive thinking cannot be disabled, and forced tool selection is unsupported. Check these assumptions in existing wrappers before redirecting production traffic. Migration guide.
What would justify switching?
Our recommendation is a small comparison using difficult tasks from the actual project: a reproducible bug, a change spanning several modules and a document question with a verifiable answer. Set acceptance criteria before running either model. Keep the inputs and tool permissions comparable, record failed attempts, and include review time in the result.
A model that produces more code is not necessarily the model that delivers more usable work. For a game team, for example, the relevant outcome is whether the change survives tests and fits the existing architecture. This release is worth evaluating where current tools stall, while a measured rollout is more informative than replacing every workflow at once.
Comments