Summary
- Anthropic has released Claude Fable 5.2 and Mythos 5.2, with Fable 5.2 currently ranked first on the Artificial Analysis benchmark.
- Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens, a 20% reduction from Fable 5.1.
- A known token-exhaustion bug that previously affected Opus 5.5 has reappeared in Fable 5.2 at the maximum reasoning level, causing the model to run out of tokens without returning output.
- According to corroborating sources, Mythos is being integrated into Cisco and IBM products, broadening the enterprise attack surface for AI-dependent workflows.
- Security teams evaluating these models should account for the token-exhaustion behaviour before deploying high-reasoning configurations in production pipelines.
What was released
Anthropic has released two new models: Claude Fable 5.2 and Mythos 5.2. Fable 5.2 currently holds the top position on the Artificial Analysis leaderboard. Anthropic has described the model as capable of solving what it characterises as two open problems in the physics of black holes, suggesting its internal evaluation criteria have grown considerably more demanding. The new models are available via Anthropic's API and through the llm-anthropic tooling ecosystem.
Pricing and positioning
Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens. That represents a 20% reduction compared with Fable 5.1, which was priced at $10 and $50 respectively. For organisations running high-volume inference workloads, the cost reduction is material, though the per-output pricing remains substantial at scale. CISOs building cost models for AI-assisted security tooling should update their assumptions accordingly.
A returning bug at maximum reasoning
Fable 5.2 supports five reasoning levels: low, medium, high, xhigh, and max. At the maximum reasoning level, the model has been observed exhausting its token budget, in one documented case consuming 128,000 tokens, without returning any output. This is not a new class of problem. The same behaviour was previously observed with Opus 5.5. What is notable is that it now affects the Fable series for the first time; both Fable 5 and Fable 5.1 handled maximum-effort tasks without triggering this failure mode. The xhigh setting has been reported as producing the best results without hitting the ceiling.
Enterprise integrations via Cisco and IBM
According to reporting from The Register, Mythos is being incorporated into products from Cisco and IBM. The specific implementation details are not fully described in the available source material, but the integrations signal that Mythos-class capabilities are moving into enterprise security and infrastructure tooling. For CISOs, this means AI model behaviour, including edge cases like token exhaustion, may surface inside vendor platforms rather than in direct API usage, where they are harder to observe and govern.
RSI concerns noted in coverage
TechCrunch's coverage of the release references repetitive strain injury concerns alongside the model launch. The source material does not elaborate on the specifics of this issue, and it would be inappropriate to speculate beyond what is reported. Security leaders should monitor further coverage for clarification on what this involves and whether it has any bearing on workforce or vendor risk considerations.
Why it matters
As frontier AI models are embedded more deeply into enterprise platforms from vendors such as Cisco and IBM, the operational characteristics of those models, including failure modes like token exhaustion, become part of the enterprise risk surface. A model that silently consumes its entire compute budget and returns nothing is a reliability and potentially a safety concern in any pipeline where the output drives a decision or automated action. CISOs need visibility into which reasoning configurations their vendors are using, what guardrails exist around runaway inference, and how failures are surfaced to operators. The price reduction in Fable 5.2 may also accelerate adoption inside development and security teams, making internal AI governance policies more urgent to enforce.
What to do now
- If your organisation uses Anthropic models directly via API, avoid the maximum reasoning level in Fable 5.2 for production workloads until the token-exhaustion behaviour is resolved or officially acknowledged by Anthropic.
- Review AI usage policies to account for high-reasoning model configurations, which carry different cost and reliability profiles than standard inference settings.
- Where Cisco or IBM products are incorporating Mythos, raise with those vendors what reasoning level configurations are in use and what failure handling exists if the model exhausts its token budget.
- Update cost models for AI inference workloads to reflect the revised Fable 5.2 pricing of $8 per million input and $40 per million output tokens.
- Monitor further reporting on the RSI concerns cited in TechCrunch coverage before drawing conclusions about any workforce or vendor risk implications.
