Anthropic released Fable 5 and Mythos 5 yesterday, new AI models that significantly outperform Claude Opus 4.8 in agentic coding, cybersecurity, and spatial reasoning. Mythos 5 access is currently restricted to vetted Project Glasswing partners, while Fable 5 remains publicly available for immediate use.
The coding and logic performance jump
The new models provide a substantial leap in handling complex, multi-step tasks. According to benchmarks, the models show significant improvements over the previous flagship:
- Fable/Mythos 5 scored 80.3% on SWE-Bench Pro, compared to 69.2% for Opus 4.8.
- Fable/Mythos 5 achieved 38.6% on Blueprint-Bench 2 spatial reasoning, outperforming GPT-5.5's 36.2%.
- In FrontierCode Diamond long-horizon coding, Fable/Mythos 5 reached a score of 29.3, dwarfing the 5.7 score from GPT-5.5.
For developers, this translates to higher-performance AI capabilities for autonomous coding and spatial logic. One notable example: Fable 5 independently completed the video game Pokemon Fire without hints, a task that previously required data unavailable to average players.
Safety measures and access restrictions
Because Mythos 5 is exceptionally proficient at identifying vulnerabilities in code, Anthropic is limiting its deployment. The model is currently available only to vetted Project Glasswing partners (Google, Apple, Microsoft, and trusted defenders including US and allied partner organizations).
Fable 5 is the public-facing option, though it includes specific safety guardrails:
- Cyber/Bio/Chemistry requests are routed to the older Claude Opus 4.8 model via classifiers.
- Fable 5 shows approximately 78% on ExploitBench without this routing, highlighting the underlying power of the new architecture.
- Anthropic maintains mandatory monitoring and retention for high-risk queries.
Pricing and what to do next
The cost for processing requests is higher than standard Claude Opus models but remains competitive. Fable 5 is priced at $10 for 10 million input tokens, twice as affordable as the preview version of Claude Mythos.
If you are currently using Opus 4.8 for complex coding, switch to Fable 5 now to see immediate gains in spatial reasoning and agentic tasks. The performance improvements translate directly to faster development cycles and more reliable autonomous code generation, saving you time on repetitive debugging and logic tasks.
If you require the peak performance of Mythos 5, you must wait for Anthropic to expand its trusted-access program. No general release date has been set.
