The launch marks a significant shift for the San Francisco-based AI company, which had previously limited Mythos-class capabilities because of concerns over advanced cybersecurity and biological-risk use cases. Claude Fable 5 is being positioned as a frontier model for software engineering, knowledge work and visual analysis, with stronger performance on long, complex tasks that require sustained reasoning across large volumes of information.
The model uses new safeguards designed to block or restrict responses in high-risk areas. When a user request crosses certain thresholds, especially in cybersecurity or biology, the system can fall back to Claude Opus 4.8 rather than allowing Fable to complete the task. Internal testing showed that about 95 per cent of Fable sessions ran entirely on Fable responses without needing such fallback, suggesting that the restrictions are aimed at a relatively narrow set of sensitive cases.
Anthropic is also offering Claude Mythos 5, understood to be based on the same underlying model as Fable 5 but with fewer restrictions for vetted users. Access to Mythos 5 remains limited through the company’s controlled programmes, including organisations involved in Project Glasswing, a security initiative focused on finding and fixing vulnerabilities in critical software.
Pricing places both Fable 5 and Mythos 5 above Anthropic’s previous flagship models. The listed rate is $10 per million input tokens and $50 per million output tokens, double the price of Claude Opus 4.8, though still below the cost of some early Mythos preview access arrangements. Anthropic is betting that stronger performance on long-horizon work will reduce the number of prompts, retries and manual interventions required for complex tasks.
The release comes as frontier AI firms face growing pressure to prove that more capable models can be deployed without widening cyber, fraud and biosecurity risks. Anthropic has built its public identity around AI safety, but the Mythos line has tested how far that approach can stretch when models become highly effective at discovering software flaws and supporting technical operations that could be used defensively or offensively.
Project Glasswing has become central to that argument. Anthropic and its partners have used Mythos Preview to identify more than 10,000 high- or critical-severity vulnerabilities across important software systems. Several partners reported large increases in bug-finding rates, while open-source scanning identified thousands of possible flaws requiring triage, disclosure and patching. The company has said the main bottleneck is no longer finding vulnerabilities but verifying, reporting and fixing them responsibly.
That context explains the split between Fable and Mythos. Fable gives businesses and developers access to much of the model’s capability while placing barriers around sensitive domains. Mythos remains reserved for organisations that can demonstrate a legitimate need for stronger cyber functionality, such as vulnerability research, red-teaming or infrastructure defence.
The move may also intensify competition among AI developers serving enterprise customers. Software engineering has become one of the most valuable commercial use cases for advanced language models, with companies seeking tools that can understand large codebases, fix defects, generate tests and assist with security reviews. A model that performs better on lengthy, multi-step engineering tasks could give Anthropic an advantage among developers, cloud partners and large corporate clients.
Still, the release raises questions over transparency. Anthropic has not fully explained how the Fable and Mythos lines relate to earlier Claude Opus, Sonnet and Haiku naming conventions, or why the first public Fable release carries the number 5. The company’s decision to keep the unrestricted version behind a trusted-access framework also leaves unresolved questions about who qualifies, how usage will be monitored and how restrictions may be adjusted over time.
For customers, the immediate appeal lies in stronger reasoning and memory across demanding workflows. For regulators and security researchers, the release will be watched as a test of whether model-level safeguards can contain risks once capabilities move from preview programmes into wider public use.
