Anthropic launched Claude Mythos 5 on Tuesday, calling it the world's most capable cybersecurity model, and set its price at $10 per million input tokens.
Key Points:
- Claude Mythos 5 upgrades the restricted Mythos Preview and stays limited to Project Glasswing partners and US government cyber defenders.
- Anthropic priced Mythos 5 and the public Fable 5 at $10 per million input tokens, under half the cost of Mythos Preview.
- Fable 5 routes cybersecurity and biology requests to Claude Opus 4.8, with safeguards firing in fewer than 5% of sessions.
Mythos 5 Unlocks Glasswing
Anthropic unveiled the system alongside Claude Fable 5, a tightly guarded version built for general public use, while reserving the unrestricted Mythos 5 for vetted partners. The model runs on the same underlying engine as Fable 5 but strips away the safeguards that limit cyber and biology work.
For now Mythos 5 reaches only Project Glasswing members and US government cyber defenders, the same group that tested its predecessor. The company first debuted Mythos Preview in April, when its ability to find software flaws drew Wall Street and federal attention. Anthropic said then that it would keep the model off the open market, citing the risk of misuse.
Also Read: OpenAI Joins $852B IPO Race As Anthropic And SpaceX Charge Ahead
Outpacing Opus And Rivals
Anthropic ranks Mythos 5 as the strongest model anywhere at discovering and exploiting software vulnerabilities, ahead of every competing system it tested. Fable 5, by contrast, routes sensitive cyber and biology queries to the older Claude Opus 4.8, which scored roughly half as well on exploit tasks.
That fallback appears in under 5% of sessions, so most users notice no difference between the two models. On some benchmarks Fable 5 scored more than 10% higher than Opus 4.8, its public fallback model. One early tester reported that it matched in 36 hours what OpenAI's GPT-5.5 reached after four days on physics research.
Caution shaped the entire rollout.
Anthropic ran more than 1,000 hours of bug-bounty testing and found no universal jailbreak across the new safeguards. The firm will also retain business customer data for 30 days to study novel attacks, though it says the records will not train future models.
Mythos Preview landed in April as what Anthropic billed a watershed for cybersecurity, with monitored access held to a few dozen critical-software providers. The company has since widened Glasswing toward roughly 150 organizations across more than fifteen countries that build critical infrastructure. Tuesday's dual launch pushes that effort toward a broader trusted-access program for outside firms.
Read Next: Bitcoin Drawdown Reveals Why Big Buyers Still Want More BTC