- Read Anthropic’s Mythos announcement here.
Anthropic has launched two Mythos-class models on June 9, 2026: Claude Fable 5, the first such model it has made publicly available, and Claude Mythos 5, restricted to vetted cybersecurity partners. The two run on the same underlying model, and Anthropic attached a mandatory 30-day data-retention requirement to both, overriding existing zero-retention agreements.
The two models, side by side: Both run on identical weights, capabilities, and pricing. What separates them:
- Claude Fable 5: Public. Anyone can use it, with classifiers that intercept high-risk prompts and route them to the less capable Claude Opus 4.8 model.
- Claude Mythos 5: Restricted. Goes only to Project Glasswing cyber defenders and critical-infrastructure partners, with cybersecurity safeguards lifted. Access depends on who the user is, not on a paid upgrade.
Anthropic will retain all Mythos-class traffic for 30 days: Anthropic will hold all traffic on both models for 30 days, on its own and third-party surfaces, even for enterprises that previously had zero-retention deals. It said it will:
- Use the data only to defend against novel attacks and reduce false positives, not to train models.
- Log all human access to the data and delete it after 30 days in almost all cases.
The policy could set an industry precedent where access to frontier models comes bundled with mandatory retention framed as a safety measure, a direct concern for Indian enterprises and developers on the Application Programming Interface (API).
How the safeguards work: Fable 5 sits behind a layer of classifiers, separate Artificial Intelligence (AI) systems that screen prompts before they reach the model.
- When a request touches cybersecurity, biology, chemistry, or model distillation, Fable 5 routes it to Opus 4.8 rather than answering.
- Anthropic said more than 95% of sessions trigger no fallback, so most users get performance equal to Mythos 5.
- In MediaNama’s own testing on June 10, Fable 5 routed a routine request to audit an open-source web application to Opus 4.8, an example of the classifiers catching standard defensive security work.
Screenshot: Fable 5 switched a routine request to audit an open-source web application to Opus 4.8, rather than answering directly. (MediaNama testing, June 10, 2026)
Anthropic says it could not break its own safeguards: Anthropic said it ran over 1,000 hours of external testing and found no universal jailbreak, a method that lets a user bypass a model’s safeguards entirely, though it acknowledged the United Kingdom’s AI Security Institute made partial progress towards one in a brief window. The claim matters because Anthropic has struggled to contain the technology before: on the original Mythos model’s launch day, a group exploited contractor credentials to reach it through a third-party vendor environment.
India recently gained restricted access: Anthropic deploys Mythos 5 through Project Glasswing in collaboration with the United States government and plans a broader trusted-access program. India was shut out of the program through April, but Anthropic’s mid-2026 expansion to over 15 countries brought in a single-digit number of Indian entities, spanning government and private sectors, according to sources cited by Livemint.
Neither Anthropic nor the organisations have disclosed which Indian institutions hold the access. Fable 5 now offers the wider Indian public the capability in guard-railed form, while Mythos 5 stays limited to that small, unnamed cohort.
Why it matters: Fable 5 makes the most capable model Anthropic has released publicly available, but on terms that tighten control rather than loosen it. MediaNama founder and editor Nikhil Pahwa has argued in Reasoned that a tool which “compresses attack timelines without compressing defence timelines increases systemic risk before it improves security,” and that logic extends to a public release.
Wider access spreads both the defensive and offensive potential of Mythos-class capability, and the retention mandate signals that Anthropic intends to monitor how the public uses it.
Pricing: Fable 5 and Mythos 5 cost $10 per million input tokens and $50 per million output tokens, which Anthropic said is less than half the price of the earlier Mythos Preview. Anthropic will include Fable 5 in Pro, Max, Team, and seat-based Enterprise subscription plans until June 22 and then require usage credits, with a plan to restore it as a standard subscription feature when capacity allows.
Also read:
