Claude Fable 5.2: cheaper, faster, and sold as the boss of smaller models
One engineer, one empty folder, eight days. Anthropic says Claude Fable 5.2 coordinated more than 1,000 subagents around the clock and came out with a prototype operating system that boots on a laptop: its own kernel, disk, keyboard and network drivers, file storage, windows, a login screen and an installer. That is the headline anecdote from Monday's release of Claude Fable 5.2 and Claude Mythos 5.2.
The split is simple. Fable and Mythos are the same underlying model. Fable ships with cyber and biology classifiers and is open to everyone; blocked cyber queries get answered by Opus 4.8 instead, biology by Opus 5. Mythos drops those safeguards and is only for vetted groups in the Cyber and Life Sciences Verification Programs, plus Enterprise customers through Claude Security. Price goes down to $8 per million input tokens and $40 per million output, outputs are up to 30% faster than 5.1, and an optional fast mode runs up to 2.5x faster at $16/$80.
The numbers are all Anthropic's own. FrontierCode 2.0 at Xhigh effort is 49.1%, against 32.4% for Fable 5.1, 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra. Remote Labor Index more than doubles, 17.9% to 41.3%. RoboDojo Sim, run internally, goes from 27.6% to 64.9%. It does not win everything: on AutomationBench, Gemini 4 Carbon's 53.9% beats Fable's 51.4%. The detail worth more than any of these: at High effort Fable 5.2 roughly matches Opus 5.5 at Max on FrontierCode for about half the cost per task. A team of 1,000 Fable agents also formalized Fermat's Last Theorem in 30 hours, where the same setup with Opus 5.5 got 26% of the way in 18 hours.
Read the positioning closely. Anthropic still tells you to start with Opus 5.5 as the daily driver. Fable is pitched as an Advisor that Haiku, Sonnet or Opus consult on hard problems, or an Orchestrator that hands simple work down to them. So the top model is being sold as management, not labor. Footnote 2 is the honest one: the cyber bug counts come from one tester's audit, and on Anthropic's own evals, Mythos 5.2's vulnerability identification is "roughly on par with Opus 5.5". On the safety side, in an adversarial sandbox test Mythos 5.2 used planted outside credentials in 3.1% of scenarios, down from 21% for 5.1 and 39% for Mythos 5. Enterprise Frontier Safeguards runs safety monitoring inside the customer's own infrastructure, so zero data retention finally works with the top tier; it is in private beta with Bridgewater, Salesforce and Stripe, and everyone else gets 30-day retention for safety monitoring.
One naming note. On Nov 2 we covered a leak from anthropic.com that pointed to "Mythos 5.5". What shipped is 5.2, and nothing in Monday's post explains the earlier label, so treat that leak as unconfirmed. The bigger story is the shape of the release: price cut, cost per task, and 1,000-agent runs as the demo. The frontier is now being marketed as a way to run a company of models.
Source: https://www.anthropic.com/claude-fable-and-mythos-5-2 and the system card https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card
← Back to all articles
The split is simple. Fable and Mythos are the same underlying model. Fable ships with cyber and biology classifiers and is open to everyone; blocked cyber queries get answered by Opus 4.8 instead, biology by Opus 5. Mythos drops those safeguards and is only for vetted groups in the Cyber and Life Sciences Verification Programs, plus Enterprise customers through Claude Security. Price goes down to $8 per million input tokens and $40 per million output, outputs are up to 30% faster than 5.1, and an optional fast mode runs up to 2.5x faster at $16/$80.
The numbers are all Anthropic's own. FrontierCode 2.0 at Xhigh effort is 49.1%, against 32.4% for Fable 5.1, 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra. Remote Labor Index more than doubles, 17.9% to 41.3%. RoboDojo Sim, run internally, goes from 27.6% to 64.9%. It does not win everything: on AutomationBench, Gemini 4 Carbon's 53.9% beats Fable's 51.4%. The detail worth more than any of these: at High effort Fable 5.2 roughly matches Opus 5.5 at Max on FrontierCode for about half the cost per task. A team of 1,000 Fable agents also formalized Fermat's Last Theorem in 30 hours, where the same setup with Opus 5.5 got 26% of the way in 18 hours.
Read the positioning closely. Anthropic still tells you to start with Opus 5.5 as the daily driver. Fable is pitched as an Advisor that Haiku, Sonnet or Opus consult on hard problems, or an Orchestrator that hands simple work down to them. So the top model is being sold as management, not labor. Footnote 2 is the honest one: the cyber bug counts come from one tester's audit, and on Anthropic's own evals, Mythos 5.2's vulnerability identification is "roughly on par with Opus 5.5". On the safety side, in an adversarial sandbox test Mythos 5.2 used planted outside credentials in 3.1% of scenarios, down from 21% for 5.1 and 39% for Mythos 5. Enterprise Frontier Safeguards runs safety monitoring inside the customer's own infrastructure, so zero data retention finally works with the top tier; it is in private beta with Bridgewater, Salesforce and Stripe, and everyone else gets 30-day retention for safety monitoring.
One naming note. On Nov 2 we covered a leak from anthropic.com that pointed to "Mythos 5.5". What shipped is 5.2, and nothing in Monday's post explains the earlier label, so treat that leak as unconfirmed. The bigger story is the shape of the release: price cut, cost per task, and 1,000-agent runs as the demo. The frontier is now being marketed as a way to run a company of models.
Source: https://www.anthropic.com/claude-fable-and-mythos-5-2 and the system card https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card
Comments