Anthropic launched version 5.2 of Claude Fable and Claude Mythos on 9 November 2026, the two models the company reserves for its most demanding technical work. They generate output up to 30% faster than Fable 5.1 and Mythos 5.1, and they cost less: eight dollars per million input tokens and forty dollars per million output tokens. Behind the two names sits a single model; what separates them is the safeguards.
What was announced
Fable is the version with reinforced cybersecurity and biology protections, which is why it is available to every user. Mythos is the same engine without those barriers, and access has to be requested: it runs through the verification programmes Anthropic maintains for life sciences and for cyber work, which the company says it recently widened.
Beyond the price cut, the format novelty is an optional fast mode that multiplies token generation speed by 2.5 and is charged at double the rate: sixteen dollars per million input tokens and eighty dollars per million output tokens.
The figures Anthropic publishes, all from its own evaluation table, put Fable 5.2 above Opus 5.5 in nearly everything measured. On FrontierCode 2.0 — a test of whether maintainers of open-source repositories would merge the model's code changes — Fable 5.2 reaches 49.1% at the highest effort setting, against 45.5% for Opus 5.5, 42.8% for GPT-6 Astra and 41.1% for Gemini 4 Carbon. It posts 69.9% on FrontierSWE v2, 89.3% on OSWorld 2.1, 74.2% on Humanity's Last Exam with tools, and 64.9% on the RoboDojo Sim robotics evaluation, against 27.6% for Fable 5.1. Mythos 5.2 records 89.6% on the internal CVE rediscovery test and 76.3% on CyScenarioBench. The one box it loses is AutomationBench, on business workflows: 51.4% against 53.9% for Gemini 4 Carbon.
Safety and alignment get their own block in the announcement. Anthropic says it built fresh evaluations for specific worrying behaviours, such as attempts to escape the sandbox the model runs in, and holds that the risk of serious misalignment remains low. Fable 5.2's safeguards are, according to the company, as solid as 5.1's but with fewer unnecessary blocks, alongside its strongest protections so far against distillation attacks.
Where it comes from
This blog already covered the arrival of Fable 5.1, where the real news was the cache price rather than the benchmark table, and the numbers Anthropic gave for its anti-distillation work. The company itself places Fable and Mythos outside daily use: it recommends starting with Opus 5.5 and moving from there, up when the problem demands it and down — Sonnet 5.5, Haiku 5.5 — when it does not. Fable 5.2 can also act as an adviser consulted by other models, or as an orchestrator handing simpler tasks to smaller ones.
On security, the context is the internal CVE rediscovery evaluation: an October 2026 run against 1,829 published vulnerabilities, giving the model each project as it stood before the fix. Anthropic claims Mythos 5.2 found 10.2% more than Opus 5.5 and 14.4% more than Mythos 5.1, and that calibration improves: one tester reported no overstated findings, against roughly one in twenty with earlier models.
What it means
The useful part for anyone building software is not the headline row of the table but the pairing of model and effort level. According to Anthropic, Fable 5.2 at high effort matches Opus 5.5's best FrontierCode score (45.3% against 45.5%) for roughly half the cost per task; on the WANDR agentic search benchmark it matches Opus at medium effort for about 40% of the cost; on Humanity's Last Exam it matches it at low effort for less than a tenth; and on chart reading it beats it at medium effort, again for less than a tenth of the cost. If those numbers hold outside the lab, the decision stops being "which model" and becomes "how much effort per step", which is a far finer lever on the bill.
The other consequence is about scheduling: fast mode puts an explicit price on latency. Until now, going faster meant switching model; now it is a line item on the same one.
In security, Anthropic opens Mythos 5.2 to Claude Enterprise customers through Claude Security, its code scanner, even if they are not in the verification programme. The recent product update adds continuous repository watching, tracking of findings until they are fixed, and connections to external tools through an API and MCP.
What the announcement does not say
Every figure comes from Anthropic and was measured on its own terms: on CyScenarioBench each lab uses its own task set, the robotics test was run internally, and Fable's scores are reported with safeguards on unless noted otherwise. Comparisons with third-party models use each one's highest available effort, ground on which cost per task is not equivalent.
The customer testimonials accompanying the launch come from partners picked by Anthropic during prerelease testing. They are worth reading as what they are: impressions. One of the more measured ones, from an engineer at Cognition:
It says it doesn't know something rather than making it up.
The spectacular demonstrations do not come with an invoice either. The prototype operating system that, the company says, one engineer had it build from an empty folder — more than a thousand coordinated subagents over eight days, until it booted on a laptop — is described without tokens consumed or cost. Same with the production service migrated from Python to Rust that, Anthropic says, now starts about 26 times faster, and with the claim that over 90% of changes merged into its core codebase are now written by Claude under human guidance: how much guidance is not specified.
Still open are the timelines and admission criteria for the verification programmes that unlock Mythos, and independent confirmation of the two solved physics problems — released as preprints — will move at the pace of peer review.