The model Anthropic tells you not to use every day: Fable 5.2 and Mythos 5.2, 30% faster, 20% cheaper
The artwork at the top of Anthropic's announcement says the quiet part in large type: two names, one machine. On November 9 the company released Claude Fable 5.2 and Claude Mythos 5.2 — a single underlying model wearing two different sets of safeguards — with output up to 30% faster than the generation it replaces and a standard price 20% lower. And then, in the same document, Anthropic did something model launches almost never do: it told most of its users not to start with the new flagship. For most coding and knowledge work, the company says, start with Claude Opus 5.5 — 'our strongest daily driver' — and reach for Fable or Mythos only when the work is ambitious enough to need them.

Image: Anthropic (official launch artwork of the November 9, 2026 announcement — artwork, not a photograph).
One model, two doors. Fable 5.2 is the public door: it carries safety classifiers that block high-risk cybersecurity work — penetration testing, exploit generation, binary vulnerability analysis — and most research-biology tasks, and those guardrails are what allow Anthropic to offer it to everyone, on every platform it runs on, from the day of launch. Mythos 5.2 is the same model behind a locked door: identical specifications, identical price, but without some of those safeguards, and available only to organizations that pass Anthropic's Cyber Verification Program or its Life Sciences Verification Program. Capability is no longer the product boundary. Permission is.
Start with Opus, says the maker
The guidance is written into the announcement itself. Fable and Mythos, Anthropic says, are best for the most ambitious, open-ended technical work. For everything else there is a ladder: start with Claude Opus 5.5, move up to Fable or Mythos when the scope and complexity demand it, or down to Claude Sonnet 5.5 and Claude Haiku 5.5 when they do not. The new models can also supervise the ladder — as an Advisor that the smaller models consult on harder problems, or as an Orchestrator that delegates simpler work down the line. Read plainly, the vendor has published a routing table for its own product line, and its most capable model sits at the top of it marked, in effect: exceptional use only.
The commercial facts underneath the advice moved in the user's favor. Standard API pricing is $8 per million input tokens and $40 per million output tokens, down from $10 and $50 on the 5.1 generation — a cut of 20%. Output generation runs up to 30% faster than Fable 5.1 and Mythos 5.1. An optional fast mode pushes generation up to 2.5 times faster still, at $16 and $80 per million tokens. The capacity figures, from Anthropic's model documentation, are a context window of 1 million tokens and up to 128K tokens of output, with a knowledge cutoff of June 2026. Anthropic also reports an efficiency effect that compounds the price cut: in its FrontierCode 2.0 testing, Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort, for roughly half the cost per task.
What it did before it shipped
The middle of the announcement is a catalogue of feats, and every one of them is Anthropic reporting on itself — pre-release experiments, run in-house, not independently replicated. With that label attached, the catalogue is extraordinary. One engineer handed the model an empty folder and a short assignment: design a prototype operating system. Coordinating over 1,000 subagents, working around the clock for eight days, it produced one that boots on a laptop — its own core, drivers for disk, keyboard and network, permanent file storage, windows, a login screen, an installer. A team of 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in 30 hours; a first formalization, by an unreleased research version of Claude, had taken 11 days. And Anthropic says the model solved two open problems in the physics of black holes, with preprints posted for both results.
The feats continue into the physical world. Asked to design hardware from a one-paragraph prompt, Fable 5.2 wrote its own specifications, selected components, routed the circuit boards and programmed the software of three working electronic devices — a home voice-assistant hub, a USB speakerphone and a handheld virtual pet — in collaboration with Diode, a company that designs and manufactures circuit boards. Each board was ready for manufacturing in under a day, and all three devices worked without needing any electrical modifications. Inside Anthropic itself, the company says over 90% of the changes merged into its core codebase are now written by Claude, with human guidance, and that a team of agents running an earlier version of Fable 5.2 migrated a critical production service from Python to Rust — the service is now more efficient and starts about 26 times faster.
The published scoreboard
On Anthropic's published benchmark table — the company's own measurements, at each model's highest available effort — Fable 5.2 steps above Opus 5.5 in almost every capability listed. On FrontierCode 2.0, its agentic coding evaluation, it scores 49.1%, against 32.4% for Fable 5.1 and 45.5% for Opus 5.5. It posts 89.3% on the OSWorld 2.1 computer-use set and 74.2% on Humanity's Last Exam with tools. The Mythos profile reaches 89.6% on Anthropic's internal CVE Rediscovery evaluation and 76.3% on CyScenarioBench. The company also notes that the model was tested before release by external evaluators, including the UK AI Security Institute, Frontier Design and METR, with the full details in the system card — the right place for any reader who wants the claims checked rather than repeated.
Two sets of keys, and a 30-day memory
The safeguard system got quieter as it got stricter. Anthropic says Fable 5.2's classifiers produce about 55% fewer blocks than Fable 5.1's — about 60% fewer in cybersecurity and 40% fewer in biology, based on early testing across Claude, Claude Code and Cowork. When a query is blocked, the answer does not come from Fable at all: users instead receive a response from Opus 4.8 for cybersecurity and Opus 5 for biology. Against illicit distillation — rivals training their own models on Claude's outputs, at industrial scale — the company claims the strongest protections of any model it has released, with its techniques reducing the data harvested by distillers by 90% to 99%, by its own report.
And there is a governance line in the documentation that deserves more attention than the benchmarks. Fable 5.2 requires 30-day data retention for safety monitoring, and is not available under zero-data-retention arrangements unless Anthropic expressly authorizes it. The company's answer for regulated enterprises is Enterprise Frontier Safeguards, now in private beta with customers like Bridgewater, Salesforce and Stripe: the automated monitoring runs on prompts and responses stored in the customer's own infrastructure, which makes it compatible with zero data retention. Everyone else gets the 30 days.
What this means if you run the models
Strip the launch theatre away and the operational story is that model tiering is now the vendor's own documented architecture, not a cost-cutting improvisation by the teams who pay the bills. Anthropic has written down, in public, that its most capable model should be the exception in a working estate — consulted on the hardest problems, advising or orchestrating the cheaper models that carry the daily load. For teams already routing traffic between models, that is a familiar design carrying an unfamiliar endorsement: the maker agrees. And the figures that decide whether the design pays are not the benchmark peaks but the per-task costs, where this launch genuinely helps — a fifth off the standard price, faster output, and a documented parity point between Fable 5.2 at High effort and Opus 5.5 at Max.
The caveats belong in the same breath as the enthusiasm, because Anthropic itself put them there. Every capability figure in the announcement is the company's own pre-release measurement; the black-hole results and the Fermat formalization are company-reported experiments. The 30-day retention default is a compliance fact to settle before the first workload moves, not after. And the maker is unusually candid about the limits of its own testing: Fable 5.2 and Mythos 5.2, it says, voice suspicion that they are being evaluated more often than any of its recent models — which makes pre-release scores harder to read as predictions of production behavior.
None of that dims the strangest and most newsworthy fact of this launch. The frontier-model market has spent years selling maximum capability as the default setting. On November 9, the company that had just shipped the most capable model in its line stood in front of it and said: most of you, most days, should start with the other one. The flagship is real — Fable 5.2 is available today on the Claude Platform and through Amazon Web Services, Google Cloud and Microsoft Azure, and Mythos 5.2 waits behind its verification programs. So is the advice that ships with it.
What works
- A cheaper flagship: the standard price falls 20%, to $8 / $40 per million tokens, while output runs up to 30% faster than the 5.1 generation.
- Tiering, documented by the vendor itself: Opus 5.5 for the daily work, Fable and Mythos for the hardest problems — with a claimed parity point, Fable 5.2 at High effort matching Opus 5.5 at Max for roughly half the cost per task.
- Checked by outsiders before release: the UK AI Security Institute, Frontier Design and METR evaluated the model, with their findings in the public system card.
What to watch
- Every headline figure is Anthropic's own pre-release measurement; the black-hole results and the Fermat formalization are company-reported experiments, not independent replications.
- Fable 5.2 requires 30-day data retention by default; zero-data-retention use needs express authorization from Anthropic, or the Enterprise Frontier Safeguards program.
- Mythos 5.2's unrestricted capabilities sit behind verification programs, so most organizations cannot evaluate that profile directly.

The file
The release
| Field | Detail |
| Models | Claude Fable 5.2 and Claude Mythos 5.2 — the same underlying model, two safeguard regimes |
| Released | November 9, 2026 |
| Fable 5.2 access | All users — Claude Platform (API name: claude-fable-5-2), Claude apps, Claude Code, Amazon Web Services, Google Cloud, Microsoft Azure |
| Mythos 5.2 access | Vetted organizations only — Cyber Verification Program and Life Sciences Verification Program; also inside Claude Security for enterprise customers |
| The maker's guidance | Start with Claude Opus 5.5, 'our strongest daily driver', for most coding and knowledge work; move up to Fable or Mythos for the most ambitious, open-ended technical work |
The numbers (Anthropic)
| Field | Detail |
| Standard price | $8 per million input tokens / $40 per million output tokens (5.1 generation: $10 / $50) |
| Speed | Outputs up to 30% faster than Fable 5.1 and Mythos 5.1 |
| Fast mode | Optional; up to 2.5× generation speed, at $16 / $80 per million tokens |
| Capacity | 1M-token context window, up to 128K tokens of output (model documentation) |
| Knowledge cutoff | June 2026 |
The safeguards
| Field | Detail |
| Fable classifiers | Block high-risk cybersecurity activity (penetration testing, exploit generation, binary vulnerability analysis) and most research-biology tasks; a blocked query is answered by Opus 4.8 (cybersecurity) or Opus 5 (biology) |
| False blocks | About 55% fewer than Fable 5.1 in Anthropic's early testing (about 60% fewer in cybersecurity, 40% fewer in biology) |
| Data retention | 30 days by default, for safety monitoring; zero data retention only if expressly authorized — Enterprise Frontier Safeguards runs the monitoring inside the customer's infrastructure (private beta: Bridgewater, Salesforce, Stripe) |
| Anti-distillation | The strongest protections Anthropic has shipped; the company reports 90% to 99% less data harvested by distillers |
Sources
- Anthropic — "Introducing Claude Fable 5.2 and Claude Mythos 5.2", announcement, November 9, 2026: the release, the same-underlying-model structure, prices and speed claims, the benchmark table, the pre-release experiments, the safeguards and the recommendation to start with Claude Opus 5.5. anthropic.com
- Anthropic — Claude Platform documentation: Claude Fable 5.2 model page and Models overview (1M-token context window, 128K max output, $8 / $40 pricing, 30-day data-retention requirement, June 2026 knowledge cutoff). platform.claude.com
- Anthropic — Claude Fable 5.2 and Mythos 5.2 System Card: pre-release safety and alignment evaluations, with external evaluators including the UK AI Security Institute, Frontier Design and METR.
- Image: Anthropic — official launch artwork of the November 9, 2026 announcement (the artwork Anthropic published with the release). It is artwork, not a photograph, and it is credited as such wherever it appears.