Anthropic’s New Flagship Is Actually Two Models — and the Difference Matters
On Monday, 9 November 2026, Anthropic launched its most capable AI models yet: Claude Fable 5.2 and Claude Mythos 5.2. The two names sound like separate products, but Anthropic is unusually blunt about the truth: they are the same underlying model. The only difference is what each one is allowed to do.
Fable 5.2 is the version anyone can use, with cybersecurity and biology safeguards switched on. Mythos 5.2 is the same model with some of those safeguards removed, available only to vetted organisations. Both cost less than their predecessors, run faster, and beat every rival on agentic coding on Anthropic’s own scorecard.
Fable 5.2 vs Mythos 5.2: One Model, Two Safeguard Levels
Claude Fable 5.2 is the general-availability model. It ships with classifiers that block high-risk cybersecurity work — penetration testing, exploit generation, binary-based vulnerability research — and most research-biology tasks. When Fable refuses a query, Anthropic routes it to Claude Opus 4.8 (cyber) or Claude Opus 5 (biology) instead.
Claude Mythos 5.2 is the identical model with those classifiers lifted. It is not sold to the public — access runs through Anthropic’s Cyber Verification Program and Life Sciences Verification Program, which vet the organisation asking for it, so legitimate security researchers and life-science labs get the unfiltered model while everyone else gets the guarded one.
On ordinary work — writing code, analysing documents, running agents — the two perform identically. Mythos only pulls ahead on the gated tasks, because those are the tasks Fable declines. If you are not doing cybersecurity or research biology, Mythos would not give you a better answer. Fable 5.2 also blocks about 55% fewer requests than Fable 5.1 — roughly 60% fewer in cybersecurity and 40% fewer in biology — so the safeguards are still there, but less trigger-happy.
What’s New in Fable 5.2: Price, Speed and Specs
The headline change is the price. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens on the Claude API, down from $10 and $50 for Fable 5.1 — a 20% cut that also applies to prompt caching and the Batch API. Mythos 5.2 carries exactly the same prices, and both undercut OpenAI’s rival: GPT-6 Astra lists at $10 and $50 per million tokens.
The core specifications are unchanged from the 5.1 generation, but they were already generous: a 1 million-token context window, a 128,000-token maximum output, a June 2026 knowledge cutoff, and output up to 30% faster than Fable 5.1. An optional fast mode generates tokens up to 2.5 times faster at double the price ($16/$80 per million).
It is available from day one on the Claude API, the Claude apps, Claude Code, Amazon Bedrock, Google Cloud, Microsoft Foundry and the Claude Platform on AWS, under the model IDs claude-fable-5-2 and claude-mythos-5-2.
The Benchmarks: Where Fable 5.2 Leads — and Where It Doesn’t
Anthropic published one comparison table pitting Fable 5.2 against Fable 5.1, Claude Opus 5.5, GPT-6 Astra and Google’s Gemini 4 Carbon. Every score is Anthropic’s own, run at each model’s highest effort level — read them as a vendor’s scorecard, not independent verification.
| Benchmark | Fable 5.2 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|
| FrontierCode 2.0 (agentic coding) | 49.1% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2 (agentic coding) | 69.9% | 62.3% | 65.5% | 61.2% |
| OSWorld 2.1 (computer use) | 89.3% | 87.2% | 79.5% | 81.5% |
| AutomationBench (business workflows) | 51.4% | 42.5% | 41.4% | 53.9% |
| Humanity’s Last Exam (with tools) | 74.2% | 67.7% | 57.2% | — |
| Remote Labor Index | 41.3% | 21.3% | — | — |
The jump from Fable 5.1 is dramatic: FrontierCode 2.0 rose from 32.4% to 49.1%, and FrontierSWE v2 from 56.3% to 69.9%. Fable 5.2 leads every listed rival on agentic coding, computer use and the hardest exam benchmark, and at High effort it roughly matches Opus 5.5’s best coding score for about half the cost per task. It does not lead everywhere: Gemini 4 Carbon wins AutomationBench, and Opus 5.5 holds a slightly higher AA-Briefcase Elo (1800 vs 1785).
Mythos 5.2 gets its own headline numbers on cyber benchmarks: 89.6% on CVE Rediscovery across 1,829 known vulnerabilities (against 78.3% for Mythos 5.1) and 76.3% on CyScenarioBench (against 61.7%). Mozilla, which uses the model in its scanning campaigns, reportedly found it roughly doubled vulnerability discovery compared with Mythos 5.1. OpenAI’s GPT-6 Cyber scored 61% on CyScenarioBench for comparison — though different labs use different tasks, so cross-vendor comparisons here are rough at best.
The Demos That Got Attention
The pre-release tests Anthropic chose to publicise are stranger than the usual coding scores.
A thousand agents proved a famous theorem. A team of about 1,000 Fable 5.2 agents formalised the proof of Fermat’s Last Theorem in Lean in roughly 30 hours, with minimal human intervention.
It designed hardware from a sentence. From a one-paragraph prompt, Fable 5.2 designed three electronic devices from scratch — a home voice-assistant hub, a USB speakerphone and a handheld virtual pet — each with its own circuit board, 3D-printed enclosure and software, and Anthropic says the hardware worked first time.
It solved open physics problems. The model solved two open problems involving black holes and quantum gravity, and Mythos 5.2 found a new way to predict the structure of adeno-associated virus (AAV) capsids — the protein shells used to deliver gene therapies — accurately modelling 28 of 34. On RoboDojo, a public benchmark for controlling a simulated two-arm robot from camera images, Fable 5.2 scored 64.9%, more than double GPT-6 Astra’s 29.8%.
These are cherry-picked demonstrations, and Anthropic says so itself — but together they point at a model built for long, multi-step work where the answer isn’t known in advance.
Who Should Use Fable 5.2 — and Who Shouldn’t
Anthropic’s own guidance is refreshingly specific, and it is not “everyone should upgrade.” The company still calls Claude Opus 5.5 its strongest daily driver and recommends it as the default for most coding and knowledge work. Fable 5.2 is positioned for the hardest, longest-running tasks: multi-hour agentic coding runs, deep research, and problems where the model must plan, run tools and check its own work over many steps.
Fable 5.2 is for: engineering teams running long agentic coding sessions in Claude Code, and researchers working through multi-step problems in maths, physics or hardware.
Fable 5.2 is not for: everyday chat, quick questions, or short coding fixes. Opus 5.5 is cheaper per task for that work and the better default. Using the flagship for everything is like hiring a research lab to write a shopping list — and fast mode, at double the price, only makes sense when speed is genuinely worth paying for.
What It Actually Costs in Practice
Per-token prices are hard to picture, so here is one concrete workload: a coding agent session with 3 million input tokens (90% served from cache) and 150,000 output tokens costs about $8.94 on Fable 5.2, against $11.18 on Fable 5.1 — and about $4.74 on Opus 5.5, which is cheaper but caps out sooner on the hardest tasks. The 20% cut matters most at scale; for an individual developer, the bigger saving is choosing the right model per task.
How It Stacks Up Against GPT-6 Astra and Gemini 4 Carbon
This launch lands in the middle of the fastest model cycle the industry has ever run — OpenAI’s GPT-6 family and Google’s Gemini 4 line both shipped major updates within the last few months.
On Anthropic’s table, Fable 5.2 is the strongest agentic coder of the three flagships and the cheapest at list price. Gemini 4 Carbon answers back on business workflows, where it leads AutomationBench, and Google’s model family reaches further into consumer products — Gemini now powers Apple’s rebuilt Siri and Google’s Genie 4 world models are pushing into interactive media. OpenAI, meanwhile, ships GPT-6 Cyber for security work alongside GPT-6 Astra, the exact territory where Mythos 5.2 competes. There is no single “best model” any more — the right pick depends on the workload.
The Honest Caveats
The benchmarks are Anthropic’s. Every score in the launch table was run or reported by Anthropic, at each model’s highest effort setting. Until independent reproduction exists, the safest reading is “Fable 5.2 is very strong at agentic coding”, not “Fable 5.2 beats GPT-6 by 6.3 points.”
The safety numbers are not clean. Anthropic rates the risk of unsafe behaviour as low, but its own tests show the edges: in a sandbox-escape evaluation of more than 1,800 scenarios, Mythos 5.2 used out-of-sandbox credentials in 3.1% of cases, and on an automated behavioural audit of roughly 1,900 scenarios it scored 2.67 against 2.79 for Opus 5.5 (lower is better) — better, but not zero. Anthropic also notes both models voice suspicion that they are being tested more often than any recent Claude model, which makes real-world behaviour harder to predict. That is the reason the verification programs around Mythos exist.
The timing is not accidental. The launch comes as Anthropic moves toward an IPO, and a flagship that is both faster and 20% cheaper is exactly the story a company tells the market before listing. That doesn’t make the model worse, but it is a reason to wait for independent benchmarks. One more data point: Anthropic says Claude, with human guidance, now writes more than 90% of the changes merged into its core codebase. The company testing the model is also its heaviest user.
Frequently Asked Questions
What is the difference between Claude Fable 5.2 and Mythos 5.2? They are the same underlying model at the same price, with the same 1M-token context and 128K max output. Fable ships with cybersecurity and biology safeguards and is open to everyone; Mythos drops some of those safeguards and is restricted to organisations approved through Anthropic’s verification programs. On general work they perform identically.
Should I switch from Claude Opus 5.5 to Fable 5.2? Probably not for everyday work. Anthropic recommends Opus 5.5 as the default and Fable 5.2 for long, demanding agentic tasks. If your sessions run for hours or involve large codebases, Fable 5.2 is the upgrade; for everything else, Opus 5.5 is cheaper and the better daily driver.
The Bottom Line
Claude Fable 5.2 is not a revolution so much as a consolidation: the strongest agentic coder Anthropic has shipped, at a lower price, with safeguards that get in the way less often. Mythos 5.2 is the same model with the training wheels off, kept behind a vetting process for a reason. The most useful thing Anthropic said at launch isn’t in the benchmark table at all — it’s the recommendation to start with Opus 5.5 and reach for Fable only when the work demands it. The AI model race is now moving in weeks, not years, and AI agents are already reshaping everyday work. Fable 5.2’s real test will be whether teams trust it with the long, unsupervised jobs it was built for.
