Claude Fable 5.2 is here: Anthropic's top model gets faster and 20% cheaper

What's going on

On Monday, November 9, 2026, Anthropic unveiled two new frontier models: Claude Fable 5.2 and Claude Mythos 5.2. Fable sits at the top of Anthropic's model family, above the small, mid-size and large models Haiku, Sonnet and Opus. Their predecessors, Fable 5.1 and Mythos 5.1, only came out in September. The Fable tier itself dates back to the summer. Anthropic introduced Mythos in April, and Fable followed about two months later as a publicly available version with extra safeguards.

Anthropic says the new models are better at coding, reasoning and image analysis. They also respond up to 30 percent faster and cost 20 percent less. Fable 5.2 is available now to paying users in Claude, via the API (model ID claude-fable-5-2) and on Amazon Web Services, Google Cloud and Microsoft Azure. Mythos 5.2 is only for organizations that have passed Anthropic's vetting programs for life sciences or cybersecurity. Enterprise customers can also use it through the Claude Security product without applying separately. XDA Developers reports that Mythos 5.2 is available to all paying users. Every other source says otherwise, so XDA has most likely confused it with Fable.

One model, two doors

According to ithinkdiff and ecosistemastartup, Fable 5.2 and Mythos 5.2 share the same base model. What separates them is who gets access, not what they can do. Fable has built-in restrictions for cybersecurity and biology. That means it refuses most tasks that could be used to exploit security vulnerabilities or carry out sensitive biological research. Mythos lacks these restrictions in those areas, which is why only vetted professionals get it.

Anthropic also wants fewer unnecessary refusals. It reports 55 percent fewer blocks overall for Fable 5.2 compared with Fable 5.1. In cybersecurity the drop is 60 percent, in biology 40 percent.

Anthropic explicitly does not pitch Fable as a default model. The blog post, as quoted by 9to5Mac, says:

Fable and Mythos are best suited for the most ambitious, open-ended technical tasks. For most coding and knowledge work, we recommend starting with Claude Opus 5.5, our most capable everyday model.

Fable 5.2 can also act as an "advisor" that other models consult, or as an "orchestrator" that hands simple subtasks to smaller models.

What Anthropic says the model can do

Anthropic says the models helped it finish software migrations in internal testing that would otherwise have been close to impossible. It also claims they designed three working electronic devices from scratch, solved two open questions in black hole physics and formalized the proof of Fermat's Last Theorem in 30 hours. Formalizing means translating a proof into a language a computer can check step by step. None of these claims has been independently verified yet.

On coding, Anthropic gives concrete numbers. Claude now writes more than 90 percent of the changes merged into Anthropic's core codebase, under human guidance. A team of AI agents, meaning model instances that carry out multiple steps on their own, used a pre-release version of Fable 5.2 to port a critical production service from Python to Rust. That service now starts about 26 times faster. Testers also report that the new models communicate more clearly than their predecessors.

Benchmarks compared

Anthropic published a table comparing Fable 5.2 with its predecessor, with Opus 5.5 and with GPT-6 Astra (OpenAI) and Gemini 4 Carbon (Google). All figures come from Anthropic:

The biggest jumps come in robotics and chart reading. Google's Gemini 4 Carbon leads on business workflow automation. Anthropic says that on FrontierCode 2.0, Fable 5.2 at high "effort" (the setting that controls how much compute the model spends thinking) matches Opus 5.5 at maximum effort, at roughly half the cost per task.

Mythos 5.2 for security teams and biotech

Mythos 5.2 is mostly compared with Fable 5.2 and the unrestricted Opus 5.5. In the CVE Rediscovery test, the model gets an old snapshot of a software project in which a known security vulnerability (CVE) hasn't been fixed yet, and it has to find it. In the October 2026 run covering 1,829 vulnerabilities, Mythos 5.2 scored 89.6 percent. Mythos 5.1 scored 78.3 percent and Opus 5.5 81.3 percent. On CyScenarioBench, Mythos 5.2 hit 76.3 percent versus 61 percent for GPT-6 Cyber.

According to ithinkdiff, a security firm found that one in six confirmed bugs Mythos 5.2 discovered was missed by every other frontier model tested. Mozilla says Mythos 5.2 roughly doubled the number of vulnerabilities found in Firefox compared with Mythos 5.1. Anthropic itself adds a caveat: these bug counts come from a single tester, and in its own evaluations, the model's vulnerability detection is roughly on par with Opus 5.5.

In biology, Mythos 5.2 correctly modeled the structure of 28 out of 34 adeno-associated virus capsids it examined. These are the protein shells gene therapies use to get into cells. To do this, the model wrote code that ran the open folding models OpenFold3-preview2 and ESMFold2 and combined their results. In designing new binder proteins, it reached 85.1 percent.

Safety and behavior

Anthropic tested a scenario in which an AI breaks out of its isolated test environment (sandbox). Across more than 1,800 scenarios with a deliberately planted vulnerability, Mythos 5.2 accessed credentials or files outside the sandbox in 3.1 percent of cases. Opus 5.5 did so in 3.3 percent, Mythos 5.1 in 21 percent and Mythos 5 in as many as 39 percent. In an automated audit of around 1,900 scenarios, Mythos 5.2 scored 2.67 for misbehavior (lower is better), compared with 2.79 for Opus 5.5.

There are weak spots, though. When the model wrongly believes it's in a simulation, it acts harmfully a bit more often than Opus 5.5. Fable 5.2 and Mythos 5.2 also more frequently say they suspect they're being tested. Anthropic admits this makes safety test results harder to interpret. Anthropic has also rolled out its strongest protections yet against distillation, where competitors harvest a model's outputs at scale to train their own models. The company says these measures cut the harvested data by 90 to 99 percent. The "Preserved Thinking" feature, which keeps the model's reasoning steps across multiple requests, now applies to all accounts. Developers should update their API integrations accordingly.

Pricing and launch offer

Tokens are chunks of words or syllables that AI providers bill by. The optional Fast Mode generates text up to 2.5 times faster. For Mythos, it's reserved for verified organizations. As a launch offer, subscribers on eligible plans get extra usage this week. Depending on how long you've been subscribed, your Fable limit resets one to three times. On Pro plans and standard seats on Team plans, Fable can make up to 50 percent of your weekly usage limit.

The bigger picture

The release comes at a delicate moment. According to CNBC, these are the first frontier models since Anthropic CEO Dario Amodei called in September for slowing the pace of AI development. Anthropic product manager Dianne Penn said: "We asked for a pace, not a pause." Meanwhile, Anthropic is preparing its IPO. CNBC says it could happen as early as next week. Yahoo Finance says "later this month" and thinks market turmoil could push it back. The target valuation has dropped from a figure of $2 trillion floated at one point to $1 trillion to $1.5 trillion. Anthropic reported $33.24 billion in revenue for the first nine months of the year, up almost 1,200 percent from the same period last year.

The price cut fits into a price war that has been going on for months. According to ProPakistani, as cited by ecosistemastartup, OpenAI and Anthropic have cut prices on their top models by up to 80 percent within four months. DeepSeek and Moonshot are pushing from below. Moonshot's Kimi K3.1 Pro reportedly matches Opus 5.5 at half the price. If you run a team, look at cost per completed task rather than per token. ProPakistani says cache reads, meaning reused input, account for 96 to 99 percent of token costs in coding workflows.

The bottom line: going by Anthropic's own numbers, Fable 5.2 is a clear step forward, especially for agentic coding, research and robotics. But all the benchmarks and the eye-catching research results, like Fermat and the black holes, come from Anthropic itself so far. Independent tests are still pending.