On November 9, Anthropic introduced Claude Fable 5.2 together with its sibling model Claude Mythos 5.2. According to the company, both are built on the same underlying model: Fable ships with cybersecurity and biology safeguards and is available to all users, while Mythos is reserved for vetted organizations in security and life-sciences research. Anthropic promises better results in coding, reasoning and vision, output up to 30 percent faster than version 5.1, and a lower price. Most of the numbers behind those claims, however, come from Anthropic’s own testing – here is a closer look.
Claude Fable 5.2: what Anthropic announced
In its official announcement, Anthropic describes Fable and Mythos as models for “the most daunting technical challenges.” For most coding and knowledge tasks, the company explicitly still recommends Claude Opus 5.5 as the everyday model. Only when a task is very large or open-ended should users move up to Fable or Mythos – or down to Sonnet 5.5 and Haiku 5.5 for simpler work. Fable 5.2 can also act as an “Advisor” that smaller models consult on hard problems, or as an “Orchestrator” that delegates simpler subtasks to smaller models.
The key specs according to the announcement and the model documentation:
- Price: $8 per million input tokens and $40 per million output tokens. Anthropic calls this a lower price than Fable 5.1.
- Fast mode: optional, up to 2.5 times faster token generation for $16 and $80 per million tokens respectively.
- Context: a 1 million token context window and up to 128,000 output tokens.
- Thinking: adaptive thinking is always on. The API documentation lists “high” as the default effort; in the Claude apps and Claude Code, the announcement says “Medium” is the default.
- Knowledge cutoff: June 2026.
- Availability: available now on the Claude Platform (model ID
claude-fable-5-2) as well as via Amazon Web Services, Google Cloud and Microsoft Azure.
According to the documentation, Mythos 5.2 has the same specifications and pricing as Fable 5.2 but is only accessible through Anthropic’s Cyber and Life Sciences Verification Programs. Enterprise customers can also use Mythos 5.2 indirectly through Claude Security, a service that scans codebases for vulnerabilities.
The benchmarks – and what they tell us
Anthropic published an extensive comparison table covering Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini 4 Carbon. Some of the figures (each model’s highest available effort unless noted otherwise):
- FrontierCode 2.0 (agentic coding): Fable 5.2 scores 49.1 percent, compared with 32.4 for Fable 5.1, 45.5 for Opus 5.5, 42.8 for GPT-6 Astra and 41.1 percent for Gemini 4 Carbon.
- FrontierSWE v2: 69.9 percent versus 65.5 for GPT-6 Astra and 61.2 for Gemini 4 Carbon.
- AutomationBench (business workflows): 51.4 percent – here Gemini 4 Carbon leads with 53.9 percent.
- OSWorld 2.1 (computer use, offline subset): 89.3 percent.
- Humanity’s Last Exam with tools: 74.2 percent versus 57.2 for GPT-6 Astra.
- RoboDojo Sim (robot control): 64.9 percent, more than double Fable 5.1’s 27.6 percent.
The table also shows where Fable 5.2 does not lead clearly. Besides AutomationBench, the gap on the chart-reading test “Chartography” is minimal: Fable 5.2 scores 73.4 and Gemini 4 Carbon 72.5 percent. On the AA-Briefcase knowledge-work test, Opus 5.5 even beats Fable 5.2 on the Elo rating, 1,800 to 1,785.
An important caveat: Anthropic ran these numbers itself. RoboDojo Sim, for example, was run internally by its own account, and the competitors’ robotics scores come from Anthropic’s own runs. For the CyScenarioBench cybersecurity test, the company notes that each lab measures on its own unique tasks, so the scores are only partly comparable. Independent measurements were not available when this article was published.
Efficiency, not just top scores
The cost figures are more interesting than the peak scores. On FrontierCode 2.0, Anthropic says Fable 5.2 at “High” effort reaches 45.3 percent, roughly matching Opus 5.5 at maximum effort (45.5 percent) for about half the cost per task. On Humanity’s Last Exam, Fable 5.2 is said to match Opus 5.5’s best score even at the lowest effort level, for less than a tenth of the cost. Anyone running these models in production should verify this on their own workloads: cost per task depends heavily on how many tokens a model spends thinking.
Mythos 5.2 for security research
Anthropic calls Mythos 5.2 its most capable model for cybersecurity. On the internal “CVE Rediscovery” evaluation, in which the model has to find published vulnerabilities in a snapshot of a project taken before the fix, Mythos 5.2 reaches 89.6 percent. According to Anthropic, the October 2026 run covered 1,829 CVEs. Mythos 5.2 found 10.2 percent more vulnerabilities than Opus 5.5 and 14.4 percent more than Mythos 5.1.
Anthropic adds its own qualifier, though: a footnote says that on its own evaluations, vulnerability identification is “roughly on par” with Opus 5.5. Fable 5.2 blocks most security-related tasks anyway – penetration testing, exploit generation and binary-based vulnerability work are caught by classifiers. Blocked requests are then answered by an older model (Opus 4.8 for cybersecurity, Opus 5 for biology). Compared with Fable 5.1, early testing showed about 55 percent fewer unnecessary blocks.
Google is taking a similar route with Gemini 4 Carbon, which initially went only to cyber defense teams: the strongest model reaches selected defenders first.
Examples from the announcement
Anthropic illustrates the model’s capabilities with a series of projects that cannot be independently verified but show where things are heading:
- A team of around 1,000 Fable 5.2 agents reportedly formalized the proof of Fermat’s Last Theorem in the Lean proof language in 30 hours. Opus 5.5 completed only 26 percent after 18 hours in the same setup.
- According to Anthropic, agents running an earlier version of Fable 5.2 migrated a production service from Python to Rust; it now starts about 26 times faster.
- One engineer had the model build a prototype operating system from an empty folder; over eight days, Fable 5.2 coordinated more than 1,000 subagents.
- Together with circuit board maker Diode, the model designed three electronic devices including boards, enclosures and software, which Anthropic says worked without electrical modifications.
- In theoretical physics, Fable 5.2 is said to have solved two open problems concerning black holes; the results have been published as preprints.
The announcement also includes quotes from early testers such as Vercel, Shopify, Datadog, Uber, Snowflake and Mozilla. Mozilla, for instance, says Mythos 5.2 “roughly doubled” Firefox vulnerability discovery metrics compared with Mythos 5.1.
Safety, sandboxing and anti-distillation
According to Anthropic, the model was tested before release by external evaluators including the UK AI Security Institute and METR. In a test with more than 1,800 scenarios in which the sandbox had a deliberate weakness, Mythos 5.2 used credentials or files outside the sandbox in 3.1 percent of cases – about as often as Opus 5.5 (3.3 percent) and far less often than Mythos 5 and 5.1 at 39 and 21 percent. At the same time, Anthropic concedes that Fable 5.2 and Mythos 5.2 voice suspicion that they are being evaluated more often than earlier models, which makes it harder to judge how they will behave in real-world use.
What is new is stronger protection against distillation – harvesting a model’s outputs to train other models. The “preserved thinking” API feature, which prevents prior reasoning in the context from being edited, now applies to all accounts, according to Anthropic. Developers should check whether their integrations are affected. For enterprises with zero data retention, “Enterprise Frontier Safeguards” is available in private beta; other customers are subject to 30-day retention for safety monitoring.
Our take
Claude Fable 5.2 is not an everyday model – Anthropic says so itself. For developers, what matters most is whether the better ratio of results to cost holds up on their own tasks, and whether always-on thinking ends up driving the bill up after all. The benchmark table is impressive, but it comes entirely from the vendor. How solid such comparisons are was shown recently by the GDPx benchmark from SemiAnalysis for frontier agents: independent measurements often produce different rankings.
The release cadence stands out: Fable 5.1 arrived on September 1, version 5.2 a little over two months later. Anthropic is also pairing the launch with extra Fable usage for paying subscribers this week. As with Claude Personal Finance, Anthropic keeps broadening its offering – while the most powerful security capabilities stay behind verification programs.
