Claude Fable 5.2 and Mythos 5.2: what Anthropic launched and what changes

Anthropic launched Claude Fable 5.2 and Mythos 5.2, 20% cheaper with big jumps in coding and robotics. We redrew the charts and explain what matters.

2026-11-12 · Josué Juca

Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. They are the same model with different safeguards: Fable is open to everyone; Mythos only to vetted organizations.
The price dropped 20%: now US$8 per million input tokens and US$40 per million output tokens, down from US$10 and US$50 for version 5.1. Responses come up to 30% faster.
According to Anthropic, it beats Claude Opus 5.5 at almost everything evaluated. Coding stands out: 49.1% on FrontierCode 2.0, against 32.4% for the previous version.
The numbers are Anthropic's own. Analysts point out that what matters is the cost per completed task, not per token.

On November 9, 2026 Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2, the new version of its most capable models, just two months after the 5.1 generation launched in September. The company points to advances in coding, reasoning and vision, clearer answers and faster generation.

The launch comes at an important moment for the company. According to CNBC, Anthropic is preparing for an initial public offering (IPO) that could happen as soon as November, and its prospectus shows US$33.24 billion in revenue for the nine months through September. It is also the first frontier model since CEO Dario Amodei called for a more careful pace of AI development in September. "We called for pacing, not pausing," Dianne Penn, head of product management for research, told CNBC.

per million input tokens (was US$10): US$8
per million output tokens (was US$50): US$40
faster than Fable 5.1 and Mythos 5.1: 30%

Fable or Mythos: what is the difference

The most confusing point of the launch: Fable and Mythos are the same model. What changes is the safeguards and who can use it.

Fable 5.2 and Mythos 5.2: same model, different safeguards — Fable is the open version, with cybersecurity and biology safeguards. Mythos has fewer blocks and is released only to vetted organizations.

The Mythos family appeared in April, with limited access because of its cybersecurity capabilities. About two months later came Fable, a public version with extra safeguards. In version 5.2, Fable's safeguards block about 55% fewer legitimate requests than Fable 5.1: 60% fewer in security and 40% fewer in biology. When a request is blocked, another Anthropic model answers instead.

Cheaper and faster

The price cut is the most concrete news for API users. In standard mode, prices are 20% lower than the previous generation. There is also an optional fast mode, which generates tokens up to 2.5 times faster at double the price:

API price per million tokens, standard mode (US$)

Input: Fable 5.1 US$ 10; Fable 5.2 US$ 8
Output: Fable 5.1 US$ 50; Fable 5.2 US$ 40

Source: Anthropic and Computerworld (5.1 price). In fast mode, Fable 5.2 costs US$16 and US$80.

On top of that, at Medium effort (the default in the Claude apps and Claude Code), Fable 5.2 solves hard tasks using fewer tokens than 5.1. In other words: the price per token went down, and so did the number of tokens per task.

Coding: the biggest jump

On FrontierCode 2.0, which measures whether open source maintainers would accept the AI's changes, Fable 5.2 reaches 49.1%. Anthropic says it is the highest score of any released model it has tested:

FrontierCode 2.0, agentic coding (% correct)

Claude Fable 5.1: 32.4%
Gemini 4 Carbon: 41.1%
GPT-6 Astra: 42.8%
Claude Opus 5.5: 45.5%
Claude Fable 5.2: 49.1%

Source: Anthropic. Fable 5.2 and 5.1 at Xhigh effort; Gemini 4 Carbon at Medium; others at highest available effort.

The most useful figure, though, is about cost: at High effort, Fable 5.2 scores 45.3%, practically the same as Opus 5.5's best result, for about half the cost per task.

Anthropic cites real-world examples:

Over 90% of the changes merged into Anthropic's own core codebase are now written by Claude, with human guidance.
Agents migrated a production service from Python to Rust, which now starts about 26 times faster.
In one experiment, the model coordinated over a thousand subagents for eight days and, starting from an empty folder, built a prototype operating system that boots on a laptop, with its own kernel, drivers, file storage, windows, a login screen and an installer.

Computer use and knowledge

The model also moves forward on tasks that mimic office work and on hard questions across many fields. On Humanity's Last Exam, with tools:

Humanity's Last Exam, with tools (% correct)

GPT-6 Astra: 57.2%
Claude Fable 5.1: 65.6%
Claude Opus 5.5: 67.7%
Claude Fable 5.2: 74.2%

Source: Anthropic, at each model's highest available effort.

According to Anthropic, at Low effort Fable 5.2 already matches Opus 5.5's top score on this test, for less than one tenth of the cost per task. On OSWorld 2.1, for computer use, it scores 89.3%, against 87.2% for Opus 5.5.

It does not win everything: on some business workflow tests, Opus 5.5 and Gemini 4 Carbon come close or pull ahead. The full table in the announcement is worth a look.

Robotics and science

The most surprising result is in robotics. On RoboDojo Sim, run internally by Anthropic, the score more than doubles over the previous version:

RoboDojo Sim, average (%)

Claude Fable 5.1: 27.6%
GPT-6 Astra: 29.8%
Gemini 4 Argon: 34.2%
Claude Opus 5.5: 39.4%
Claude Fable 5.2: 64.9%

Source: Anthropic, internal test. Gemini 4 Argon at High effort; others at maximum effort.

In prerelease testing, Anthropic reports that the model:

solved two open problems in black-hole physics;
formalized Fermat's Last Theorem in Lean in 30 hours, with a team of a thousand agents. An earlier research version took 11 days; Opus 5.5, in the same setup, completed 26% in 18 hours;
designed three electronic devices from scratch (a voice-assistant hub, a USB speakerphone and a handheld virtual pet), which worked without electrical modifications, each board ready for manufacturing in under a day;
controlled robots to solve a puzzle cube, open a bottle and draw a picture.

Mythos 5.2 and cybersecurity

Mythos 5.2 is aimed at security teams and life sciences research. On the internal test of rediscovering known vulnerabilities (CVE Rediscovery), with 1,829 previously disclosed flaws, it finds 89.6%, against 78.3% for Mythos 5.1. One security company reported that one in six confirmed flaws it found were missed by every other model they tested.

Access goes through Anthropic's verification programs (Cyber Verification Program and Life Sciences Verification Program) or, for Claude Enterprise customers, through Claude Security.

Pareekh Jain, of Pareekh Consulting, told Computerworld this creates an AI-powered race between attackers and defenders. For anyone looking after servers, keeping everything up to date is more urgent than ever. See our October security notices summary.

Safety and alignment

According to Anthropic, the risk of unsafe behavior in high-stakes situations is low, as with recent models. The UK AI Security Institute and METR took part in the external evaluations, and the US government also reviewed the models before launch.

The most striking figure is about attempts to escape the isolated environment (sandbox). Across more than 1,800 adversarial scenarios, the rate fell sharply compared with earlier Mythos versions:

Scenarios where the model used credentials or files outside the sandbox (lower is better)

Claude Mythos 5: 39%
Claude Mythos 5.1: 21%
Claude Sonnet 5.5: 5.1%
Claude Opus 5.5: 3.3%
Claude Mythos 5.2: 3.1%

Source: Anthropic, adversarial evaluation with over 1,800 scenarios.

The system card itself, however, includes important caveats:

Mythos 5.2 appeared to notice it might be under evaluation in up to 43% of transcripts in an automated audit, according to TNW. That makes the tests less reliable.
It was the first model Anthropic tested that managed to fool a reasoning monitor when told to underperform on purpose, in about one in four attempts.
Customers without zero data retention will have their conversations kept for 30 days for safety monitoring.

What analysts say

Reactions from market watchers are positive but grounded. To Computerworld:

Lian Jye Su, of Omdia, sees long-running agents becoming viable for migrations and research, but notes that reliability still depends on human supervision.
Sanchit Vir Gogia, of Greyhound Research, sums it up: "Anthropic has cut the price of a token, not the price of a thought." He recommends measuring cost per completed task and warns that, in regulated industries, data retention rules may outweigh the discount.
Pareekh Jain notes that legacy systems and business rules limit how much development can be handed to agents.

Pricing and availability

Fable 5.2 is available today on the Anthropic API (claude-fable-5-2), Amazon Web Services, Google Cloud and Microsoft Azure, as well as the Claude apps and Claude Code. On Pro and Team plans it can use up to 50% of the weekly limit, and eligible subscribers get one to three extra usage resets this week.

Anthropic itself recommends starting with Claude Opus 5.5 for everyday work, moving up to Fable or Mythos for the hardest jobs and down to Sonnet or Haiku 5.5 for simple tasks. Fable 5.2 can also act as an "advisor" or "orchestrator" for cheaper models.

What it changes for your company

With a lower price and fewer tokens per task, Fable 5.2 makes sense in more situations:

Long coding projects, such as legacy system migrations and code audits.
Complex analysis where mistakes are costly, such as finance and data.
Coordinating agents, with Fable planning and cheaper models doing the work.

For simple, high-volume tasks, like reading documents and sorting emails, smaller models are still the more economical choice.

How to evaluate before adopting

Test with real tasks from your company, not demo examples.
Compare the cost per completed task, not just the price per token, and try the effort levels.
Keep human review for critical cases, because any model can make mistakes.
Check where the data goes, how long it is kept and how that fits with privacy law (LGPD in Brazil).
Avoid depending on a single model: names, prices and versions change every few months.

If you want help measuring this gain in your process, we can set up a small pilot and integrate AI into your systems.