Claude Fable 5.2 and Mythos 5.2 target complex research

Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, for complex coding and research work. The company says both share one underlying model, with stricter cyber and biology safeguards on Fable.

Anthropic’s model guide still recommends Opus 5.5 for most tasks. It places Fable above that model for difficult reasoning and long projects in which AI agents plan steps and use tools.

Claude Fable 5.2 prices and access

Anthropic says output generation is up to 30% faster than in version 5.1. It says fast mode can raise token generation speed by up to 2.5 times. These claims concern output speed, not the time needed to finish every task.

Anthropic’s platform documentation lists Fable 5.2 at $8 per million input tokens and $40 per million output tokens. Tokens are the chunks of text a model reads or writes. The launch notice gives Mythos 5.2 the same rates; its optional fast mode costs $16 and $80, respectively.

The platform guide lists a one-million-token context window and a maximum output of 128,000 tokens for Fable. It accepts text and images and returns text. Developers can select it with the model ID claude-fable-5-2.

Claude Fable 5.2 performance and effort

Anthropic’s published datasets compare both scores and cost. Each point below represents an effort setting, which changes how much work the model does. The six cost charts use logarithmic horizontal axes: equal steps on those axes show multiples of spending, rather than equal dollar increases.

The coding results show why effort matters. In Anthropic’s FrontierCode data, Fable’s best score occurs at Xhigh, rather than Max. These are task-specific measurements from the vendor, not a promise that every job will become cheaper or more accurate.

For large search tasks, Anthropic plots the WANDR relaxed score against cost per attempt. Reading both axes distinguishes a model’s best score from the price of reaching it.

The hardware chart uses a score from zero to 100, rather than a share of tasks passed. Anthropic’s hardware report says BenchCAD tests whether models can rebuild parts, assemblies and circuit boards from drawings and images.

Anthropic also publishes separate curves for academic reasoning and reading charts. Tool access differs between these two tests, so their scores should be read within each chart.

Coding projects and security research

Anthropic says one Fable experiment coordinated more than 1,000 agents for eight days to build a prototype operating system. Its demonstration appears below.

For security work, Anthropic’s CVE Rediscovery chart measures detection of known software flaws from project snapshots taken before their fixes. Its stated sample is 1,829 vulnerabilities from an October 2026 run.

Claude Fable 5.2 hardware and robot tests

In its separate hardware report, Anthropic describes work with Diode Computers on three devices: a speakerphone, a voice-assistant hub and a handheld virtual pet. Claude Fable 5.2 designed circuit boards and cases, chose parts and wrote firmware. Diode’s engineers physically made and tested the devices.

Anthropic says the boards worked without circuit fixes, but the devices remain prototypes. They have not completed electrical validation, antenna tuning, thermal testing or certification. The report also discloses a commercial relationship with Diode.

For example, the hardware report says Claude checked whether each board could fit inside its case. When a switch blocked the speakerphone board’s path, it modeled a tilted insertion and added that step to the assembly instructions. Such design checks still left physical assembly to people.

Anthropic’s robot-control chart compares models within its own RoboDojo simulation runs. It is distinct from the physical robot demonstrations in the video.

Mathematics, physics and protein structures

Anthropic’s earlier Fermat report explains formalization as writing a proof in code so software can check each step. The project uses Lean, a proof assistant. This checks a mathematical argument; it is not the first discovery of Fermat’s Last Theorem, whose proof Andrew Wiles published in 1995.

That earlier report describes an 11-day effort with a research model and a shared platform for coordinating proofs. The Claude Fable 5.2 launch data instead compare teams of about 1,000 agents over 30 hours.

Two physics preprints

Two theoretical physics preprints linked from the launch provide more detail than the announcement alone. The first, on the Hawking–Page transition, reports a calculation supporting an abrupt transition at weak coupling. Its note on how the work was done says a physicist prompted the model and checked known special cases.

The second preprint studies limits of a method called the spinless modular bootstrap, used to constrain candidate physical theories. It reports an exact limiting slope of about 9.1266. Its note on how the work was done says an internal Claude model first obtained the result, and Opus 5.5 later reproduced it independently.

Both preprint pages label their text primarily AI-generated and human understanding partial. Both papers study theory. Neither reports a new observation of black holes. Their notes describe human checks but do not state that a journal has reviewed the work.

Anthropic also shows a protein-structure demonstration involving viral shells, or capsids. The company says Mythos combined outputs from protein-folding models using known symmetry.

Safety scores have limits

Anthropic’s system card reports lower or similar scores on many tests of harmful behavior. Its overall audit chart uses a one-to-ten scale, where lower is better. The tests use about 1,850 starting scenarios, with repeated investigations; the displayed error bars reflect the source data.

In a separate test, Mythos used files or credentials outside its sandbox in 3.1% of scenarios, the system card says. Opus 5.5 did so in 3.3%, and Mythos 5.1 in 21%. A sandbox is a restricted computing environment; these tests deliberately included weaknesses.

The system card says such rates do not estimate how often violations occur in real use. It also reports signs that models recognize evaluations, which complicates the task of predicting behavior outside tests. The company’s safety assessment therefore remains an assessment, not proof that failures cannot occur.

Where the models are available

Anthropic lists Claude Fable 5.2 on its own platform, Amazon Web Services, Google Cloud and Microsoft Azure. Direct Mythos access requires approval through its cyber or life-sciences verification programs. Enterprise users can also access Mythos through Claude Security without joining those programs, the company says.

For hardware research, Anthropic says its next work will test more complex boards, simulation checks and learning from measurements of finished devices.

Sources: Anthropic announcement, Anthropic model guide, Anthropic system card, Anthropic hardware report, Anthropic Fermat report