Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026. The headline is a model built for unusually hard, open-ended technical work; the more interesting evidence is in what Anthropic says it did with a real circuit board and a hard protein-structure problem. This guide pairs three videos from Claude's official channel with the underlying research notes, benchmarks and access rules. It also separates a working prototype from a finished product, and a structure prediction from a treatment.
Fable 5.2 is broadly available; Mythos 5.2 uses the same underlying model but is restricted to verified cybersecurity and life-sciences work. Anthropic lists Fable at $8 per million input and $40 per million output tokens through its API. It recommends Opus 5.5 for most daily coding and knowledge work, reserving Fable for the hardest projects. The hardware and capsid demonstrations are valuable evidence of range, but each relied on specialist tools, people or existing models.
Watch Anthropic Introduce Fable 5.2
Video: Claude's official YouTube channel, published 9 November 2026. The linked Anthropic release is the source for the model claims in this article. The two deeper videos below are also from Claude's channel.
Where the New Model Is Stronger, and Where the Table Is Mixed
Anthropic reports improvements in coding, reasoning and vision. Its launch evaluation table shows Fable 5.2 ahead of Opus 5.5 on many technical tests, though scores use different effort settings and some evaluations are internal or partial. These are vendor-reported scores, not measurements reproduced by JQ AI SYSTEMS.
| Evaluation | Fable 5.2 | Opus 5.5 | Read it carefully |
|---|---|---|---|
| FrontierCode 2.0 | 49.1% at Xhigh | 45.5% | An agentic coding test, not a guarantee that every repository task improves. |
| FrontierSWE v2 | 69.9% | 62.3% | A software-engineering gain in Anthropic's table. |
| Remote Labor Index | 41.3% at High | 21.3% | A large difference on the reported knowledge-work tasks, not an employment forecast. |
| OSWorld 2.1, partial offline set | 89.3% | 87.2% | Only a 2.1-point lead on this computer-use evaluation. |
| AA-Briefcase v1.1 Elo | 1785 | 1800 | Opus still leads this view of a knowledge-work evaluation. |
The release is not an across-the-board win against every competitor either: Anthropic's AutomationBench row has Gemini 4 Carbon at 53.9% and Fable at 51.4%. More importantly, benchmark quality must be weighed against task cost and the review needed to accept an answer. Anthropic reports that Fable at High effort roughly matched Opus at Max on FrontierCode 2.0 for about half the cost per task in its test. That is a test-specific efficiency result, not a blanket promise of cheaper Fable usage.
The Working Voice-Assistant Hub
Anthropic and Diode Computers asked Fable 5.2 to design three devices from short prompts: a speakerphone, a voice-assistant hub and a handheld virtual pet. For the hub, Anthropic says Claude wrote specifications, chose parts, routed the four-layer circuit board, modeled the enclosure in CAD and wrote firmware. Diode then fabricated, assembled and physically tested the board. Anthropic reports that all three prototypes worked without electrical modifications. This is far more concrete than a rendered product concept, but human engineers and manufacturing equipment were essential to completing the job.
The electronics case study explains why the demonstration matters. Claude checked component datasheets, ran design-for-manufacturing and sourcing checks, resolved flagged solder-mask geometry, and wrote a custom routing algorithm when the available PCB router was too laborious. Diode says it built the boards directly from Claude's files without a manufacturing-change round. The work illustrates linked electrical, geometric and software reasoning rather than a single impressive image.
The limit: Anthropic calls these early prototypes. They have not passed electrical validation, antenna tuning, thermal testing or certification, and they were not designed for mass production. Anthropic also discloses a commercial relationship with Diode. The video should therefore be read as a documented collaborator-assisted prototype, not as independent proof that Fable can ship a consumer device alone.
The AAV Capsid Prediction: A Different Kind of Engineering
An adeno-associated virus (AAV) has a protein shell called a capsid. Its geometry matters for gene-therapy research, but an AAV capsid is a large symmetric assembly, not just one protein chain. Anthropic says Mythos 5.2 wrote code that used two existing open-source folding models on overlapping groups of chains, then used known capsid symmetry to assemble a prediction of the full structure. The video describes groups of five neighboring chains and a 60-copy shell. The key contribution is the computational strategy that made those established predictors useful on the full assembly, not a new foundation folding model created from scratch.
In Anthropic's reported tests, the method accurately modeled 28 of 34 natural and engineered capsids under its 3-angstrom cutoff. For the AAV2 reference, Anthropic reports 1.2 angstrom C-alpha RMSD using OpenFold3-preview2 and 1.4 angstrom using ESMFold2. Those are aggregate structural error measures against an experimental reference, not a statement that every atom was placed within one atom's width. Anthropic reports approximately 58 minutes of inference on one H200 GPU for the OpenFold3-preview2 route and seven minutes for ESMFold2 on that AAV2 example. See the original OpenFold3 and Biohub ESMFold2 projects for the underlying tools.
The limit: a predicted protein structure is not a working gene therapy. It does not show delivery, immune response, efficacy or safety in patients. Anthropic's launch article describes the computational test; readers should not turn its promising geometry into a clinical claim. Mythos 5.2 access is also restricted, so ordinary Fable users should not assume they can reproduce this biology workflow.
Other Results Worth Reading, Without Flattening the Caveats
The same release describes an eight-day prototype operating-system build involving more than 1,000 subagents, robotics demonstrations, and a 30-hour formalization of Fermat's Last Theorem using roughly 1,000 agents. Those are ambitious, heavily tooled experiments, not ordinary one-chat outputs. The physics section links to two research preprints about black-hole-related problems. The first preprint's provenance says a physicist prompted and checked the model's work against known special cases. That provenance matters more than a headline saying AI “solved physics.”
How to Use Fable 5.2, and When Not To
Anthropic says Fable 5.2 is available across Claude products and developer platforms, including AWS, Google Cloud and Microsoft Azure. The API model ID is claude-fable-5-2. Standard API pricing is $8 per million input tokens and $40 per million output tokens; optional fast mode is $16/$80 and is advertised as generating output up to 2.5 times faster. Those are API rates, not subscription prices. Anthropic's temporary launch promotion offers eligible paid subscribers extra Fable usage or resets, with different terms and expiry dates by plan.
For a sensible first test, choose one difficult task with an external check: a failing integration test, a circuit or CAD design an engineer can review, or a research synthesis with sources that can be verified. Run Opus 5.5 as the baseline, then Fable 5.2 with the same tools, acceptance criteria and review budget. Record total tokens, wall time, corrections and the final validated outcome. If a task is routine, Anthropic's own recommendation is to start with Opus. If it is advanced cybersecurity or biomedical work requiring Mythos, follow the Cyber Verification Program or Life Sciences Verification Program rather than assuming that a paid Claude plan alone grants Mythos access.
Primary Sources and Research Tools
Anthropic's 9 November launch is the central source for access, pricing and evaluations. Its system card documents methodology and safety testing. The electronics write-up details Diode's role, design process, prototypes and commercial relationship. The OpenFold3 and ESMFold2 repositories identify the two pre-existing folding systems behind the capsid demonstration. The three embedded videos are official Claude uploads, not independent replications. Source check: 10 November 2026.