
Claude Fable 5.2 review: when the harder task is worth it
What Claude Fable 5.2 does well, who needs it, API costs and a brief comparison with Opus, GPT and Gemini. With official images. Updated November 14, 2026.
A new tier for demanding technical work
Anthropic released Claude Fable 5.2 on November 9, 2026. It is aimed at demanding reasoning and agent sessions that can run for hours: the sort of work where a model must inspect a system, try an approach, notice a failure and continue. That makes it relevant to developers, researchers and people building ambitious projects through vibecoding. [1] [2]
Fable 5.2 and Mythos 5.2 share the underlying model, with different safeguards and access. Fable is the broadly available version; Mythos requires verification for specialized work. Cybersecurity or biology results labeled Mythos should therefore stay attached to Mythos. This review covers Fable, using documentation and published evaluations checked on November 14. PAS7 has not run its own inference tests. [1] [6]
Fable accepts text and images and produces text. Its context window is 1 million tokens, with up to 128,000 output tokens. Those limits let an application supply substantial material, but they do not guarantee perfect recall or let the model operate a computer without an agent and connected tools. [2]
Where it could help a vibecoding project
A difficult repository change rarely fits in one generated function. An agent may have to read several packages, reproduce a bug, change code, run a build and repair a regression. Fable’s intended role is to handle that extended reasoning and tool loop. Developers working on migrations, unfamiliar codebases or interdependent features are more plausible beneficiaries than someone asking for a button’s CSS. [2]
Our interpretation is that Fable is worth shortlisting when repeated corrections and supervision dominate the work. For a hobby app, that might be an import feature touching the UI, validation, persistence and tests. For a team, it could be a migration whose acceptance criteria span multiple services. These are proposed use cases, not projects we have completed with Fable.
Access is through Claude apps and Claude Code where the account includes Fable, or through the API and supported cloud platforms. In the API the identifier is claude-fable-5-2. Plan limits and platform availability still apply. You are choosing both a model and the environment that gives it files, tools and permissions. [1] [6]
A concrete example: software that becomes hardware
Anthropic and Diode published an experiment in which Fable designed a handheld virtual pet, a voice-assistant hub and a USB speakerphone. The work connected circuit layouts, enclosure geometry and firmware; human engineers manufactured and assembled the devices. This is useful evidence of coordinating constraints across tools. [3]
The photograph shows prototypes, rather than consumer products you can buy. Anthropic says they have not completed electrical validation, antenna tuning, thermal testing or certification. An engineer still owns those steps. That boundary matters when a striking demo is used to infer production readiness. [3]
Anthropic / Diode prototypes. The original English labels name the handheld pet, home voice-assistant hub and USB speakerphone. Humans built and tested the devices; this is not PAS7 hardware. [3]
Section hardware-example screenshotA short comparison with other frontier models
Selected launch figures reported by Anthropic, in percent. Highest available effort unless noted: FrontierCode uses Fable xhigh and Gemini medium. These are vendor-reported comparisons, not a matched-budget PAS7 test. FrontierCode 2.0 is a different version from the earlier FrontierCode tests. [1]
| Benchmark | Claude Fable 5.2 | Claude Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|
| FrontierCode 2.0 | 49.1% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2 | 69.9% | 62.3% | 65.5% | 61.2% |
| AutomationBench | 51.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% | 58.7% | 68.1% | 64.8% |
| Chartography · no tools | 73.4% | 64.4% | 71.0% | 72.5% |
The lead changes with the workload
Fable leads these coding and science rows, while Gemini 4 Carbon leads AutomationBench. This is already enough to reject a universal-winner reading of the launch table. A five-day-old release also needs more independent testing across real workloads.
Artificial Analysis reports an Intelligence Index score of 65 for Fable at max effort with default fallback. Its weighted index workload costs $26.54 per task. These numbers describe that specific route and workload, not the price of your next chat. The current index is v5.0.1; our October comparison used v4.3.2, so their scores cannot be compared directly. [4]
Within Claude, Opus 5.5 remains Anthropic’s recommended starting point for most work, and Sonnet 5.5 targets faster everyday use. GPT-6 Astra is a demanding-work alternative in OpenAI’s ecosystem; GPT-6.1 Sol offers a lower-cost option for repeated calls. Gemini 4 Carbon deserves a look when native audio or video input matters, since its documented inputs extend beyond text and images. These are selection criteria, not a fresh ranking from our own tests. [7] [8] [9] [10]
Who gains from the engineering side
Research developers and hardware prototypers may benefit when a task spans calculations, technical documents and executable checks. The hub illustration makes the dependency chain visible: the enclosure must fit the board, the circuit must match the components, and firmware must operate the resulting device. A fluent explanation alone cannot validate that chain.
In Anthropic’s example, Fable used datasheets and revised its layouts as it found constraints. For your project, the corresponding evidence would be simulation results, manufacturing checks and measurements. That is a more useful target than asking the model to produce an impressive-looking engineering report. [3]
Anthropic’s original design-to-prototype figure. The English labels identify CAD, board layout, code, assembly and the finished device. Each stage needs its own checks. [3]
Section engineering-use screenshotWhat a Fable session can cost
Standard Fable 5.2 API rates are $8 per million input tokens and $40 per million output tokens. Fable 5.1 charged $10 and $50, so both base rates fall by 20%. Optional fast mode doubles the new rates to $16 and $80; Anthropic advertises up to 2.5× output speed, which is not a guarantee that a whole project finishes 2.5× sooner. [2] [6]
Here is our controlled calculation: 100,000 uncached input tokens plus 10,000 total billable output tokens cost $1.20 on Fable, $0.60 on Opus 5.5, $0.30 on Sonnet 5.5 or GPT-6.1 Sol, and $1.50 on GPT-6 Astra. For Fable the arithmetic is 0.1 × $8 + 0.01 × $40. This is one fixed-volume API example, excluding cache writes, tool fees, premium modes and taxes. Reasoning tokens and repeated agent calls change the actual bill. [2] [7] [8] [9]
Effort is another cost control. Fable defaults to high in the API, while the launch announcement gives medium as the default in Claude apps and Claude Code. The prompting guide recommends medium for interactive work, high for harder agent tasks, and max mainly for exploring the ceiling or rare open-ended problems. More reasoning is useful only when it improves the result enough to justify the time and tokens. [1] [5]
What to check before switching
Long-running does not mean reliably self-finishing. Anthropic documents early stops on multistep tasks, particularly at low and medium effort, and cases where prompts encouraging code ownership trigger broader changes. Give the agent a bounded scope and require it to report what remains incomplete. Review the actual diff as well as its summary. [5]
Custom API clients also need a compatibility check. Changing earlier history, the system prompt or tools can invalidate preserved thinking and produce a 400 error. Older computer-use integrations need migration. Supported clients handle much of this, but swapping a model identifier in a home-built proxy is not a complete integration test. [6]
For company data, the retention rule is material: the documentation requires 30-day retention for Fable 5.2, with zero data retention available only when Anthropic expressly authorizes it. Check that against your organization’s requirements before sending sensitive files. Safeguards may also cause refusals or a configured fallback, so record the model route when evaluating results. [6] [11]
A first task worth giving it
Use a small piece of your own difficult work. Here is a PAS7 example brief, proposed for evaluation: “Update our CSV importer while preserving saved column mappings. Handle UTF-8 names, quoted separators and line breaks. Show malformed-row errors without saving a partial import. Keep the existing interface and provide evidence for each requirement.”
Accept the result only after fixtures cover those cases, an invalid file leaves storage unchanged, the existing mappings still load, and the project’s build and relevant tests pass. Record total spend, elapsed time and human corrections. Give the same starting state to your current model. We have not executed this trial; it is a reusable acceptance brief, not a benchmark result.
For routine changes, begin with your existing model or a cheaper option. Try Fable when a well-defined task repeatedly stalls, or let it advise on one difficult design decision. Keep it in the workflow if it reduces the effort needed to reach an accepted result. That gives the new model a concrete job and gives you a reason to keep paying for it.
Sources and image credits
Related Articles

GPT-6 Astra vs Claude 5.5: which model earns the next task?
GPT-6 Astra, GPT-6.1 Sol, Claude Opus and Sonnet 5.5 compared across 69 benchmark metrics, API costs and coding workflows. Data checked October 24, 2026.