Meet Claude Fable 5.2: Features and Benchmarks

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday, November 9. Fable 5.2 is open to everyone. Mythos 5.2 runs on the same model without the cybersecurity and biology limits, so access goes through Anthropic’s verification programs.

The main improvements are in robot control, engineering design, chart and document work, and long agent runs. Let’s take a closer look.

Fable sits at the top of Anthropic’s lineup, above Opus, Sonnet and Haiku. Anthropic’s own guidance hasn’t changed: start most coding and knowledge work on Opus 5.5, and move to Fable when a task is genuinely open-ended or runs for hours.

*The 30% speed figure applies to ordinary requests. Fast mode is a separate setting.

Before we introduce the comparison, it's worth noting that every figure below comes from Anthropic’s own docs.

Fable 5.2 beats Opus 5.5 on eleven of the twelve rows. The size of the win varies a lot by task. RoboDojo went from 27.6% to 64.9%, BenchCAD from 44.4 to 83.5 points, and Remote Labor Index more than doubled.

Two results go the other way. Gemini 4 Carbon leads AutomationBench at 53.9% against 51.4%. On AA-Briefcase Elo, Opus 5.5 sits at 1800 and Fable 5.2 at 1785, so the older and cheaper model comes out ahead on that one.

Read the FrontierCode 2.0 row before quoting it. Fable 5.2’s 49.1% was run at Xhigh effort and 5.1’s 32.4% was too, so that comparison holds. The competitor columns used different settings, which makes them close to a different test.

The context window hasn't changed, and neither has the output ceiling. If 1M tokens wasn’t enough for your job on 5.1, 5.2 doesn’t solve it.

Most coverage of this release focuses on coding, which somewhat undersells it.

The benchmark that moved most outside of robotics is Chartography, which tests reading a chart with no external tools. Fable 5.2 scored 73.4% against 44.8% on 5.1, which is a different class of result rather than an incremental one. If your work involves pulling numbers out of a plot, a scanned table or a dashboard screenshot and reasoning on them, that’s the row to care about.

AA-Briefcase went from 61.5% to 66.1%, and it measures exactly the kind of output most office work produces: a written brief built from source material. Humanity’s Last Exam with tools rose from 65.6% to 74.2%, and AutomationBench from 31.4% to 51.4%. Terminal-Bench-Science reached 74.6%, which covers multi-step scientific analysis run from a terminal.

Anthropic describes the intended use as multi-step deep research and analysis carried through to a finished document, spreadsheet or deck. The WANDR result supports that, at 82.1% on agentic search against 68.7% for 5.1.

One caution on all of it. These benchmarks ask the model to complete a bounded task in one sitting. A two-hour research job with forty sources has more places to go wrong than any of them measure, so treat the scores as a sign of where to experiment rather than a promise about a Monday deadline.

Medium is the default in Claude apps and Claude Code. That means most people will never see the top numbers in the tables above unless they change the setting.

Anthropic’s more useful claim is about token use. It says Fable 5.2 at High effort lands roughly level with Opus 5.5 at Max effort on FrontierCode 2.0, while spending about half the tokens. Fewer tokens means less spend wherever you run the model, and it also means faster turnaround on the same job.

Higher effort costs more per request than Medium, in both tokens and time. A routine rewrite, a summary or a formatting pass can stay on Medium. A job that reads twelve documents and has to reconcile them before writing anything is worth the extra reasoning. Xhigh is for long coding problems and multi-hour agent runs, where a wrong turn early is expensive to unwind.

Anthropic published a set of examples from its own work, and they’re worth reading (with the source in mind).

Fable 5.2 formalized Fermat’s Last Theorem in 30 hours using a team of 1,000 agents. Opus 5.5, given the same setup, completed 26% of the work in 18 hours and then stalled. Anthropic’s first pass at the same formalization, on an unreleased research model, took 11 days.

From a one-paragraph prompt, the model designed three working electronic devices: a voice assistant hub, a USB speakerphone and a handheld virtual pet. Diode fabricated the circuit boards.

An agent running an early build migrated a production service from Python to Rust, and the service now starts about 26 times faster. Anthropic also says Claude writes over 90% of the changes merged into its core codebase, with human review on each one.

Mythos 5.2 predicted full structures for 28 of the 34 capsids it attempted, which matters for gene therapy delivery. That work sits behind the safeguards Fable keeps and Mythos drops.

Anthropic published all of this itself, and no outside party has repeated any of it. That doesn’t make the numbers wrong, since some of them are verifiable claims about completed work. It does mean you’re reading the strongest case the company can make, assembled by the people with the most to gain from it. Take it as an upper bound on what’s been demonstrated, not a floor you can count on.

The robotics result is impressive. Going from 27.6% to 64.9% on RoboDojo isn’t a tuning gain, and it puts a general-purpose model ahead of most systems built specifically to control robots. If you work anywhere near hardware, that might change your toolkit.

But for everyday writing and coding, Fable 5.2 is hard to justify as a default. We recommend you start on Opus and switch only when necessary, since earlier Claude versions still handle general workflows well. Besides, they’re all available on SYNTX.AI, so you can test and decide for yourself which model fits your task.

Fable 5.2 is available on SYNTX.AI, and you don’t need an Anthropic account.

1. Sign up with your email. New accounts get welcome tokens, which let you test some tools for free.
2. Open LLM Studio and pick Claude Fable 5.2 from the model list.
3. Set effort to Medium for routine work and raise it when a task needs more reasoning.
4. Attach documents, charts or screenshots if the job involves reading them.

Below are a few more features that can help you tailor Claude to your task.

Capabilities

For a rewrite, a summary or a quick question, you can leave both off and the answer will arrive sooner.

Tools

The tools can work together. With Web search, Code interpreter and Charts on, one prompt can collect figures from the web, run the calculation and plot the result.

Usage is billed in tokens, so what a task costs depends on length, effort level and whether Fast mode is on. Unused tokens stay in your account instead of resetting at the end of the month.

Claude’s latest version is already live, available under the same subscription as ChatGPT, Gemini, Perplexity, DeepSeek and other leading LLMs. For creative work, SYNTX.AI carries 100+ models for image, video, and audio generation. Sign up today and test the tools for free with welcome tokens.

What is the difference between Claude Fable 5.2 and Claude Mythos 5.2?

The underlying model is the same. Fable 5.2 ships with strong cybersecurity and biology safeguards, so it’s available to everyone. Mythos 5.2 removes those limits and is available only to organizations approved through Anthropic’s Cyber or Life Sciences verification programs.

Is Fable 5.2 better than Opus 5.5?

On Anthropic’s published benchmarks, yes, on eleven of twelve tests. In practice, Anthropic still recommends Opus 5.5 for most coding and knowledge work because it costs much less to run. Use Fable for long agent sessions and hard open-ended problems.

Do I need to change the effort setting in Claude?

Medium is the default and suits most prompts. Raise it to High or Xhigh for dense multi-step reasoning, multi-document analysis or hard coding problems. Higher settings cost more tokens per request.

Can I use Fable 5.2 without an Anthropic subscription?

Yes, on SYNTX.AI, under the same subscription as ChatGPT, Gemini and other LLMs. New accounts get welcome tokens at signup.