We tested Anthropic’s Fable 5 for a week: warp drive for coders

Fable 5 is Anthropic’s new Mythos-class model, and it doesn’t really behave like a normal chatbot. It’s slow, expensive, and incredibly powerful—closer to a warp drive you spin up for big jobs than a daily driver you chat with all day.

What Fable 5 actually is

Fable 5 sits at the top of Anthropic’s model stack: Haiku, Sonnet, Opus, and then Mythos. Fable is a Mythos-class model—essentially the same architecture as previous Claude models, just much larger and more capable.

To make it safe enough to release, Anthropic has put strict guardrails on it. You can’t use it for cyber-related tasks, biological work, or other sensitive domains. Within those limits, though, it’s designed to be the most capable general-purpose and coding model Anthropic has shipped so far.

That power comes at a price. Fable 5 currently costs around $10 per million input tokens and $50 per million output tokens—roughly twice the cost of Claude Opus. It also loves to use tokens, so long runs can get expensive quickly.

A model that builds whole projects on its own

The most striking thing about Fable 5 is how far it can go from a single prompt. One example: a fully playable 3D browser game inspired by Jorge Luis Borges’ “The Library of Babel.”

The prompt was essentially: read the story, then plan and build an end-to-end 3D game where the player is dropped into an infinite library. From there, Fable 5:

• Read and analyzed the original text
• Planned the game mechanics and environment
• Implemented a browser-playable 3D experience
• Iterated on its own work in loops, checking and refining as it went

After 3–4 hours of autonomous execution, the result was a detailed, accurate Library of Babel simulation: hexagonal galleries stacked endlessly, shelves laid out as described in the story, and even a working bookmark system that could “find” specific texts in the library.

The key point isn’t just that it built a game. It’s that it did the whole thing in one shot, with minimal human intervention beyond the initial prompt.

How strong is it at coding?

To get a more grounded view, Fable 5 was tested on a “senior engineer benchmark.” The test: give the model a messy, real-world production codebase and ask it how it would rewrite it from first principles, then score the result out of 100.

Previous top scores looked like this:

• Claude Opus 4.8: 63/100
• GPT 5.5: 62/100

Fable 5 scored 91/100—essentially tied with a strong human senior engineer, and far beyond any previous model on this benchmark.

That gap shows up not just in raw coding, but in how it thinks: architecture choices, tradeoffs, and the ability to reason about a whole system rather than just patching individual files.

Where Fable 5 really shines

Fable 5 isn’t just “a bit better” at everything. It has a few standout strengths that feel qualitatively different from earlier models.

1. Sustained autonomous execution

The best way to use Fable 5 is to give it a big, meaty task—then walk away. It’s built for long, multi-hour runs where it can plan, execute, check its own work, and iterate.

Think of it like this: you describe the destination, it plots the route, and then it quietly works through the entire journey. That’s very different from the usual back-and-forth chat where you constantly steer the model.

2. Taste and attention to detail

One striking example was a custom mini-site built around a series of philosophy lectures. The only instruction was: find Hubert Dreyfus’ lectures on Heidegger and turn them into a small site that’s easier to consume.

From that single prompt, Fable 5:

• Found the audio lectures on its own
• Transcribed and summarized them
• Created a clean table of contents
• Built a synchronized player that highlights text as the audio plays
• Added thoughtful UX touches like skip buttons, playback speed, and follow/unfollow text

The design details were surprisingly polished: font choices, weights, drop caps, and layout all felt intentional rather than default “AI slop.” It looked more like something a careful front-end developer and designer might produce than a generic boilerplate template.

3. Deep context and research

Fable 5 is also excellent at digesting large amounts of data and surfacing clear, actionable insights.

In one test, it was given:

• Hundreds to thousands of subscriber survey responses
• Analytics data
• Access to the live website

From that, it produced a concise diagnosis: the business had a free-to-paid conversion problem, and the most promising lever was pricing transparency plus a trial offer. It framed this as a falsifiable bet—exactly the kind of output you’d expect from a strong growth lead after days of work.

This kind of cross-cutting analysis—reading qualitative feedback, pairing it with quantitative data, and then turning it into a clear, testable strategy—is where Fable 5 feels like a genuine multiplier for knowledge work.

4. Crushing backlogs and routine engineering work

Another real-world use case: a markdown editor app called Proof had accumulated a long list of GitHub issues, many of them auto-submitted by agents.

Fable 5 was pointed at the repo and told to:

• Review all issues from the last few weeks
• Close anything irrelevant
• Write fixes for the rest, including Rust code changes

It went through the backlog, closed issues, and produced working patches that were actually merged. Other models can help with this kind of work, but usually only with tight human supervision and one-issue-at-a-time prompting. Fable 5 is much closer to “tell it what you want, then let it run.”

Why it feels like a warp drive, not a car

The best mental model for Fable 5 is a sci-fi warp drive:

• You give it a big destination (a game, a full analysis, a major refactor)
• It spins up, does a lot of heavy computation, and eventually delivers you there
• It compresses weeks or months of effort into hours or days

But just like a warp drive, it’s not great for short hops around town. For quick questions, small edits, or highly interactive collaboration, it’s overkill: too slow, too expensive, and too heavy.

One useful trick: you can run Fable 5 at lower reasoning levels (medium or low) for simpler tasks. People inside Anthropic reportedly do this to keep latency and cost down when they don’t need the full power. Still, for everyday chat or rapid back-and-forth, lighter models like Claude Opus or GPT are usually a better fit.

Who should actually use Fable 5 right now?

Fable 5 is not a general-purpose upgrade for everyone. Whether it’s worth it depends heavily on how you already use AI.

Great fit: advanced users and technical builders

If you’re at the high end of AI adoption—using multiple agents, orchestrating workflows, delegating substantial chunks of work, and comfortable with code—Fable 5 can feel transformative. You’re likely to have “galaxy-scale” problems where compressing weeks of work into a night run is genuinely valuable.

For “vibe coders” and technical tinkerers, it opens up projects that used to be out of reach: one-shot games, complex tools, and polished experiences you can build mostly through prompting, as long as you can afford the runs.

Overkill: most everyday knowledge workers

If you mostly use AI like a smarter search engine, an email helper, or a writing assistant, Fable 5 will probably feel like too much. It’s slower, more expensive, and not dramatically better for many everyday tasks.

In writing, for example, it doesn’t clearly beat Claude Opus 4.8. Its prose tends to be dense, literary, and blocky—good for thinking through structure and arguments, but not ideal for crisp copywriting or fast content production. For that, Opus or GPT models are still better daily drivers.

How it compares to other frontier models

On pure coding and system-level reasoning, Fable 5 appears to leap ahead of previous Claude and GPT releases in at least some internal benchmarks. But that doesn’t mean it replaces everything else.

For many users, GPT 5.5 (and likely successors like the rumored GPT‑5.6) will remain the more balanced choice for day-to-day work: faster, cheaper, and great at interactive collaboration. If you want a deeper dive into how these frontier models stack up and what recent leaks and benchmarks actually mean, it’s worth reading a detailed comparison of GPT‑5.6 and Claude Mythos 5.

If you’re curious about how Mythos fits into Anthropic’s broader roadmap and safety posture, you may also want to check out this breakdown of what we actually know about Claude Mythos.

Limits and weaknesses

Despite the hype, Fable 5 has clear downsides and blind spots:

• Cost and token hunger: Long autonomous runs can rack up serious bills.
• Latency: It’s slow, especially at high reasoning levels. You won’t enjoy using it for chatty back-and-forth.
• Writing style: Its default prose is dense and literary, not snappy. Great for thinking, less great for marketing copy.
• Safety constraints: No cyber or bio work, and likely strict refusals around other sensitive topics.

Even for heavy AI users, it’s not a full replacement for faster, cheaper models. A realistic setup is more like: use lighter models for everyday work, and spin up Fable 5 only when you truly need the warp drive.

What Fable 5 means for the future of work

Fable 5 is a good preview of what happens when “frontier” capabilities become usable tools. It raises the floor for non-experts—letting a hobbyist build a one-shot game or a polished mini-site—and raises the ceiling for experts, who can now tackle projects that used to require whole teams.

It doesn’t remove human work; it changes its shape. Automation tends to create new, higher-level tasks: deciding what to build, how to frame problems, what to test, and how to stitch many automated systems together into something valuable.

In the near term, Fable 5 is a powerful but niche tool. It’s best for people who already think in terms of agents, workflows, and large, complex projects. Over the next 6–12 months, though, it’s reasonable to expect this level of capability to get cheaper and more accessible—turning today’s warp drive into tomorrow’s standard engine.

Should you try it?

If you’re an engineer, advanced AI user, or ambitious builder with big projects in mind—and you’re comfortable with higher costs—Fable 5 is absolutely worth experimenting with. Treat it like a special tool you bring out for major jobs, not something you keep running all day.

If you’re earlier in your AI journey or mostly need help with everyday writing and research, you’ll probably get more value from lighter, faster models for now. But Fable 5 is a clear signal of where things are headed: more autonomy, deeper reasoning, and a lot more that a single prompt can do.