What Is Claude Fable 5? Everything Developers Need to Know

Anthropic launched Claude Fable 5, the most capable model the company has ever made generally available. Claude Fable 5 is a “Mythos-class” model made safe for general use, and it posts state-of-the-art scores on nearly every AI benchmark it’s been tested against, from software engineering to scientific research. If you’ve worked with Claude Opus 4.8, think of Fable 5 as the new premium frontier tier that sits above it. This guide walks through what the model is, how it differs from its restricted sibling Claude Mythos 5, what it can do, how its safety system works, and who should actually use it.

TL;DR

Claude Fable 5 is Anthropic’s new frontier model, launched June 9, 2026, as a “Mythos-class” model made safe for general use. It posts state-of-the-art results across coding, knowledge work, vision, and science, runs autonomously across millions of tokens, and costs $10 per million input tokens and $50 per million output tokens via the claude-fable-5 API ID.

What is Claude Fable 5?

Claude Fable 5 is the public frontier model in Anthropic’s June 2026 release. The company describes it as “a Mythos-class model made safe for general use,” which is the single most important phrase for understanding what you’re getting. There’s a more powerful underlying model, called Claude Mythos 5, and Fable 5 is that same model with safety guardrails applied so it can ship to everyone.

“Mythos-class” refers to a tier of capability, not a separate architecture. Fable 5 and Mythos 5 share the same underlying model. The difference is the safeguards layered on top, which we’ll cover in detail below. For now, the headline is this: Fable 5 is the most capable model Anthropic has ever released for general availability, and it’s state-of-the-art on nearly all tested benchmarks.

You can call it today through the Claude API using the exact model ID claude-fable-5. It’s also fully available on consumption-based Enterprise plans. If you’ve been building on Anthropic models already, adding Fable 5 is mostly a matter of swapping the model string, much like the upgrade path described in our guide to using the Claude Opus 4.8 API. The Anthropic models overview documentation lists it alongside the rest of the current lineup.

Where Fable 5 stands out is breadth. It’s not specialized for one task. Anthropic reports strong performance in software engineering, knowledge work, vision, and scientific research at the same time. That combination is what makes it a “frontier” model rather than a point solution, and it’s why the launch matters beyond the coding crowd.

Claude Fable 5 vs Claude Mythos 5

This is the distinction most people get wrong, so it’s worth being precise. Anthropic launched two models on June 9, 2026, not one. Claude Fable 5 is public. Claude Mythos 5 is restricted.

Both run on the same underlying model. Fable 5 ships with safety safeguards in place. Mythos 5 has some of those safeguards lifted, which makes it more capable in sensitive areas but also more dangerous in the wrong hands. So Anthropic does not make Mythos 5 publicly available.

Instead, Mythos 5 is deployed through Project Glasswing, a collaboration with the US government that puts the model in the hands of a small group of cyberdefenders and infrastructure providers, with the cybersecurity safeguards lifted. Biology researchers will soon get access to a version of Mythos 5 with the biology and chemistry safeguards removed, while the cyber protections stay on for that group. Anthropic has said a broader trusted-access program is planned, and existing Claude Mythos Preview users can upgrade to Mythos 5.

Here’s the side-by-side.

The practical takeaway: if you’re a developer reading this, the model you can actually build on is Fable 5. Mythos 5 is a controlled-access program, not a product you sign up for. The two share pricing and capability ceilings, but only one is on the menu.

What Claude Fable 5 can do

Fable 5’s headline capability is endurance. Anthropic says it “can work autonomously for longer than any previous Claude” and “stays focused across millions of tokens in long-running tasks.” That second part matters more than raw benchmark numbers for anyone building agents. A model that loses the thread halfway through a long task forces you to chunk work, add checkpoints, and babysit the loop. Fable 5 is built to hold context across very long horizons without drifting.

The clearest demonstration of this came from a game test. Anthropic had Fable 5 play Slay the Spire, a roguelike deck-builder that rewards long-term planning across many turns and runs. When the model was given persistent file memory, so it could write notes to itself and read them back, its performance improved 3x over Claude Opus 4.8. That’s a concrete signal that Fable 5 uses long-running memory well rather than just tolerating a bigger context window.

The capability surface is broad:

• Software engineering. Strong enough to take on large, multi-step engineering work, not just isolated snippets. This is where the long-horizon focus pays off most.
• Knowledge work. Document reasoning, analysis, and synthesis across long inputs.
• Vision. Image and visual understanding as part of the same model.
• Scientific research. Reasoning over technical and scientific material.

If your current workflows use Opus 4.8 for agentic coding, the patterns you’ve already built carry over. Our walkthrough of dynamic Claude Code workflows on Opus 4.8 covers the kind of multi-step agent setups that benefit directly from a model that stays focused longer. The main change with Fable 5 is that you can hand it bigger, longer tasks and trust it to keep going.

Benchmarks and real-world results

Anthropic positions Fable 5 as state-of-the-art on nearly all tested benchmarks. The named results, which come from independent and partner evaluations, back that up. The numbers themselves weren’t published in detail, so the placements matter more than any single figure.

• FrontierCode (by Cognition). Fable 5 scored highest among frontier models at medium effort. FrontierCode targets realistic coding tasks, so a top placement here speaks to practical software work, not toy problems.
• Finance Benchmark (by Hebbia). Fable 5 posted the highest score, with notable gains in document reasoning, charts, and tables. That’s a direct signal for anyone doing financial analysis or working with dense structured documents.
• CursorBench. Fable 5 reached state-of-the-art and, in Anthropic’s words, “opened up a class of long-horizon problems that were out of reach.” This is the long-running-task strength showing up in a coding-agent benchmark.
• FrontierBench (by Cognition). Fable 5 was the highest-scoring model, excelling at long-horizon reasoning.

The most striking real-world result comes from Stripe. Using Fable 5, Stripe ran a migration across a 50-million-line Ruby codebase and finished it in a single day. The company estimated the same task would have taken its team over two months by hand. A migration at that scale isn’t a clever demo. It’s the kind of work that normally ties up engineers for weeks, and it’s exactly where a model that holds focus across millions of tokens earns its price.

You can read the full set of claims in Anthropic’s official Fable 5 and Mythos 5 announcement. The pattern across all of these results is consistent: Fable 5 wins most clearly on long, complex, multi-step tasks rather than short one-shot prompts.

How the safety safeguards work

The safeguards are what separate Fable 5 from Mythos 5, and they work in a way that’s worth understanding before you deploy.

Fable 5 doesn’t refuse sensitive queries with a flat error. Instead, certain queries are routed, or fall back, to Claude Opus 4.8 rather than being answered by Fable 5 itself. So when a request touches a protected area, you still get a capable answer from a strong model; you just don’t get Fable 5’s full frontier capability applied to that specific topic. Anthropic reports these safeguards trigger in “less than 5% of sessions” on average, so for the vast majority of normal use you’re getting Fable 5 end to end.

There are three protected areas:

1. Cybersecurity. Exploit development, offensive cyber tasks, and agentic hacking. An external partner that tested the model found Fable 5’s safeguards against harmful cyber queries among the strongest they’d evaluated.
2. Biology and chemistry. Sensitive biological and chemical queries, including things like AAV design and bioweapons-related requests.
3. Distillation. Attempts to extract the model in order to train competing models.

The design choice here is pragmatic. Rather than degrade the model for everyone or block whole topics outright, Anthropic routes the small slice of risky sessions to a model with a different risk profile. If your application happens to live near one of these areas, plan for the possibility that some responses come from Opus 4.8 instead of Fable 5, and test accordingly. For everyone else, the routing is mostly invisible.

This is also the cleanest way to understand the Fable-versus-Mythos split one more time: Mythos 5 is the same model with some of these safeguards removed, which is precisely why it stays inside controlled programs like Project Glasswing instead of shipping to the public API.

Pricing at a glance

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens. Anthropic notes this is “less than half the price of Claude Mythos Preview,” so despite being the new top tier, it lands cheaper than the previous preview-grade frontier option.

Fable 5 sits at roughly double the per-token cost of Opus 4.8, which is the trade you’re making for frontier capability and long-horizon endurance. For a deep breakdown of how to estimate real bills, batch discounts, and cost control, see our dedicated Claude Fable 5 pricing guide. If you’re still costing out the tier below it, the Claude Opus 4.8 pricing breakdown is the companion piece. The current numbers also live on Anthropic’s pricing page.

Claude Fable 5 vs Claude Opus 4.8

For most teams, the real decision isn’t Fable versus Mythos. It’s Fable 5 versus Opus 4.8. Opus 4.8 is still an excellent, cheaper model, and it’s the one Fable 5 routes to under the safeguards. Picking between them comes down to how hard your tasks are and how long they run.

A reasonable rule of thumb: reach for Fable 5 when you have long, complex, multi-step work where staying on task across millions of tokens is the whole game, such as large codebase migrations, deep document analysis, or extended agent runs. Stay on Opus 4.8 when the task is shorter or more routine and the lower price wins. If you want a wider field, our Claude Opus 4.8 vs GPT-5.5 vs Gemini 3.5 comparison frames where Anthropic’s lineup sits against the other frontier labs, and the same logic applies a tier up. And if budget is the constraint, the free Opus 4.8 access options are worth a look before you commit spend.

Testing the Claude Fable 5 API with Apidog

Once you decide to build on claude-fable-5, you’ll be sending requests to the Anthropic Messages API. That’s where an API client helps. Apidog is an API design and testing tool, and it gives you a clean place to develop against Fable 5 before you wire anything into production code.

A few things it handles well for this kind of work:

• Auth without friction. Fable 5 calls use bearer-token authentication. You set your Anthropic API key once as an environment variable in Apidog and reuse it across every request, so the key never gets pasted into a dozen scattered scripts.
• Streaming responses. Fable 5 is built for long outputs, and you’ll usually want to stream them. Apidog can send the request with streaming enabled and show you tokens as they arrive, which makes it easy to confirm your streaming setup works before you build it into an app.
• Token and cost tracking. Because pricing is $10 input and $50 output per million tokens, the usage field in each response is money. Inspecting the input and output token counts per call helps you estimate real spend while you’re still experimenting.
• Saved, repeatable requests. Keep your model string, system prompt, and parameters in a saved request so the whole team tests against the same setup.

The flow mirrors how you’d test any modern LLM endpoint. If you’ve followed our guide on testing the ChatGPT API with Apidog, the steps for Fable 5 are nearly identical; you swap the base URL, auth header, and model field for Anthropic’s. When you’re ready to move from inspection into real integration code, the Claude Fable 5 API guide covers request bodies and parameters in depth. Download Apidog to follow along while you experiment.

Beyond its capabilities,the government-ordered suspension that took Fable 5 offlineis a reminder that even the most production-ready models can be subject to regulatory intervention.

Once you have a clear picture of what the model can do, structuring prompts to get more out of each Fable 5 request is the practical next step—especially when working within a token or cost budget.

Computer use is emerging as a key differentiator across frontier models, andHolo3's dedicated computer use API is worth benchmarking directly against Fable 5 if you're choosing a model to drive browser or desktop automation.

FAQ

Is Claude Fable 5 free?

No. Claude Fable 5 is a paid frontier model priced at $10 per million input tokens and $50 per million output tokens through the Claude API and consumption-based Enterprise plans. There’s no free tier for Fable 5 itself. If you want to try a strong Anthropic model without paying, the cheaper Opus 4.8 has some no-cost access paths, and a free Fable 5 access guide covers any trial-style options for the new model.

What does “Mythos-class” mean?

“Mythos-class” describes a capability tier, not a separate model architecture. The underlying Mythos 5 model defines the ceiling. Fable 5 is that same Mythos-class model with safety safeguards applied, so it can be released for general use while keeping the most sensitive capabilities gated.

Can I get Claude Mythos 5?

Not through normal channels. Mythos 5 isn’t publicly available. It’s deployed through Project Glasswing, a US government collaboration, to a small group of cyberdefenders and infrastructure providers with cyber safeguards lifted. Biology researchers will get a version later with biology and chemistry safeguards removed, and a broader trusted-access program is planned. Existing Claude Mythos Preview users can upgrade to Mythos 5.

Is Fable 5 better than Opus 4.8?

On raw capability, yes. Fable 5 is the more capable frontier model and posts state-of-the-art benchmark results, including a measured 3x improvement over Opus 4.8 in the Slay the Spire memory test. But Opus 4.8 costs half as much ($5/$25 versus $10/$50) and is excellent for shorter or routine tasks. Better depends on whether your work needs frontier, long-horizon performance or value. For a side-by-side on the tier below, see what Claude Opus 4.8 is.

What is the Fable 5 API model name?

The API model ID is exactly claude-fable-5. You pass that string in the model field of your Anthropic Messages API request. It’s available today via the Claude API and fully available on consumption-based Enterprise plans.

Why did my Fable 5 query get a different-feeling answer?

It may have hit a safeguard. Queries touching cybersecurity, biology and chemistry, or model distillation are routed to Claude Opus 4.8 instead of being answered by Fable 5. Anthropic says this happens in less than 5% of sessions on average, so it’s uncommon, but it’s the reason a sensitive query might feel slightly less capable.

Does Fable 5 have a bigger context window than Opus 4.8?

Fable 5 is built for long-running tasks and stays focused across millions of tokens. Opus 4.8 has a 1M-token context window. Both handle long inputs well, but Fable 5’s edge shows up most in sustained, multi-step work where it keeps track of context over a very long horizon.

How much can Fable 5 actually do in one task?

A lot. In Anthropic’s most cited example, Stripe used Fable 5 to migrate a 50-million-line Ruby codebase in a single day, a job the team estimated would have taken over two months by hand. That scale is the clearest illustration of what long-horizon autonomy unlocks.

Claude Fable 5 is Anthropic’s new frontier ceiling for everyone outside controlled programs: a Mythos-class model, made safe, priced at $10/$50, and built to stay focused across the kind of long tasks that used to break smaller models. The fastest way to know whether it fits your workload is to send a few real requests and watch the responses and token counts come back. Download Apidog, point a request at claude-fable-5, and see how it handles your own use case before you write a line of integration code.