What is Claude Fable 5?
The safeguarded, generally available half of a new Mythos-class family. Powerful enough that the mind directing it matters more, not less.
Claude Fable 5 is Anthropic's newest and most capable model, released June 9, 2026, as the safeguarded, generally available version of a new Mythos-class model. It is state-of-the-art on nearly all tested benchmarks, especially software engineering, knowledge work, vision, and scientific reasoning, and works autonomously for longer than any previous Claude. It runs on a one-million-token context window at 10 dollars per million input and 50 per million output, double the prior Opus 4.8 flagship. Because a frontier model amplifies the mind directing it, the structured First Brain you bring still decides what you get.
Claude Fable 5 is Anthropic’s newest and most capable model, released on June 9, 2026, and it is the safeguarded, generally available member of a new model family Anthropic calls Mythos-class. In plain terms, it is the model most people and developers will actually use: state-of-the-art on nearly all of the capability benchmarks Anthropic tested, unusually strong at software engineering, knowledge work, vision, and scientific research, and able to work autonomously for longer than any previous Claude. It runs on a one-million-token context window, costs 10 dollars per million input tokens and 50 dollars per million output tokens, and answers to the model id claude-fable-5. The deeper point, the one this whole site is about, is that a model this powerful amplifies the mind directing it, so the structured First Brain you bring still decides what you get out. Here is what Fable 5 is, what it does well, and where its limits are.
What exactly is Claude Fable 5?
It is the public, safety-enabled version of a single underlying model. Anthropic shipped that model in two configurations: Claude Fable 5 and Claude Mythos 5. Fable 5 has safeguards switched on and is broadly available; Mythos 5 is the same model with some safeguards lifted, restricted to vetted partners, and not something the general public can use. So when people say “the new Claude,” they almost always mean Fable 5.
What makes it a generational step rather than a point update is autonomy and breadth. Anthropic describes it as state-of-the-art on nearly all tested benchmarks, with a lead that grows as tasks get longer and more complex, which is a polite way of saying it holds a goal in mind across many steps better than earlier models did. If you have followed how large language models actually work, the headline is the same skill, next-token prediction, scaled and tuned until sustained, multi-step work becomes reliable enough to trust on real projects.
What is Fable 5 actually good at?
Four areas stand out, and the evidence is concrete rather than vibes. In software engineering, Anthropic reports it completed a codebase-wide Ruby migration in a single day that a team had estimated at two months, and it posted the highest score among frontier models on Cognition’s FrontierCode evaluation at medium effort. In knowledge work, it took the top score on the Hebbia Finance benchmark and cleared 90 percent on a widely shared analytics benchmark, a ten-point jump over the prior Opus generation.
Vision is the second surprise. Fable 5 can rebuild a working web app from a screenshot, and Anthropic demonstrated it playing Pokemon FireRed through a vision-only harness, which sounds like a stunt but is really a test of perceiving a changing environment and acting on it over time. The third strength is scientific reasoning, and the fourth is the duration of autonomous work itself: the model sustains a task across more steps before drifting, which is what turns a chatbot into something closer to a junior collaborator.
| Capability area | What Anthropic reports | Why it matters in practice |
|---|---|---|
| Software engineering | Top FrontierCode score; one-day codebase migration | Larger, multi-file changes with less hand-holding |
| Knowledge work | Top Hebbia Finance score; 90%+ analytics, +10 over Opus | Research and analysis you can lean on, with checking |
| Vision | Rebuilds apps from screenshots; plays a game vision-only | Reads interfaces and documents, not just text |
| Autonomous duration | Lead grows as tasks get longer | Holds a goal across many steps before drifting |
How is it kept safe, and what are the limits?
Fable 5’s safeguards are the entire reason it can be public. Anthropic built three classifier areas around the model. The first blocks offensive cybersecurity tasks. The second is deliberately cautious on biology and chemistry, falling back to a smaller model on most such requests even at the cost of being over-careful. The third blocks large-scale attempts to extract the model’s capabilities. When a safeguard triggers, the system falls back to Opus 4.8 rather than refusing outright, and Anthropic reports that more than 95 percent of Fable sessions involve no fallback at all.
The honest limits follow from how these models work. Fable 5 is more capable, but it is still a system that produces fluent, plausible text, so it can be confidently wrong, a tension explored in whether large language models really understand language. It does not eliminate the need to verify facts, it does not have goals or judgment of its own, and longer autonomous runs mean a small early error can compound. Anthropic also applies a 30-day data-retention requirement to all Mythos-class traffic. None of this is a reason to avoid the model. It is a reason to treat it as a powerful instrument rather than an oracle.
Where does it sit next to Opus 4.8?
Fable 5 is the step above the prior flagship. Opus 4.8, until now Anthropic’s most capable model, costs 5 dollars per million input tokens and 25 dollars per million output tokens; Fable 5 doubles that to 10 and 50, which tells you Anthropic positions it as a meaningfully stronger tier rather than a free upgrade. Both carry the one-million-token context window at standard pricing. The practical read is that Opus 4.8 remains an excellent, cheaper default for most work, and Fable 5 earns its premium on the hardest, longest, highest-stakes tasks where its autonomy and benchmark lead actually pay for themselves. The full breakdown lives in what Claude Fable 5 costs, and the way to reach it is covered in how to access Claude Fable 5.
Who should actually reach for Fable 5?
The honest answer is fewer people than the launch excitement suggests, and that is fine. Fable 5 earns its premium on work that is genuinely hard, long, or high-stakes: a multi-file refactor or migration that would otherwise eat a week, a deep research or financial-analysis task where being wrong is expensive, an agent that has to hold a goal across many steps without drifting, or vision work like turning interfaces and documents into structured output. In those cases the autonomy and the benchmark lead translate into real saved time.
For the large middle of everyday use, drafting, summarizing, ordinary coding help, quick questions, the cheaper models are the smarter buy. Opus 4.8 at half the price covers most serious work, and Sonnet-class models cover the rest at a fraction again. Paying double for Fable 5 on a task a smaller model handles cleanly is spending for headroom you will not use.
A useful rule is to match the model to the hardest part of the job, not the average part. If most of your work is routine with occasional hard spikes, keep a cheaper model as the default and reach for Fable 5 on the spikes. The skill of choosing well, knowing which problems actually need the frontier and which do not, is itself a First Brain skill: it comes from understanding your own work clearly enough to judge its real difficulty.
Why a more powerful model makes your First Brain matter more
Here is the part the launch coverage misses. A frontier model is an amplifier, and an amplifier multiplies whatever signal it is given. Point Fable 5 at a vague, half-formed request and you get fluent, confident, average output faster. Point it at a precise question from someone who genuinely understands the problem, holds the relevant context, and can judge the answer, and it becomes a genuine extension of that person. The model did not supply the understanding. It scaled the understanding that was already there.
This is the case for First Brain before Second Brain, and it gets stronger, not weaker, as the tools improve. The durable advantage is a connected internal model of what you know, a biological knowledge graph dense enough that you can frame sharp questions, supply real context, and catch the moments when a confident answer is wrong. People worried that a model this capable makes their own thinking redundant have it backwards: the more powerful the instrument, the more the quality of the hand on it decides the result, which is the same reason building a First Brain matters in the age of capable AI. The method for building that internal model is the core of Building Your First Brain, free for the first 1,000 readers.
Key takeaways: what Claude Fable 5 is
Claude Fable 5 is Anthropic’s newest and most capable model, launched June 9, 2026, as the safeguarded, generally available half of a new Mythos-class family, with Mythos 5 being the restricted, safeguard-lifted counterpart most people will never touch. It is state-of-the-art across software engineering, knowledge work, vision, and scientific reasoning, and its lead grows on longer, more autonomous tasks, backed by concrete results like a one-day codebase migration and top scores on FrontierCode and the Hebbia Finance benchmark. It runs on a one-million-token context window at 10 dollars per million input and 50 per million output, double the prior Opus flagship, with three safeguard classifiers and a fallback to Opus 4.8. The honest framing is that it is a powerful amplifier, not an oracle, so the structured First Brain directing it still decides what you get. The limit worth keeping in mind: it can be confidently wrong, so verification and your own judgment remain essential.
Frequently asked questions
What is Claude Fable 5?
Claude Fable 5 is Anthropic’s newest and most capable model, released on June 9, 2026, as the public, safety-enabled version of a new Mythos-class model. It is state-of-the-art on nearly all the benchmarks Anthropic tested, especially strong at software engineering, knowledge work, vision, and scientific research, and able to work autonomously for longer than any previous Claude. It uses a one-million-token context window and the model id claude-fable-5. Because it amplifies the mind directing it, the most reliable way to get value from it is to bring a structured First Brain, which is what the Build First Brain approach builds.
Is Claude Fable 5 better than Claude Opus 4.8?
Yes, on capability, and Anthropic prices it accordingly. Fable 5 is state-of-the-art across more benchmarks and holds longer autonomous tasks better than Opus 4.8, which was the prior flagship. It also costs twice as much, at 10 dollars per million input tokens versus 5, and 50 versus 25 for output. For most everyday work Opus 4.8 remains a strong, cheaper choice; Fable 5 earns its premium on the hardest, longest, and highest-stakes tasks where its lead actually changes the outcome.
What does Claude Fable 5 cost?
The API price is 10 dollars per million input tokens and 50 dollars per million output tokens, with the full one-million-token context window at standard pricing and a 50 percent discount through the Batch API. That is double the price of Opus 4.8. On Anthropic’s paid plans it was included at no extra cost for a launch window, then moved to usage-credit access. A full breakdown, including cache and batch pricing and how it compares to alternatives, is covered separately in the cost guide.
Is Claude Fable 5 safe to use?
For general use, yes, which is the whole reason this version exists. Fable 5 ships with three safeguard classifiers covering offensive cybersecurity, biology and chemistry, and large-scale capability extraction, and it falls back to Opus 4.8 when a safeguard triggers rather than refusing. Anthropic reports more than 95 percent of sessions involve no fallback at all. The realistic caveat is not safety policy but reliability: like any model it can be fluently and confidently wrong, so verify important facts and keep your own judgment in the loop.
Can anyone use Claude Fable 5?
Effectively yes. Fable 5 is the broadly available version, reachable through the Claude API and Anthropic’s apps. The restricted counterpart, Claude Mythos 5, is the same underlying model with some safeguards lifted and is limited to vetted partners through a trusted-access program, so the general public will only ever use Fable 5. For nearly everyone, Fable 5 is the model in question, and the differences between the two are worth understanding before assuming you need the restricted one.
Dive deeper in
How do I access Claude Fable 5?
Claude Fable 5 is on the API (claude-fable-5) and claude.ai now; paid plans are phased. Bedrock, Vertex, and Copilot are unconfirmed; Mythos 5 is not public.
What is the difference between Claude Fable 5 and Claude Mythos 5?
Fable 5 and Mythos 5 are the same model governed two ways: Fable is safeguarded and public; Mythos has safeguards lifted and is restricted to vetted partners.
Claude Fable 5 vs GPT-5.5: which is better?
No clean overall winner: Fable 5 leads Anthropic's benchmarks, GPT-5.5 is cheaper at $5/$30 vs $10/$50, both have 1M context. Test on your own task.