Anthropic has launched two models at once: Claude Fable 5 and Claude Mythos 5. In short, Fable 5 is the public version of a new Mythos-class model, while Mythos 5 is almost the same model with broader capabilities enabled for trusted cybersecurity and biology teams.
The main story is not that the benchmarks went up again. That part is expected now. The more important shift is that Anthropic is trying to give regular users a model above the Opus class without opening the riskiest capabilities to everyone by default.
The one-minute version
- Fable 5 is available today. Developers can use
claude-fable-5through the Claude API. - Mythos 5 is not a normal public model. It is the same underlying capability tier, but with selected safeguards lifted for verified Project Glasswing participants and future trusted access programs.
- The strongest areas are long engineering tasks, document analysis, vision, long context, and scientific work. The longer and more complex the task, the more visible the gap from earlier Claude models should become.
- Fable 5 has a safety layer. If a request falls into sensitive areas like cybersecurity, biology/chemistry, or model distillation, the answer may automatically fall back to Claude Opus 4.8. Users are supposed to be told when that happens.
- Pricing is $10 per million input tokens and $50 per million output tokens.
- There is an important data-policy change. For Mythos-class models, Anthropic is introducing 30-day retention for business customer traffic for safety purposes. According to Anthropic, this data will not be used to train new Claude models.
Why this is not just another Opus
Anthropic calls Fable 5 a Mythos-class model. That tier sits above Opus. In recent months, this level of capability was available only to a limited set of cyberdefenders through Mythos Preview. Now part of that tier is moving into regular Claude.
The idea behind Fable 5 is roughly this: give most users near-Mythos capabilities, but put automated classifiers around the riskiest directions. So Fable 5 is not a weak version. It is closer to a public version of the same model line with an additional safety wrapper.
That is a meaningful shift. A strong-model launch used to be easier to describe: the model got smarter, everyone gets the upgrade. Now a strong-model launch is also a product and policy structure: who gets full access, who gets guarded access, where fallback is required, what logs are retained, and how fast false positives can be reduced.
Where Fable 5 should stand out
Based on Anthropic's description, Fable 5 is especially strong when the model is not just answering one question but holding a goal over time.
Software engineering. The clearest use cases are large migrations, reviews, prototyping, and agentic work in big codebases. Anthropic cites early feedback where the model handled work that was not just a single file edit, but a multi-step reshaping of a system. For developers, this is probably the central scenario: less micromanagement, more complete tasks.
Knowledge work. The model is better at working with documents, tables, charts, and complex analysis. That sounds less dramatic than "it completed a game from screenshots," but it may matter more in real work. Finance, legal documents, internal reports, research briefs: anything where the task is not to summarize one text, but to build a conclusion from several sources.
Vision. Anthropic highlights vision specifically: extracting numbers from scientific figures, understanding interfaces from screenshots, and rebuilding a web app from an image. The most vivid demo is Pokemon FireRed, where the model completed the game using only visual input, without special map or game-state hints.
Memory and long context. Fable 5 is supposed to stay focused across millions of tokens and use its own notes better. For agentic work, this is critical. The weak point is often not that the model "doesn't understand," but that it loses the thread after several hours.
Science and biology. This is where the wording needs more care: Anthropic describes the strongest capabilities mostly in the context of Mythos 5 and trusted access. But the direction is clear. Models are moving beyond literature-summary assistance toward choosing tools, forming hypotheses, and running chains of experimental work.
The main feature is the access boundary
The most practical part of the launch is the safeguards.
Fable 5 uses separate classifiers that inspect requests and decide whether a task falls into a risky domain. If it does, Fable 5 does not answer directly: the request is handled by Claude Opus 4.8 instead. Anthropic says fallback triggers in less than 5% of sessions on average, so most everyday usage should still be real Fable 5.
The covered areas are:
- cybersecurity, especially exploitation and offensive tasks;
- biology and chemistry, where dual-use risk is high;
- attempts to distill the model's capabilities into other models.
This is a compromise. For regular users it will sometimes be annoying: a harmless request may accidentally hit a classifier. For Anthropic, this is the cost of launching quickly. The company says it tuned the layer conservatively and wants to reduce false positives after launch.
Mythos 5: the same strength, not for everyone
Mythos 5 is the model for trusted use cases. At launch, access goes to Project Glasswing participants and selected research groups. In cybersecurity, some restrictions are lifted because the intended users are infrastructure defenders who need exactly those capabilities.
That does not mean regular users "did not get the best model." It suggests the shape of future launches: the same underlying model may have different access modes depending on risk, audience, and domain.
Pricing, availability, and the catch
Fable 5 is available now in the Claude API as claude-fable-5. Pricing for Fable 5 and Mythos 5 is the same: $10 per million input tokens and $50 per million output tokens.
For subscriptions, Anthropic chose a temporary rollout. From June 9 through June 22, Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost. Starting June 23, usage is expected to require usage credits unless Anthropic extends the included window because capacity allows it.
It is also worth checking the data policy. For Mythos-class models, Anthropic is requiring 30-day retention of business customer traffic on first-party and third-party surfaces. The company says this is for defending against complex attacks and reducing false positives, not for training new models. But if you operate under strict data requirements, this is not a footnote. It is a security-review item.
What this means in practice
I would not treat Fable 5 as "the model for every task." On short prompts, the difference may not be dramatic. The real value should appear where tasks used to collapse into dozens of clarifications: a project migration, analysis across a large document set, agentic work with memory, complex UI work from screenshots, or research with several dependent steps.
The most interesting test is to give the model a goal, not a question. Not "write a function," but "inspect the codebase, find the cause, fix it, verify it, and explain the risk." If Fable 5 really holds a longer horizon, that is where it should separate itself from earlier Claude models.
At the same time, this is a launch where restrictions are part of the product. Fable 5 promises more autonomy, but in some areas it will intentionally hand control to Opus 4.8. That is the new normal for frontier models: the stronger the model gets, the more important it becomes not only what it can do, but who can access which abilities, when, and under what mode.