Anthropic has released Fable 5 and Mythos 5, the public face of the model class it had been teasing since the Opus 4.8 launch under the name Project Glasswing. The release is unusual in a way that has not been fully digested in the early coverage: it is not one new model but two, sharing the same underlying brain, separated by a safety routing layer that determines which one a given user actually gets to talk to.
On benchmarks, the new family is in a class of its own. Both Fable 5 and Mythos 5 outperform every frontier model currently in the public market by margins that are not within the usual statistical noise. They handle long-horizon autonomous tasks more reliably, use tokens more efficiently than previous Anthropic models, and post particularly strong numbers on knowledge work, vision, and long-context memory. Underneath the benchmark wins, the architectural choice is the more interesting story.
One brain, two doors
The shorthand is this: Fable 5 and Mythos 5 are literally the same model, served through two different access doors.
Mythos 5 is the unmediated version, the model Anthropic described in the Opus 4.8 launch post as requiring “stronger cyber safeguards” before general release. It remains private. Access is restricted to a small group of Anthropic partners selected for cybersecurity work, the same Project Glasswing program flagged in earlier announcements. Most users will never interact with it directly.
Fable 5 is the public version. It is the same model with a routing layer in front. In roughly 95 percent of conversations, talking to Fable 5 is functionally identical to talking to Mythos 5. The remaining 5 percent is where the door does its work.
The routing layer is a detection system. When a conversation appears to be heading toward sensitive categories, including cybersecurity, biology, or chemistry, or toward content that conflicts with the model’s alignment policies, the system intervenes. The conversation pauses. The user is either rerouted to Opus 4.8, which continues the conversation under a more conservative safety profile, or the conversation ends entirely.
Anthropic’s framing is that the underlying capability has not been blunted. The model has not been made less capable. A lock has been placed in front of it. The bet is that the lock holds.
This is a meaningfully different approach to safety than the industry norm. The standard pattern is to train a single model to refuse certain content while retaining the ability to discuss those topics in benign contexts. The Fable/Mythos pattern keeps full capability available but routes uncertain conversations to a different model entirely. Whether it proves more robust than the older approach is now an open empirical question.
The false-positive problem
The early experience with the routing layer suggests it intervenes more often than the 5 percent figure implies in practice. Users testing the model on routine work, including content with no plausible safety concern, have reported being rerouted to Opus 4.8 mid-conversation. One early tester, in the middle of iterating on a self-introduction page for the model, was rerouted while simply asking Fable to dial down its humor.
This matters because the rerouting is not silent. It changes which model you are working with, and Opus 4.8, while excellent, is not Fable 5. A task started in Fable can end up finished in Opus 4.8 without the user choosing the switch.
The companion question to the false-positive rate is the false-negative rate. Active communities have, historically, sustained efforts to jailbreak every frontier model that has been released. Fable 5 is too new to know how its routing holds up under sustained pressure. The next several weeks of public stress testing will be informative.
Pricing, access, and the things to read carefully
The access mechanics are worth reading carefully because they will change.
Through June 22, Fable 5 is available inside all paid plans: Pro, Max, Team, and enterprise tiers. The catch is consumption. Fable 5 burns roughly twice the credits of Opus 4.8 per equivalent task. Early users have reported running through 20 percent of their daily allowance in a single session because the model was given a vague goal and allowed to run.
From June 23 onward, the model moves to a pay-per-credit usage system. Anthropic has stated that Fable 5 will be removed from the standard subscription plans and eventually reintroduced once computing capacity catches up with demand. In practice, that means a window where Fable is metered separately, with costs scaling directly to use.
There is a third detail that has received less attention and probably should receive more. Anthropic does not normally retain conversations. For this release, conversations involving Fable 5 will be stored on Anthropic’s servers for at least 30 days. The stated rationale is monitoring the rollout of a model class new enough that the company wants visibility into how it is being used. Whatever the reasoning, users should know.
Treat Fable 5 like an expensive specialty tool until the dust settles. Two-times credit consumption, an access model that is about to change, and 30-day conversation retention together point toward selective use for high-value work, not adoption as a daily driver.
The usage pattern that actually works
The economic and architectural reality of Fable 5 points toward a specific way of using it: as the planner, not the executor.
The right default is to use Fable 5 for the part of a project where reasoning quality matters most. Strategy. Architecture. Design decisions. The plan for a complex implementation. Anything where a small improvement in reasoning translates into a large downstream payoff. Then switch to Opus 4.8 to actually execute the plan. Opus 4.8 is not a fallback. It remains one of the most capable models in production. It is also half the per-task cost of Fable 5 in the current pricing.
The cost-quality split this produces is favorable. Most projects spend the majority of their tokens on execution. By keeping execution on Opus 4.8 and using Fable 5 only for the planning steps that benefit from extra reasoning, the bill stays manageable while the work still benefits from frontier-level thinking where it matters.
How to prompt Fable 5
The early consensus on getting good results from Fable 5 converges on four habits.
Be direct. Fable 5 slows down when given vague context. Stating the goal, the constraints, and the expected output crisply tends to produce better results than narrative framing.
Name the tools you want used. Fable 5 is noticeably more conservative than Opus 4.8 about deploying its full toolkit. If you want web search, sub-agents, or dynamic workflows, ask for them explicitly. Otherwise the model is likely to default to a more restrained execution path.
Give it a stopping criterion. The model can run for a long time on open-ended goals. A long run means a large bill. Specifying when the work is done, in concrete terms, is how you keep the meter under control.
Force it to ask questions rather than invent. A capable model with insufficient context will produce something plausible by inferring details. With Fable 5’s cost structure, that “something plausible” can become expensive. Constraining the model to ask clarifying questions before generating anything substantial pays back.
The unifying principle is constraint. Fable 5 rewards specificity in goals, in tools, in stopping conditions, and in input. The frontier capability is genuine. The cost of an unconstrained Fable 5 session is also genuine.
What Fable 5 can actually build
The benchmark numbers are abstract. The hands-on demonstrations are more useful for understanding the capability level.
Early testing has produced a meditation and breathing application complete with cardiac coherence exercises, guided sessions, ambient sound, and a working five-minute breathing animation, all generated from a single prompt. The output bears a clear visual resemblance to Headspace and functions as a complete app.
A second test built a 3D roller coaster construction tool in a cartoon visual style, with the ability to place obstacles, add loops, and switch into a first-person ride view. The result is rough on design but mechanically functional, including physics constraints.
A third test produced a living 3D city with buildings, traffic-aware cars, pedestrians, day-night cycles, weather changes, and a user-triggered blackout. The visual polish is noticeably above what single-prompt generation from previous-generation models could produce.
The most distinctive test was a browser-based operating system with draggable, resizable windows, a working file explorer, a notepad, a paint application, and one “surprise” application chosen by the model itself. The result was a fully interactive desktop with all the requested features, plus, as the surprise, a working Minitel emulator. Built from a single prompt.
None of these is a polished product. They are demonstrations of one-prompt construction capability that did not exist at this level six months ago, and the gap with the next-best model is now meaningful enough to be visible without instrumentation.
Frequently Asked Questions
What is the difference between Fable 5 and Mythos 5?
They are the same model. Mythos 5 is the unmediated version, restricted to a small set of Anthropic partners doing cybersecurity work under Project Glasswing. Fable 5 is the public version, served with a safety routing layer in front of it. In roughly 95 percent of conversations, talking to Fable 5 is functionally identical to talking to Mythos 5.
What happens when Fable 5 reroutes a conversation?
The safety routing layer monitors for sensitive categories, primarily cybersecurity, biology, and chemistry, and for content that conflicts with the model’s alignment policies. When triggered, the conversation either continues with Opus 4.8 under a more conservative profile, or ends. False positives are common in early use, so it is worth knowing the rerouting happens and what it changes about your session.
How much does Fable 5 cost compared to Opus 4.8?
Roughly twice the credit consumption per equivalent task. Through June 22, it is available in paid plans (Pro, Max, Team) with no additional charge beyond the higher consumption. From June 23, it moves to a pay-per-credit usage model and is removed from standard plans, with reintroduction expected once compute capacity catches up.
Is my conversation with Fable 5 stored?
Yes. For this release, Anthropic is retaining Fable 5 conversations on its servers for at least 30 days. This is a departure from the company’s usual policy. The stated rationale is monitoring the rollout of a new model class. Users handling sensitive content should factor this in.
Should I use Fable 5 for everything?
No. The pattern that produces the best cost-quality outcome is using Fable 5 for the planning and reasoning steps of a project (strategy, architecture, hard decisions) and switching to Opus 4.8 for execution. Opus 4.8 remains one of the most capable models in production, and the per-task cost difference adds up quickly.
How do I get good results from Fable 5?
Four habits consistently improve outcomes. Be direct rather than narrative. Name the tools you want used (web search, sub-agents, dynamic workflows) because the model is more conservative than Opus 4.8 about deploying them. Give it a clear stopping criterion to prevent runaway costs. Constrain it to ask clarifying questions rather than infer missing context.
