The split turns the Claude safety classifier into a separately priced feature, and a new Claude API fallback option makes the friction a variable a builder can route around.
Anthropic released two flagship models on Tuesday, June 9, 2026: Claude Fable 5 and Claude Mythos 5. Anthropic says the two share the same underlying capability. The advertised difference is that Fable 5 ships wrapped in a safety classifier and Mythos 5 does not.
That single distinction is the story. Everything else is the cost of acting on it.
The split, per Simon Willison's writeup of the launch, is a deliberate product posture. A frontier lab is choosing to commercialize the safety layer as a separately priced feature rather than a default property of the model. Builders using the Claude API can now pick which version to buy. The catch is that Anthropic prices Fable 5 at $10 per million input tokens and $50 per million output tokens, with a 1 million token context window, a 128,000 token maximum output, and a knowledge cutoff of January 2026. Anthropic's pricing for Mythos 5 is the missing piece a builder would want before choosing.
The friction this creates is real, not theoretical. Willison, who had no early access and spent roughly 5.5 hours testing Fable 5 the day of release, reports that the stricter guardrails fire often enough to be a noticeable speed bump in practice. His verdict on the underlying capability is that Fable 5 is "something of a beast": slow and expensive, but able to handle everything he threw at it. The guardrail layer is the variable. The model underneath, in his read, is the same one shipping without the safety classifier in Mythos 5.
Two API additions shipped with the launch that make the split more than a marketing distinction. First, the Claude API now surfaces when a request has been rejected by the new guardrails, giving developers a programmatic signal they can log, alert on, or pass through to end users. Second, according to Willison, there is a new option to request automatic fallback to another model if a request gets rejected. That second piece is the one builders should pay attention to. Anthropic is not just shipping two models. It is shipping the mechanism to route around the safety classifier in production. Whether you read that as a feature or as a tell depends on what you think the classifier is for.
Willison's hands-on test surfaces one useful texture detail. In a project-list generation test, Fable 5 hedged in places where Opus 4.8 also hedged, suggesting the two models share a recognizable conservative instinct. That is not a benchmark verdict, and Willison does not present it as one. It is a signal that the underlying training produced a model with a stable behavioral fingerprint, and the safety classifier is layered on top of that fingerprint rather than baked into it.
The Mythos release is the part of the launch that deserves more scrutiny than it is getting. Anthropic positions Mythos 5 as sharing Fable 5's capabilities without the safety classifier. That framing raises a question Willison flags implicitly: if the underlying model is the same and the safety layer is detachable, what is the safety layer actually doing? It is not necessarily doing nothing. Classifiers catch requests the base model would happily answer, and they shape the model's behavior in subtler ways too. But the existence of a same-day twin without the classifier puts a spotlight on what the classifier adds that the base model does not already provide. Anthropic has not, in this launch, made that case in detail to independent reviewers.
For a builder deciding what to build on, the calculus now has three knobs: cost, latency, and how often the safety layer will intervene in the actual workload. The new fallback option makes the third knob less binary. A developer can pay Fable 5 prices, eat the latency, and route around rejections to a cheaper model when the classifier fires. They can also choose Mythos 5 if Anthropic prices it as a cheaper, classifier-free alternative, accepting responsibility for whatever the base model would have produced without the filter. Or they can stay on an earlier Claude and skip the question entirely.
The open question is whether other frontier labs follow the pattern. OpenAI, Google DeepMind, and the open-weight community have all shipped safety layers in different shapes: system prompts, RLHF, constitutional methods, post-hoc filters. None of them, to date, have shipped a same-day classifier-free twin of their flagship with the API plumbing to route around the classifier in production. If Mythos 5 gets traction, the bifurcated launch stops being an Anthropic quirk and starts being a template.
What to watch: Anthropic's published pricing for Mythos 5, which Willison's writeup does not include, and the first independent benchmarks comparing the two models on tasks where classifiers are known to interfere, like creative writing at the edges of policy. If the underlying capability really is the same, those benchmarks should show it. If they do not, the bifurcation story is more complicated than the launch framing suggests.