Anthropic shipped Claude Fable 5 today (June 9, 2026), and it's the first generally available model from a new tier the company calls Mythos-class — a rung above the Opus line. The short version for engineers: it's the strongest coding and agentic model Anthropic has released, it's cheaper than the preview that preceded it, and it ships with a new safety layer that occasionally routes your request to a different model. Here's what actually matters if you build software for a living.
One model, two names
Fable 5 and Mythos 5 are the same underlying model. The only difference is the safeguards. Fable 5 is the version you get everywhere, with classifiers that intercept a narrow band of high-risk requests. Mythos 5 is the unrestricted variant, available only to vetted cyberdefenders through Project Glasswing and, soon, a small set of biology researchers. For day-to-day development work, Fable 5 is the frontier model — Anthropic notes that in more than 95% of sessions, Fable behaves identically to Mythos because nothing triggers a fallback.
The API model string is claude-fable-5.
Coding is the headline
The claim Anthropic leads with is long-horizon autonomy: the longer and more complex the task, the bigger Fable 5's lead over previous models. The supporting anecdote is concrete. During early testing, Stripe ran Fable 5 against a 50-million-line Ruby codebase and had it perform a codebase-wide migration in a single day — work the company estimates would have taken a team more than two months by hand.
Two things stand out for production use. First, token efficiency: on Cognition's FrontierCode evaluation — which grades whether a model can pass hard coding tasks while meeting production-quality standards — Fable 5 scores highest among frontier models even at medium reasoning effort. That's a cost signal as much as a capability one. You don't always need to crank effort to the maximum to get a passing result, and lower effort means fewer output tokens.
Second, the early-access customer quotes are unusually coding-heavy. Cursor calls it state-of-the-art on CursorBench and credits it with opening up "a class of long-horizon problems that were out of reach." GitHub highlights complex, multi-turn coding tasks handled with more autonomy and reliability. Cognition's Scott Wu says it's the highest-scoring model on their frontier coding eval and "generalizes to unfamiliar tools out of the box" — which, if it holds up, is the interesting part for anyone wiring a model into a custom toolchain or MCP server.
Memory, vision, and the agentic story
The capabilities that make agents actually work get real attention here.
Long-context and memory. Anthropic says Fable 5 stays coherent across millions of tokens and improves its own output using notes it writes to disk. Their illustration: when the model played the deck-builder Slay the Spire with persistent file-based memory, the memory boost helped it three times more than it helped Opus 4.8. The takeaway for builders is that giving Fable a scratchpad or a memory file pays off more than it did with prior models — the model knows how to use persistent state.
Vision. Fable 5 is the new state-of-the-art on vision tasks. The demo that matters for web developers: it can rebuild a web app's source code from screenshots alone. It also needs far less scaffolding — it beat Pokémon FireRed start to finish on a vision-only harness, where earlier Claude models needed elaborate helper tooling. Less scaffolding generally means less brittle agent plumbing.
Knowledge work and research
Outside of code, the pattern repeats. On Hebbia's senior-level finance benchmark, Fable 5 posts the highest score of any model, with gains in chart and table interpretation. On the research side — and this is the Mythos variant doing the heavy lifting — Anthropic reports roughly 10x acceleration on parts of the protein-design pipeline, novel molecular-biology hypotheses that its own scientists preferred ~80% of the time in blind comparisons, and a week-long autonomous genomics project that produced a model outperforming a recently published one at 1/100th the size. Most of this sits behind the trusted-access programs, but it's a useful signal of where the ceiling is.
On alignment, Anthropic's automated assessment puts Fable/Mythos 5's misaligned-behavior rate low and comparable to Opus 4.8.
The safeguards you'll actually notice
This is the part to plan around. Fable 5 ships with classifiers covering three areas: cybersecurity, biology and chemistry, and distillation (attempts to extract the model's capabilities to train competitors). When a classifier fires, your request is handled by Claude Opus 4.8 instead of Fable, and you're told it happened.
The honest framing from Anthropic: the safeguards are deliberately tuned conservative, so they'll sometimes catch benign requests. They say fallback triggers in under 5% of sessions on average, and over 95% of sessions see no fallback at all. For most application development that's a non-issue — but if your work lives near security tooling, exploit research, or anything that looks like offensive cyber, expect to get bumped to Opus 4.8 sometimes. A fallback is a degraded experience, not a refusal: Opus 4.8 is still very capable, and your call still returns an answer.
There's also a data-retention change worth flagging for anyone with compliance obligations. All traffic on Mythos-class models — Fable 5 included, first- and third-party — carries a mandatory 30-day retention window. Anthropic says it won't train on this data or use it for non-safety purposes, logs all human access, and deletes after 30 days in almost all cases. If your data-handling policy currently assumes zero-retention inference, read the fine print before you switch.
Pricing and availability
Fable 5 is $10 per million input tokens and $50 per million output tokens — less than half the price of the earlier Mythos Preview. It's available everywhere as of today via the Claude API and on consumption-based Enterprise plans.
Subscription access is being staged because demand is expected to be high and hard to predict. From launch through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23, it moves to usage credits on those plans, with Anthropic saying it intends to fold Fable 5 back into standard subscription access once capacity allows.
Bottom line
If you're building agents, doing large refactors, or running anything long-horizon, Fable 5 is the obvious thing to test next — the token-efficiency story alone makes it worth benchmarking against your current Opus or Sonnet setup. Budget a little engineering time to detect and gracefully handle Opus 4.8 fallbacks, confirm the 30-day retention policy fits your compliance posture, and treat the June 23 pricing change as a calendar item if you're on a subscription plan rather than the API.
