Claude Fable 5: Anthropic's Most Powerful Public Model, and What It Actually Changes
Anthropic put a Mythos-class model in public hands with Claude Fable 5. Past the benchmark wins in software, knowledge work, vision, and science, the real shift is what happens to long-horizon agentic work and how safety is now built in by fallback.
On June 9, 2026, Anthropic did something it had spent months signaling it was nervous about: it handed the public a frontier model from the tier it had previously kept behind closed doors. Claude Fable 5 is the safe-for-general-use version of a Mythos-class model, a step above the Opus line, and by the company’s own account it is state of the art on nearly every benchmark it was tested on. Its restricted twin, Claude Mythos 5, is the same underlying model with the cyber safeguards removed, available only to a small set of vetted defenders and government partners.
The benchmark headlines will get the attention. The more interesting question is what changes for the rest of us when a model this capable becomes something you can call from an API for the price of a normal request.
“The gap between the model labs keep for themselves and the model you can build on just narrowed to almost nothing.
”
What was released
There are two models, and the distinction matters. Fable 5 is generally available right now, callable as claude-fable-5 on the Claude API and through AWS Bedrock, Google Vertex AI, and Microsoft Foundry. Mythos 5 is not public; it is offered in limited availability to approved partners in Anthropic’s Project Glasswing, aimed at cyberdefenders and infrastructure providers who need the safeguards loosened to do their work.
The pricing is the part that should make builders sit up.
Per million input tokens
Per million output tokens, under half the price of the Mythos preview
Of sessions where the safeguards trigger a fallback, on average
A frontier-class model at that price is not a research demo, it is an economic event. It changes which products are viable, because the thing you could not afford to run on every request last month you can run on every request now.
What it is good at
Anthropic’s announcement is specific about where Fable 5 pulls ahead, and a pattern runs through all of it: the harder and longer the task, the bigger its lead over previous models. That single sentence is the most important one in the whole release, and I will come back to it.
Software engineering
The company reports it is the most token-efficient Claude yet and tops Cognition’s FrontierCode benchmark. The flagship anecdote: a partner ran a codebase-wide migration on a 50-million-line Ruby codebase in a day, work quoted at two months by hand.
Knowledge work
Highest reported score on Hebbia’s finance benchmark, with strong document, chart, and root-cause reasoning, validated on trading-analysis tasks. This is the unglamorous office work that quietly runs companies.
Vision
A new state of the art, pulling precise numbers out of scientific figures, rebuilding a web app’s source from a screenshot, and clearing a game with a vision-only harness. Reading the world, not just text about it.
Memory and long context
It stays coherent across millions of tokens and improves itself using its own written notes, with file-based memory reportedly tripling its results on a long-horizon game relative to Opus 4.8.
Life sciences (via Mythos 5)
The restricted model reportedly designs proteins at the level of skilled humans, generates novel research hypotheses experts prefer most of the time, and ran an autonomous genomics project that produced a far smaller model beating a published result.
Safety as a fallback, not a refusal
The most engineering-interesting design choice is how the safety works. Instead of the model simply refusing in high-risk areas, new classifiers watch for misuse and jailbreak attempts in domains like cybersecurity, biology, chemistry, and model distillation. When something trips a classifier, the request quietly falls back to Claude Opus 4.8, the previous-generation model, and the user is told it happened.
“You do not get refused. You get a slightly less capable model for the dangerous slice of work, and full power for everything else.
”
The reported numbers say this is tuned to stay out of the way: the fallback triggers in under five percent of sessions on average, so more than ninety-five percent of the time you are getting full Mythos-class performance. Anthropic says it red-teamed the model hard, with a thousand-plus-hour bug bounty that found no universal jailbreak, though an external safety body reportedly made early progress on one. It is a more honest posture than pretending the model is harmless, and a more usable one than blanket refusals. There is also a data condition attached to the high-capability traffic: a 30-day retention window, logged human access, deletion after the window, and use limited to defending against attacks rather than training.
So what do you actually do with it
The temptation with a release like this is to swap the model name in your config and call it a day. That leaves most of the value on the table, because Fable 5’s strengths point at a different way of building.
Give it longer tasks, not just harder ones
Since the lead grows with task length, the win is to hand it work you would previously have chopped into tiny supervised steps. Let it run the whole migration, the whole audit, the whole research pass, and check the result.
Lean on the memory
A model that self-improves from its own notes across millions of tokens is built for persistent, file-based memory. This is the same pattern that makes agent loops compound, and it now has a model that rewards it.
Re-run your unit economics
At this price, calls you rejected as too expensive may now be the cheapest part of your stack. Recompute cost per finished task, not cost per call, especially inside anything that loops.
Plan around the fallback, not against it
If your product touches security or bio-adjacent work, expect the occasional drop to Opus 4.8 and design the experience so it degrades gracefully instead of breaking.
The honest caveats apply as they always do. The headline capability claims are Anthropic’s own, measured on benchmarks the company chose, and benchmarks are not your workload. The subscription access has already been bumpy: Fable 5 was bundled into the Pro, Max, and Team plans for free through June 22 and then pulled on June 23 in favor of usage credits, which tells you the capacity is tight and the economics are still being worked out. None of that changes the core fact, though. A model from the tier labs were treating as too powerful to ship is now a line in your config, priced like an ordinary one.
