Fable 5, the most capable AI ever released to the public.

On June 9, Anthropic did something it had never done before: it handed the public a model it had previously reserved for cyberdefense teams working alongside the US government.

That model is Claude Fable 5, the first publicly available Mythos-class model: a new tier above Claude Opus, until now the top of the range. In Anthropic’s own words, “Fable 5’s capabilities exceed those of any model we’ve ever made generally available.”

At Veya, we read every model announcement with one question: does this change what a Brain can carry for its owner? This time, the answer is an unambiguous yes.

Where it comes from. In April, Anthropic quietly gave its first Mythos-class model to a small circle of cybersecurity organisations, judging it too capable for open release. Fable 5 is that same class of intelligence made safe for everyone. It shares its underlying model with the restricted Claude Mythos 5, but ships with a new layer of safety classifiers: when a request touches sensitive ground (offensive security, certain corners of biology and chemistry), it is answered by Claude Opus 4.8 instead, and you are told. Anthropic reports this happens in fewer than 5% of sessions.

The name is a quiet wink: fable comes from the Latin fabula, “that which is told” — a cousin of the Greek mythos. The same story, told safely.

The headline is autonomy, not intelligence. Every model generation gets smarter on benchmarks. What sets Fable 5 apart is how long it can work alone, and Anthropic is blunt about it: the longer and more complex the task, the wider its lead.

The early-access numbers read like fiction. Stripe watched it migrate fifty million lines of Ruby in a single day — work estimated at two months for an entire engineering team. Hebbia, whose benchmark tests senior-analyst financial reasoning, recorded the highest score it has ever measured. A leading legal AI company ran blind reviews in which its own lawyers found Fable 5’s contract redlines matched or beat their production system every single time.

And in a demonstration equal parts charming and significant, it finished Pokémon FireRed from raw screenshots alone: no maps, no helper tools, no game-state data. That is not a party trick. It means the model can operate software the way a person does, with far less scaffolding holding its hand.

What changes for an expert practice. Until now, delegating to AI worked best on bounded tasks: classify this email, extract this clause, draft this summary. Long-horizon reliability changes the unit of delegation from tasks to projects. A full due-diligence document review. A multi-source research memo. A quarter’s worth of reporting reconciliation, planned, self-corrected, finished.

Document work gets a direct upgrade too. Expert practices run on PDFs, scans, charts and tables, and Fable 5 is the new state of the art on vision: it extracts precise figures from dense documents, and it once rebuilt a web application’s source code from screenshots alone.

The economics reward architecture, not blind upgrades. Fable 5 costs double what Opus 4.8 does. The wrong response is to route everything through it. The right one is the tiered architecture a well-built Brain already uses: fast, inexpensive models for routing and classification, Opus-class models for everyday reasoning, and Fable 5 reserved for the genuinely hard, long-running work where its lead is widest. Remember Stripe’s two months compressed into a day; that is where the price pays for itself.

The honest word on safeguards. Do the new classifiers get in the way of normal business use? Almost certainly not. They cover offensive cybersecurity, sensitive biology and chemistry, and attempts to clone the model into competitors. They are tuned conservatively for launch, so an occasional false positive will happen — and even then the fallback is not a refusal, it is Opus 4.8, itself a frontier model. For legal, financial and consulting work, the practical impact rounds to zero.

One change does belong in your vendor review: Mythos-class traffic now carries a mandatory 30-day data retention period, used solely for safety monitoring, never for training, with logged human access and deletion after thirty days. Well-bounded, but new.

The window. Fable 5 is available now via the Claude API, Claude Code and the major clouds. Until June 22 it is included in Pro, Max, Team and seat-based Enterprise plans at no extra cost; after that, access shifts to usage credits while capacity catches up with demand. If you want to evaluate it on your own work, the next two weeks are the moment.

Most model releases deserve a fraction of their hype. This one is different, not because the benchmarks impress (they do), but because of what just crossed into general availability: an AI that can carry long, complicated work to completion with minimal supervision. For practices that already own a well-architected Brain, Fable 5 slots in as a more powerful engine. For those still doing everything by hand, the gap just widened again.