From a coding agent to a thought partner.

The frame

Fable 5 can work autonomously for longer than any Claude before it, stay focused across millions of tokens, and reflect on and validate its own work — so the unit you hand it grows from a turn to a whole task.

You drive the same way you always have. What changes is how much one prompt can carry.

What it means · 1 of 5

What it means · 2 of 5

Fable 5 plans the work, carries it across a long autonomous run, then reflects on and validates what it produced before handing it back. The million-token window means it keeps the whole task in view — so the bigger job, not the quick turn, is where it does its best work.

What it means · 3 of 5

“It really has this ‘big model smell’ that I haven’t felt before.” Boris Cherny · Creator of Claude Code at Anthropic

What it means · 4 of 5

The common thread across all three: the work that used to take a team a week now resolves in a run. That compression — not any single number — is the shift worth planning a year around.

What it means · 5 of 5

That is the whole mechanism, drawn to scale of its own data: per Anthropic’s early figures, about 95% of sessions run entirely on Fable 5, the safeguards trigger in under 5% on average, and they are intentionally over-broad at launch — they catch some harmless queries too. Some security researchers have noted the launch guardrails can block legitimate defensive work, consistent with that posture. Beneath it sits the documented floor: ASL-3 protections, classified CB-1 — capable with non-novel chemical and biological information, below the threshold for novel-weapon uplift, as reported by the safety documentation. The takeaway for a district is plain: a classifier plus a fallback to a more conservative model is a documented design, not a promise — a real floor under everyday work.

Anthropic’s post arrow_outward   How the routing works arrow_outward   On the safety design arrow_outward   The safeguards, reported arrow_outward

The receipts

Drag the dial. The two endpoints are Anthropic’s published figures; the curve between them — and the spend bar — are illustrative.

Thinking is adaptive: the effort parameter sets how deep Fable goes, and accuracy rises with it — from roughly 11.5% at low effort to 30.9% at max on FrontierCode Diamond. Pricing runs $10 per million input tokens and $50 per million output, so the dial is also where you tune the spend.

On cost

Boris’s arithmetic, drawn: tokens are sized by price, the counts illustrate one complex task. Some tasks land the other way — the point is that the bill follows the task, not the sticker.

So the sticker price reads higher, but the bill that lands depends on the task. Hand Fable the kind of work where its efficiency compounds, and the per-token premium can come back out in the wash — sometimes below what Opus would have cost.

How to use it

Kevin’s rule of thumb: if the work has a plan, a build, and a check — hand Fable the whole arc. If you’d just do it yourself in five minutes, you don’t need a model that runs for an hour.

The close

The close

A more capable instrument is not a replacement for the read of the person holding it. Fable 5 plans further and checks itself — but what to build, and whether it’s right for the room, is still the human’s call.