Hand it a spec, and
the migration finishes overnight.
Until now, the brain behind Claude Code was Opus 4.8. On June 9, a new model joins it: Claude Fable 5. The aim is autonomous execution that keeps running for hours on end. Anthropic itself says it compressed a code migration that took roughly two months down to a single day—we unpack that shift with diagrams.
It got faster.
But the "smarts" stayed put
Claude Code is a coding agent that runs in your terminal. Its core model so far has been Opus 4.8. That was plenty for everyday code completions and questions, but on tasks that need to keep running for hours—large refactors or migrations spanning many files—some users reported limits: losing the logical thread partway through, or dropping the context of the instructions.
May's Colossus-1 deal doubled the rate limits, so you could "throw more at it." But the model's own reasoning power stayed at Opus 4.8. In other words, the "number of times" went up while the "smarts per run" did not—that was the situation until now.
| Until now (Opus 4.8) | After Fable 5 is added |
|---|---|
| The only core model is Opus 4.8 | Fable 5 is also selectable |
| Rate limits doubled (under May's deal) | Reasoning power itself is raised |
| Reports of dropped context on long tasks | Designed for long-horizon autonomous execution |
| Parallel generation from specs was trial-and-error | Parallel tasks generated from multi-page specs |
Running fast and staying on course for a long time are
two different abilities.
On June 9, Fable 5
landed in Claude Code
The new model released yesterday (6/9 PT) is now usable in Claude Code too. The intent is clear: a design that pays off the longer you let it run.
Hand it the spec
You feed it the multi-page spec describing "what to change and how," as is. The starting point is to cut the manual work of slicing it into tasks by hand.
Break it into parallel tasks
With Fable 5's reasoning power, it automatically splits the spec into several tasks that can advance independently, and runs them at the same time. This flow only works because the design assumes long-horizon autonomous execution.
Turn it into code in one sweep
The agent writes each task forward and runs it through verification. What matters is "how long it can keep running" more than how long the output is—so the larger the migration, the bigger the payoff.
The claim: a two-month
migration became one day
As supporting evidence, Anthropic points to a real code migration it ran in-house. The numbers are striking, but how you read them matters.
"Claude Fable 5," released June 9, 2026, is now available in Claude Code as well, supporting long-horizon autonomous execution and parallel task generation from multi-page specs—that is the heart of this update. As its own case study, Anthropic says it cut a Stripe code migration from roughly two months to one day.
That said, this is a showcase the vendor chose itself. The preconditions and the nature of the migration may not match your own project, so it's healthy to take it with a grain of salt. Even so, the direction—"hand over a spec and it writes the code in parallel"—has clearly become more plausible thanks to the boost in reasoning power.
Who it helps, and who won't notice
Even with the same update, the experience changes completely depending on how you use it. First figure out which side you're on.
The long-session crowd
For engineers who run large refactors or migrations several times a week, the ability to keep running pays off. This is where the experience really changes.
Spec-driven PMs and teams
For teams that want to advance several tasks in parallel starting from a spec, the bridge from planning to implementation becomes a realistic option.
People with only short use cases
If you mostly use completions or one-off questions, the difference between models is basically noise. There's little reason to force a switch—your usual workflow is enough.
"Smarts" and "volume"
finally line up
May's 2× rate limit was an expansion of "quantity." This Fable 5 update is an expansion of "quality." Only when both come together does the use case of long-horizon autonomous execution start crossing into the practical zone. Hand over a spec and leave the rest to it—that picture is starting to become something worth considering, not an exaggeration. That's the significance this time.
Even so, rather than swallowing the vendor's numbers whole, the sound move is to run it once on your own typical task and see for yourself. Whether it pays off depends on how long you keep Claude Code running.