Anthropic shipped the Claude 5 family this year, and Fable 5 is its flagship: the first model in a new tier that sits above Opus, the strongest model Anthropic makes generally available. We've been running it inside real delivery work since release. This is what's actually different, from practice rather than from the launch page.
A different class of model
Model generations usually feel like the same assistant with better grades. Fable 5 feels like a different colleague. The improvements that matter aren't benchmark points; they're behavioural:
- It holds a large codebase in its head without losing the thread, so a change made in one corner respects decisions made in another
- It plans long, multi-step work and actually finishes it, instead of stalling halfway through with something plausible
- It verifies as it goes: running builds, checking its own diffs, and saying plainly when something failed rather than declaring victory
- It pushes back. Given a flawed instruction, it flags the flaw instead of politely implementing it
That last one is underrated. An assistant that executes bad ideas quickly is a liability. One that catches them is leverage.
What it means for software teams
The ceiling on what AI assistance can carry just moved. Work we used to slice into small supervised steps (a large refactor, a migration across dozens of call sites, a review of an unfamiliar subsystem) can now be delegated in bigger pieces, with verification at the seams instead of hand-holding in the middle.
That doesn't reduce the need for experience; it concentrates it. The engineering shifts toward the decisions the model can't make for you: what to build, what good looks like, what risk is acceptable. The typing was never the hard part.
How we adopted it
New models don't touch client work here on launch day. Fable 5 ran on internal work first, side by side against the previous generation on tasks we knew well, and moved into delivery only after it cleared that bar. It cleared it faster than anything we've evaluated.
Senior engineers still review every line that ships. The model got better; the guarantee stays the same. If you want to see what this generation of tooling looks like on your codebase, that's exactly what a workshop is for.