Is Claude Fable 5.1 Actually As Good As Claimed?

My first reaction the day Anthropic announced Fable 5.1 was skeptical, because "state-of-the-art on nearly all tested benchmarks" is a sentence we hear at every launch now, and it starts to sound like a cliché. But after going through the announcement and the published data line by line, I think there's genuinely something concrete behind part of the claims — not all of them, but an important part.

The most striking claim is about long-running tasks. According to Anthropic's own announcement, Fable 5.1 can stay focused across hours-long tasks that span an entire codebase — code review, performance work, not in one sitting but sustained across a session that runs for hours. They shared a customer example: a codebase migration that would normally take two months got done in 24 hours. I always read numbers like that with some distance, because company-shared examples tend to be best-case scenarios — but if the underlying claim is true (that the model doesn't forget a decision made three or four steps earlier and then act inconsistently), that solves exactly the thing that usually erodes trust in a long AI session.

On Cognition's FrontierCode evaluation, Fable 5.1 reportedly scored highest among all frontier models at medium effort levels. On analytics tasks there's a claimed 10-point jump over the previous Opus model, crossing 90% on complex analytical work. That's interesting to me because analyzing why a system (say, a reconciliation system stitching together eight different payment providers) was designed a certain way is a different skill from just writing code — and real progress is being claimed here.

There's also an interesting test: the model reportedly completed the game Pokémon FireRed using vision input alone, with no extra scaffolding — something earlier Claude models couldn't do without additional help. That might not look directly relevant to daily work, but it's actually a meaningful signal: the model can work toward a long-horizon goal, periodically checking back on where it stands — which is exactly what "what was the current state, where was I going" means in a software project.

On pricing: $10 per million input tokens, $50 per million output — Anthropic positions this as less than half the price of the previous "Mythos Preview" model. I don't think price should be judged in isolation, but alongside how many tokens it takes to finish a task; a cheap-but-token-hungry model can end up costing more than an expensive-but-efficient one. I don't have an independent measurement that puts both of those together yet, honestly — that's the kind of thing that takes time to become clear.

My overall take: Fable 5.1's strongest claim is making a real difference specifically in long, multi-file work where "not losing context" matters. For short, one-off questions, you won't notice much difference no matter which model you use — the real test is whether the model still remembers what it's doing after hours in the same session. The published numbers are promising on that front, but it would be premature to call it settled without seeing independent user experience beyond the company's own examples.