Benchmarks are one thing. How a model feels in real use is another. Since Fable 5 went live for everyone, the first people whose judgment I trust have put it through its paces — and the impressions are clear.
«Big model smell»
Simon Willison coined a phrase that nails it: Fable 5 has that «big model smell». He doesn’t just mean speed or cost, but how much the model simply knows. It feels like a big model — one that misses less on broad questions and brings more context than you’d expect.
The hard numbers behind it: a one-million-token context window, up to 128,000 tokens of output, and knowledge up to January 2026. Pricing sits at $10 per million input tokens and $50 per million output tokens — roughly double Opus 4.8. Fable 5 is strong, but it’s no bargain.
A real version jump
Ethan Mollick, who’s watched AI in practice for years, puts it even more plainly: Fable 5 «outperformed basically every other public model I have used by a considerable margin». And Andrej Karpathy, recently joined Anthropic himself, calls the release «super exciting» and a «major-version-bump-deserving step change forward». In the same breath, though, he warns: the model «still has quirks that people will run into», and the safeguards are set «a little too trigger-happy» at launch.
My take
This lines up with what I wrote at launch: the lead grows with the complexity of the task. For a quick one-off question, the premium isn’t worth it — Opus handles that easily. But on long, multi-step work, where deep knowledge and stamina have to come together, you feel the difference immediately. My advice: aim Fable 5 at the heavy lifting and leave the everyday questions to the cheaper model. And keep those oversensitive safeguards in mind — right now they still cost you some patience.
Sources: Simon Willison, Ethan Mollick, Anthropic