Cursor is speechless! Claude Fable 5 sets a new record with 72.9% on a brutal internal benchmark

Hi, it's me, Shiichan!

Claude Blog claude.com

What was announced?

Claude's blog just published a story about how Cursor evaluated Claude Fable 5 for real engineering work. Cursor is a coding agent that supports all the major frontier models, and Nate Schmidt, an engineer who works on evals there, walked through the results of CursorBench, the team's own internal benchmark. Claude Fable 5 hit 72.9% on CursorBench's Max effort setting — a new high score.

Why it matters

Schmidt says public benchmark scores don't always line up with whether developers actually accept a model day to day. Many evals present a clean, well-defined problem with clear constraints to fix. Real prompts from actual users rarely look that tidy.

That's exactly why Cursor's team built CursorBench to capture that messiness — tasks with nothing but a stack trace and the word "fix," or tasks where the model is fed a wrong module name to see whether it pushes back or just goes along with the user's hunch. Scoring 72.9% on a benchmark built like that carries real weight.

What changes

The standout part of the story is what Schmidt calls "global reasoning." Where a lot of models lean on "local reasoning" — thinking mainly about what just happened and what comes right after — Fable 5 steps back and reasons about the whole mission before acting.

Schmidt put it this way: "If you're at A and you have no idea where B is, Fable is an excellent choice." Cursor is now splitting work accordingly — lighter models handle tasks with a clear path, and Fable 5 gets reserved for the genuinely hard problems where the route isn't obvious. Looking ahead, the team plans to let Fable 5 run backend systems unattended for days or weeks at a time, hunting for performance bottlenecks and user pain points before anyone reports them.

Dive Deep

The clearest illustration of global reasoning is a space-flight simulator test with a one-line prompt: "build a rocket that lands on the moon."

• Claude Opus (the older model): spent 12 to 16 hours iterating without ever landing. It kept cycling between adding fuel when it ran short and then finding the rocket too heavy to clear the stratosphere once it did.
• Claude Fable 5: succeeded in just a few hours. Instead of aiming straight for the moon, it plotted a multi-step strategy on its own — first an orbital mission to gather telemetry, then a moon-landing mission informed by that data.

"I don't feel like I have to bootstrap Claude Fable 5 to understand the world I exist in and the problem I'm trying to solve."

Schmidt says starting from that kind of situational awareness is a big part of why the model can move so much faster through the hardest, messiest problems.

Wrap-up

• Claude Fable 5 set a new record of 72.9% on CursorBench's Max effort setting, Cursor's internal benchmark.
• Its strength is "global reasoning" — solving problems without a predefined correct path.
• In a space-flight simulator test, Claude Opus failed for 12-16 hours trying to land on the moon, while Claude Fable 5 pulled it off in a few hours using a multi-step strategy.
• Cursor plans to push Fable 5 further, including unattended backend tests that run for days or weeks.

If you're the kind of engineer who wants to hand off a problem when you have no idea what the path forward looks like, this one's worth a read!