Sonnet finishes 7 out of 10 jobs on its own. The old Sonnet finished 1.
This test is called Terminal-Bench. The AI gets a real job, like "build this" or "fix this." Nobody helps it. It only counts if the end result actually works.
Customers have to get their "low balance" alert within 5 seconds. If the alert shows up late, the job fails.
Every new lead has to be checked and saved correctly, with no duplicates.
Pages have to load under a time limit, and nothing is allowed to break.
Decide which claims to approve or reject, like a real claims auditor would.
A skilled person would need hours for each one. A checker tests the result at the end. It passes only if it works.
It even edged out Opus 5.5, the bigger model that costs twice as much. Opus finished about 6.6 out of 10.
Both scores use each model's highest effort setting
An app builder called Base44 said it rarely stopped in the middle of a build to wait for an answer. So fewer builds get stuck.
The old Sonnet did things one at a time. The new one groups them together, so it finishes in fewer steps.
Testers said it quickly understood big existing projects, and kept going on tasks that ran for hours.
- Higher up means it finished more jobs.
- Further right means each job cost more money.
- Find "Med" on the blue line. That's the default setting. Each job costs under $1, and it still beats the old Sonnet's best score, which cost over $10 a job.
