06Benchmark · updated 9 November 2026
Which AI model completes a work task best?
How well an AI agent completes an entire work task, in real software.
How to read it
- Each bar is one model. Under the name is the software that runs the model (for example Codex or Claude Code). That is called a harness: the agent's working environment.
- The test consists of 122 tasks from the work of domain experts: multiple steps, in real software, with a result that a computer checks.
- There are two figures. The pass rate only counts a task if everything is correct. The score also counts how much is correct. A report with ten parts of which nine are correct counts as failed in the pass rate and as 90% in the score. Higher is better.
- Switch between the two at the top right, or choose "Beide" (both): the hatched bar is the score, the solid bar the pass rate.
- Point at a bar for the details, including the cost per task.
What it means for you
Even the best agents do not fully succeed at every task. A process that is 90% right still has to be checked. So look mainly at the pass rate and build in control steps. The chosen model makes more difference than the software around it, so compare models for your own process.
More benchmarks
Also take a look at the other benchmarks
Next
A benchmark does not choose a model for you. We do.
Which model is best depends on your work, your data and what it may cost per month. In an AI Partnership we choose per use case, and switch when a better model comes out.
More on how we approach that: the AI Partnership.
Free to use: all benchmarks are licensed under CC BY 4.0. Reuse or quote them, crediting: NinA AI Benchmarks, nina-ai.nl/benchmarks.
Which model fits you?
Let us compare them on your work.
A 15-minute intro call. Afterwards you know which models and which approach fit your processes.
- 160+ organizations
- Sessions rated 9.3
- Amsterdam
