Choose evidence for your workload
- Repository changes
- Start with CursorBench and SWE-sweep. Check the agent harness because tooling can affect the outcome.
- Terminal workflows
- Review Terminal-bench and check which metric is reported. Record your own commands, timeouts and manual interventions.
- Charts and documents
- Look at Chartography for chart understanding. For production extraction, test the exact documents and schema you use.
Run a fair comparison
Keep prompts, tools, test fixtures and acceptance criteria fixed. Record the model ID, thinking level and date. Separate wall-clock time from review time, and compare cost per accepted result. Keep failed and retried runs in the log.
Download the benchmark worksheet (CSV)
Source record
Reviewed November 19, 2026. The model card is the source for benchmark coverage. This site has not run or reproduced the published scores.
Primary sources
Google launch announcement · DeepMind Neon model card