What do the published results show?
These are results from separate DeepMind model cards. Matching labels make them useful for screening candidates. Release dates, thinking settings and harnesses can still differ.
| Evaluation | Neon | Argon |
|---|---|---|
| RiemannBench | 84.0% | 76.0% |
| Chartography | 73.5% | 71.6% |
| LVBench | 89.0% | 91.7% |
| GraphWalks (256k–1M, BFS F1) | 95.2% | 84.2% |
Sources: DeepMind neon model card and DeepMind argon model card. Results were not independently reproduced.
Where the comparison breaks down
Terminal-bench 4.1 and 4.0 are different versions. CWE-Bench v2 and v1 also differ. The older Agents’ Last Exam pass rate should not be equated with the Neon card’s v1.1 score. Argon leads on the LVBench row, so a general recommendation for Neon does not settle video-heavy work.
Compare on the work you actually do
- Repository maintenance
- Give both models the same issue, starting commit and test suite. Compare whether the patch passes tests, how many follow-up prompts it needs and time to an accepted diff.
- Long research handoff
- Give both models the same source pack. Count unsupported claims and missing citations before comparing completion time.
- Batch document work
- Score factual accuracy and required fields on the same set of files. Log total cost per accepted output, including retries.
What Google published
Google describes Argon as a model for deep reasoning and Neon as an option for complex, high-frequency work. The Neon model card also compares certain capabilities with Argon in a safety assessment. That comparison does not provide an overall quality ranking for your workflow.
DeepMind Argon model card · Benchmark worksheet