Google DeepMind reports that Neon scored 84.0% on RiemannBench and 39.0% and 12.8% on SWE-sweep in two different agent environments, so consider both the reasoning tier and the agent environment when choosing a model.
Tasks these results can help assess: Mathematical and scientific reasoning, agent tasks in codebases, chart and video understanding, and graph traversal in long contexts.
What requires separate evaluation: Success rates on real-world business tasks, factual accuracy, latency, cost, and performance in other clients.
Applicable model version: Gemini 4 Neon, adapted from Gemini 4 Carbon. The evaluation materials do not specify the API model revision for each benchmark.
Test environment or client: The methodology uses the Gemini API by default and documents the agent environment, tool configuration, or source of public scores for each benchmark.
Reasoning tier and parameters: The methodology says it uses the highest thinking setting by default, without naming the default tier. FrontierCode 2.0 explicitly uses High. The materials do not specify sampling parameters such as temperature.
The official results are current as of November 2026. By default, the methodology reports the single-attempt pass rate (pass@1), with majority voting and parallel inference disabled for single-attempt evaluations. Some small benchmarks average results across multiple runs to reduce variability. The model card combines DeepMind's own evaluations with external benchmark scores. The table gives the setup for each result. Official evaluation methodology PDF
| Task | Benchmark | Neon | Source and configuration |
|---|---|---|---|
| Math | RiemannBench | 84.0% | Benchmark's official leaderboard |
| Science | Terminal-Bench-Science 0.1 | 70.5% | Evaluated by DeepMind using mini-swe-agent vision, with the reward averaged over 5 runs per question |
| Agentic coding | CursorBench 4.1 | 65.4% | Score reported by Cursor |
| Agentic coding | FrontierCode 2.0 | 46.2% | Benchmark's official leaderboard, with Neon using High |
| Agentic coding | SWE-sweep | 39.0% / 12.8% | Evaluated by DeepMind using Antigravity CLI 1.5.9 and mini-swe-agent vision, respectively, with scores weighted by subtask |
| Chart understanding | Chartography | 73.5% | Evaluated by DeepMind using Surge's public leaderboard methodology, with tools disabled |
| Video understanding | LVBench | 89.0% | Evaluated by DeepMind with tools disabled, with Gemini sampling 1 frame per second at low resolution |
| Long context | GraphWalks | 95.2% | Evaluated by DeepMind on 200 questions with contexts ranging from 256k to 1M tokens, using F1 for breadth-first search |
The scores come from the model card's Evaluation table, and the configurations come from the Methodology and Additional Details sections of the companion PDF. The model card specifies a combined input and output context limit of 1M tokens. The default synchronous output limit is 64K tokens. Long decoding in the background can output up to 1M tokens, subject to the shared context limit. Model card · Official evaluation methodology PDF
The two SWE-sweep scores each correspond to a specific agent environment. The 95.2% GraphWalks score is an F1 score for graph traversal over the specified range of context lengths. Accuracy on real-world long-document question answering needs to be measured separately. The benchmarks define their scores differently, so each percentage is meaningful only within its task and scoring method.
The model card notes that Neon can hallucinate. Its knowledge cutoff is March 2026, and its knowledge in some areas may be older. The science, multimodal, and long-context results apply to their respective test sets. Fact-checking still requires consulting primary sources.
To verify the results, use the November 2026 results table in the model card and check the PDF for each benchmark's run environment, tools, and scoring method. If rerunning the evaluations, record the model revision, thinking setting, benchmark version, agent environment, and tool permissions. The public materials let you verify aggregate scores and disclosed configurations. They do not provide per-question results, complete invocation parameters, or confidence intervals for every benchmark.
Gemini 4 Neon