OutYet reporting
Gemini 4 Neon’s benchmark case depends on the task and harness
Google positions Neon between faster Hydrogen and deeper-reasoning Carbon. Its published comparison shows strong results across work and science tasks, with some coding and chart-reading measures still favoring competitors.
Google introduced Gemini 4 Neon on November 18 for complex software engineering, knowledge work, and science workflows. The model is available through the Gemini API and Google AI Studio, and Google says it is rolling out to select Gemini app, business, and Antigravity users. Google lists a one-million-token context window and API prices of $3 per million input tokens and $15 per million output tokens. That gives developers a concrete cost and context-size baseline, though actual spend will depend on request size, caching, and how much reasoning a task needs.
Neon fills a middle position in Google’s Gemini 4 lineup: the company describes Hydrogen as its fast everyday model and Carbon as its model for harder, longer-horizon work. Google’s launch post reports a score of 61 on the Artificial Analysis Intelligence Index at the xhigh thinking level and 70.9% on the Vals Index. These are provider-reported comparisons against named external evaluations, not a guarantee that Neon will be the best choice for a particular workload. The model card provides a broader benchmark table and notes that results vary with evaluation setup.
The comparison is mixed in ways that matter to coding teams. In Google’s model card, Neon scores 65.2% on Terminal-bench 4.1, just below Claude Opus 5.5 at 65.8%, while its 46.2% on FrontierCode 2.0 is below Opus 5.5’s 47.3%. On SWE-sweep, however, Neon records 39.0% with Google’s provider harness, compared with 10.1% for Opus 5.5 and 8.3% for GPT-6.2 Sol. Under the Mini-SWE harness, the scores shift to 12.8% for Neon and 11.6% for Sol. The contrast is a reminder that harness choice can change the apparent size of a model gap.
Neon’s strongest reported results are not confined to code. Google’s model card lists 56.3% on AutomationBench, ahead of the compared Opus 5.5 and GPT-6.2 Sol results, and 70.5% on Terminal-Bench Science. But the same table has Sol ahead on Chartography, a chart-reading benchmark, and Opus ahead on Terminal-bench 4.1 and FrontierCode 2.0. That makes the release more useful as a task-specific option than as a simple overall winner. Teams should match the benchmark to the work they actually run, especially when an evaluation uses a provider’s own harness.
For API users, the practical tradeoff is a substantial context window and a price point below Google’s higher-tier Carbon, paired with results that differ by task and evaluation method. Google’s model card is the most detailed source for those comparisons, while DataCamp’s independent review also flags the harness differences in coding results. Neither source establishes how Neon will perform on a team’s private codebase or documents, and benchmark scores cannot substitute for testing the prompts, tools, and latency constraints of a production workflow. The announcement and card describe intended uses and evaluations; they do not settle those deployment questions.
Related models
Sources
- Gemini 4 Neon Model Card · Google DeepMind
- Introducing Gemini 4 Neon · Google
- Gemini 4 Neon: Features, Benchmarks, Pricing, and Access · DataCamp