OutYet reporting

Gemini 4 Neon’s benchmark lead depends on the workload

Google’s model card shows strong results across several tasks, while independent launch-day evaluations expose tradeoffs in price, speed, and task type.

OutYet Editorial Desk

Google’s November 18 announcement positions Gemini 4 Neon between Carbon and Hydrogen for complex coding, analysis, and enterprise work. The API price is $3 per million input tokens and $15 per million output tokens, with a one-million-token context window. Google reports an Artificial Analysis Intelligence Index score of 61 at xhigh and says Neon costs less per task than Carbon at its default high setting. The model card also lists four thinking levels, so users can choose a different balance of reasoning depth and spend.

The provider’s comparison table shows why a single overall score is not enough. On its cited evaluations, Neon leads Claude Opus 5.5 and GPT-6.2 Sol on AutomationBench and Agents’ Last Exam, but trails Opus on Terminal-Bench 4.1 and FrontierCode 2.0. Google’s SWE-sweep figure is 39.0% using its Antigravity harness, while the same card lists 12.8% on the Mini-SWE harness. These scores measure different setups, and the harness difference makes the headline figure a poor stand-in for expected performance in another coding environment.

Two independent write-ups add useful context, though their tests are not interchangeable. Eesel’s review reports Artificial Analysis results where Neon at high scored 59.6 for $3.74 per benchmark task, while Astra scored 60.1 for $2.07; it also notes that Neon generated substantially more output tokens. Spring Prompt tested five business benchmarks through OpenRouter at medium reasoning, rather than Google’s high default, and found Neon ahead on factual product listings but behind GPT-6.2 Sol on several other tasks. Spring Prompt says most of those task benchmarks ran once, so those results should be treated as an early signal rather than a stable ranking.

For teams choosing a model, the useful question is whether the setting, harness, latency, and cost resemble the work they plan to run. Neon’s million-token context and adjustable thinking levels may suit long inputs and jobs where extra reasoning is worth its output-token bill. The independent tests also show why price per token alone can mislead: task cost changes with output length and reasoning effort. Google’s benchmark card is provider-authored, and the independent samples are small or benchmark-specific, so a local evaluation remains necessary before relying on the reported advantage.

Related models

Sources