OutYet reporting

Gemini 4 Neon makes a cost and speed case, with benchmark caveats

Google positions Neon between its everyday and flagship Gemini models. Its model card confirms the Carbon lineage and one-million-token context, while much of the performance case rests on provider-reported evaluations.

OutYet Editorial Desk

Google introduced Gemini 4 Neon on November 18 for complex coding, analytical work, science, and enterprise workflows. The API price is $3 per million input tokens and $15 per million output tokens, with cached input priced at $0.15 per million. Google says the model is rolling out through its API, Google AI Studio, Gemini for business, Antigravity, and the Gemini app, though access depends on the product and subscription tier. This is a practical mid-tier pitch: teams can try a model aimed at substantial multi-step work without defaulting to Google’s larger Carbon model.

The model card adds a useful detail that the launch post’s positioning alone doesn’t convey: Neon is adapted from Gemini 4 Carbon. It accepts text, images, audio, and video in a combined context window of up to one million tokens, and supports up to 64,000 synchronous output tokens by default. The card also lists a March 2026 knowledge cutoff and says Neon may hallucinate; Google’s full safety and security report was still forthcoming when the card was published. Those disclosures matter for users weighing long-document capacity against freshness and reliability.

Google’s comparisons are promising but need to be read at the settings and harnesses used. Its model card reports an Artificial Analysis Intelligence Index score of 61 for Neon, versus 59 for Claude Opus 5.5 and 57 for GPT-6.2 Sol, and lists results across coding, science, and knowledge-work benchmarks. Google also says Neon costs less per task than Carbon in specified evaluations. VentureBeat’s launch coverage points out that headline token prices are not the whole cost comparison: Neon’s listed input and output rates are higher than those of Claude Sonnet 5.5 and GPT-6.2 Sol, while its reported advantage depends on benchmark performance and task-level cost. These are useful indicators, not a substitute for testing the same workload and tools.

For developers, the strongest case to investigate is frequent, tool-heavy work where latency, context, and repeated task cost all count. Google reports that Neon can work alone or coordinate Hydrogen subagents in Antigravity, and its card shows a strong one-million-token long-context result on GraphWalks. That doesn’t establish how it will perform on a particular repository or agent stack. Benchmark scores, provider evaluations, and the Carbon lineage provide a starting point; teams should compare completion quality, wall-clock time, and total token use on their own tasks. The limited synchronous output allowance, March cutoff, product-specific access, and pending full safety report are relevant constraints for production decisions.

Related models

Sources