DEV Community

IA Compa
IA Compa

Posted on Originally published at iacompa.com

Google launches Gemini 4 Neon, its Gemini 4 model for coding and knowledge work, at $3/$15 per million tokens

Google has launched Gemini 4 Neon, positioning it as the Gemini 4 model for coding and knowledge work — and, in effect, the replacement for Gemini 3 Pro in Google's lineup.

The company says Neon is its best model for agentic coding and knowledge work, spanning coding, tool use, multimodal reasoning and web search. It is available now through Gemini in AI Studio, Vertex AI, Gemini CLI, the Gemini app and Gemini API. Gemini 4 Neon is also being released to developers and enterprise customers via the Gemini API and Vertex AI.

What changed

  • Gemini 4 Neon, also written as Gemini 4.0 Neon, replaces Gemini 3 Pro as Google's flagship model for these workloads.
  • Google describes the model as built for agentic coding and knowledge work.
  • Pricing is $3 per million input tokens and $15 per million output tokens.
  • Pricing for low, medium and high reasoning effort is unchanged.
  • The model supports a 1M-token context window.

Metrics

Benchmark Gemini 4 Neon
GDPval 60.9%
SWE-bench Pro 49.1%
Terminal-Bench 2.0 62.4%
Terminal-Bench 2.0 (with effort: high) 66.3%
BrowseComp 72.3%
GPQA Diamond 95.8%

Benchmarks were reported from public sources.

Reasoning effort

Google reports that Gemini 4 Neon includes a reasoning-effort setting that trades tokens and speed for higher accuracy:

Reasoning effort GDPval score Median tokens
Low 46.8% 1,450
Medium 55.9% 5,850
High 60.9% 13,700

Benchmarks & resources

Why it matters

Neon matters because it is a like-for-like price upgrade over Gemini 3 Pro, with no pricing change for low, medium and high reasoning effort. In other words, Google is positioning Neon as a direct replacement: the same price, with higher intelligence, broader capabilities and a reasoning-effort setting that lets teams trade cost and speed for accuracy.

The benchmarks point the same way. Its 49.1% on SWE-bench Pro, 62.4% on Terminal-Bench 2.0 (66.3% with high reasoning effort) and 72.3% on BrowseComp suggest a model built for the day-to-day work of agents: writing and fixing code, completing terminal tasks and researching the web. Its 95.8% on GPQA Diamond and 60.9% on GDPval also place it among the strongest models for expert-level knowledge work.

Sources

Originally published at IA Compa.

Top comments (0)