Google DeepMind has released Gemini 4 Neon, marking its fourth model introduction in seven weeks. The aggressive release cadence places the model directly between the company's existing Carbon and Hydrogen offerings. Neon provides a one-million-token context window for text, images, audio, and video inputs, while scoring 61 points on the Intelligence Index at its highest reasoning setting, "xhigh." That score edges past Gemini 4 Carbon's 60 points, leaving Neon in third place overall on the Artificial Analysis benchmark behind OpenAI's GPT-6.2 Astra and Anthropic's Claude Fable 5.2.
The model's pricing structure shows that efficiency gains stem directly from lower list prices rather than reduced token usage per query.
Google DeepMind is altering the economics of inference by lowering costs on complex processing tasks while applying immediate pressure on rival providers. Through Google AI Studio and Gemini for Business, one million input tokens for Neon cost $3.00, or $0.15 when cached, while one million output tokens cost $15.00. According to Artificial Analysis, per-task costs for Neon at the xhigh setting total $6.44, representing a 17 percent reduction from Carbon's undiscounted cost of $7.76.
Inference Economics and Benchmarks
The fundamental shift in task pricing is driven by raw rate cuts rather than optimized token usage. Neon uses approximately 15 percent more output tokens per task than Carbon, meaning financial savings depend entirely on the reduced list price. At its second-highest reasoning setting, "high," Neon matches Carbon's intelligence score while cutting per-task operational costs by half. Benchmark processing times also drop from roughly 36 minutes to 21 minutes, as Neon generates approximately 110 tokens per second compared to Carbon's 57 tokens per second.
Google's iterative release sequence follows previous infrastructure expansions focused on multimodal outputs. Neon extends this visual capability focus, hitting a top score of 33.4 percent on the GDP.pdf document test at the high reasoning level, outperforming Carbon by five points. On the AA-Briefcase test, Neon takes the top spot for presentation quality with 1771 Elo points, placing it ahead of Anthropic's Claude Opus 5.5.
Enterprise Coding and Security Performance
In coding environments, Neon matches or exceeds rival benchmarks while substantially undercutting per-task operational costs.
Competitors may point out that top-tier benchmarks like the Artificial Analysis index still place rival models from OpenAI and Anthropic ahead of Google's latest option. However, in specific developer workloads, Neon alters the spending equation. Inside the Google Antigravity execution environment, Neon at the xhigh setting achieves a score of 59 points on the Coding Agent Index. That score surpasses Carbon's 55 points as well as the 58 points recorded by both Claude Fable 5.2 and Claude Opus 5.5. An average coding task costs $6.86 with Neon, compared to $12.61 for Opus 5.5 and $15.47 for Fable 5.2. Google reports that the Antigravity setup also triples Neon's hit rate on the SWE-sweep test to 39.0 percent.
In broader performance testing, Google reports that Neon leads individual categories on the Vals Index, reaches 56.3 percent on Zapier's AutomationBench, and scores 84.0 percent on science evaluations like RiemannBench. For enterprise security applications, Google is deploying a cyber variant via the Fairwind Program that reaches 90.5 percent on SEC-bench Pro and 51.26 percent on WebExploitBench v1.2 at half the price of competing models. Additionally, Neon's model card indicates that the system blocks 100 percent of indirect prompt-injection attacks on the Gray-Swan benchmark, bringing processing execution times down.
What it signals
The model actually consumes about 15% more output tokens per operation than Carbon, yet offsets that appetite strictly through slashed API rates in Google AI Studio.