Gemini 4 Neon scores 61 at $6.44 per measured task
Google’s Gemini 4 Neon scored 61 on Artificial Analysis’s Intelligence Index, ran at 110 tokens per second and cost $6.44 per task, but its higher-effort setting did not win every evaluation.
By Nathan Brooks
4 min read

What's News
- November 18, 2026: Neon (xhigh) scored 61, one point above Carbon, and ranked third behind GPT-6.2 Astra (max) and Claude Fable 5.2 (max).
- Neon cost $6.44 per Intelligence Index task and averaged 21 minutes, versus $7.76 and 36 minutes for Carbon; output ran at 110 tokens per second versus 57.
- Neon scored 59 on the Coding Agent Index and averaged $6.86 per task, versus 58 for both Opus and Fable at $12.61 and $15.47, respectively.
- On GDP.pdf, Neon (high) achieved a 33.4% all-pass score, beating Carbon (high) at 28.4% and Neon (xhigh) at 28.0%.
- Neon’s API costs $3 per million input tokens and $15 per million output tokens; cached input costs $0.15 after a 95% discount.
By Nathan Brooks
Gemini 4 Neon (xhigh) scored 61 on Artificial Analysis’s Intelligence Index and cost $6.44 per task, according to the evaluator on November 18, 2026. Neon averaged about 21 minutes per task, versus about 36 minutes for Gemini 4 Carbon (high), which cost $7.76 at standard, non-discounted pricing.
Artificial Analysis scored Carbon (high) at 60, one point below Neon. At standard, non-discounted pricing, Neon was $1.32 cheaper per task and completed tasks about 15 minutes sooner. Artificial Analysis placed Neon third overall, behind GPT-6.2 Astra (max) and Claude Fable 5.2 (max).
Artificial Analysis’s evaluation also found Neon used about 15% more output tokens per task than Carbon. Its cost advantage came from pricing, not lower token consumption. Neon produced output at about 110 tokens per second, against 57 for Carbon, which accounted for the shorter completion time.
Google’s release notes said Neon became generally available on November 18 for complex software engineering, knowledge work and science workflows. The model offers four thinking levels, a unified 1M-token context window and long decode support of up to 1M tokens.
Google described Neon as a proprietary reasoning model that accepts text, image, speech and video input and produces text. High is its default effort setting.
At high effort, Artificial Analysis recorded a score of 60 and 14.8 minutes per Intelligence Index task. It matched Carbon (high) on intelligence at less than half Carbon’s cost per task under standard pricing, according to the evaluator.
Coding and knowledge work
Using Antigravity CLI, Artificial Analysis measured a 59 on its Coding Agent Index for Neon (xhigh). Carbon (high) scored 55, while both Claude Fable 5.2 (xhigh) and Claude Opus 5.5 (max) scored 58.
Artificial Analysis put Neon’s average coding-agent cost at $6.86 per task, versus $12.61 for Opus and $15.47 for Fable. That made Neon about 46% cheaper than Opus and 56% cheaper than Fable on the evaluation’s measured task cost.
Artificial Analysis found a different result on GDP.pdf, a professional-document reasoning evaluation. Neon (high) posted a 33.4% all-pass score, compared with 28.4% for Carbon (high) and 28.0% for Neon (xhigh). The highest effort setting therefore trailed high by 5.4 percentage points.
The all-pass metric counts the share of 500 attempts per model on which every criterion passes. Artificial Analysis used 100 tasks across 10 professional domains for the evaluation.
On AA-Briefcase, Artificial Analysis measured an overall rating of 1,700 Elo for Neon (xhigh) on a comparative scale for agentic knowledge work. It also measured 1,771 Elo for Presentation Quality, 45 points above the previous leader, Claude Sonnet 5.5 (max), at 1,726.
Artificial Analysis’s separate Multimodal Knowledge Agent results put Neon’s Climb WSR at 56%, two points above Carbon and one point below Claude Fable 5.2 (max) at 57%.
Artificial Analysis defines Climb WSR as the mean weighted satisfaction rate on harder criteria in Google DeepMind’s evaluation of agent work on visually dense project files. It said MKA is a standalone evaluation rather than an Intelligence Index component.
Pricing and access
Artificial Analysis stated Google’s paid-tier API price for Neon at $3 per million input tokens and $15 per million output tokens. Cached input costs $0.15 after a 95% discount.
Carbon’s discounted rate is $2 per million input tokens and $10 per million output tokens through November 30, 2026. Neon sits 50% above that temporary rate, but 25% below Carbon’s standard $4/$20 input/output rate, Artificial Analysis said.
At launch, Google made Neon available through its API and the Gemini Enterprise Agent Platform. Pro and Ultra subscribers could use it in Antigravity and the Gemini app.
What the benchmark covers
Artificial Analysis cautioned that its Intelligence Index methodology may not apply directly to every use case. The index combines nine evaluations across knowledge work, coding, science and general knowledge. Each category carries a 25% weight.
The suite is primarily English-language and text-based, although image-capable models receive images in some evaluations. Artificial Analysis evaluates visual reasoning, speech input and multilingual performance separately.
What does Neon change for enterprise buyers?
Artificial Analysis’s results point to a measured operating gain, not a decisive intelligence lead. High produced the strongest GDP.pdf result, while xhigh led the Coding Agent Index and AA-Briefcase results; the better setting depends on the workload.
The practical test is the cost of completed work, not the token rate alone. Buyers must compare Neon and Carbon at the effort level their tasks require, under the pricing available when they buy. Carbon’s temporary discount ends November 30, so the immediate price comparison will change.
Neon’s clearest documented advantage is faster output and lower measured task costs at standard prices, not a decisive intelligence gain. Its enterprise case will rest on whether those savings hold on real workloads after Carbon’s November 30 discount ends.
Source: unite.ai
More from Nathan Brooks
Show full bio
News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.
2,460 articles