Gemini 4 Neon Evaluation: Slight Improvement in Intelligence Index, Lower Cost per Task
Artificial Analysis measured Gemini 4 Neon (xhigh) at an intelligence index of 61, one point higher than Carbon (high), with a cost per task of $6.44 and an average completion time of about 21 minutes. Neon outputs approximately 110 tokens per second, but uses about 15% more output tokens per task, meaning its cost advantage primarily comes from pricing. It scored 59 on the Coding Agent Index, while performance at different reasoning effort levels did not consistently improve on evaluations such as GDP.pdf.
Why read this
SelectedAI score 82
The evaluation presents the intelligence index, task cost, and differences across reasoning effort levels, making it useful for comparing model efficiency on specific workloads.
Topics
Body
Artificial Analysis reported on November 18, 2026 that Google’s Gemini 4 Neon (xhigh) scored 61 on its Intelligence Index, one point above Gemini 4 Carbon (high). Neon cost $6.44 per Intelligence Index task and averaged about 21 minutes, versus $7.76 and about 36 minutes for Carbon at standard, non-discounted pricing.
The evaluation placed Neon third overall, behind GPT-6.2 Astra (max) and Claude Fable 5.2 (max). It also found Neon used about 15% more output tokens per task than Carbon, so the cost saving came from pricing rather than reduced token use. Output speed of about 110 tokens per second, against about 57 for Carbon, accounted for the shorter completion time.
Google’s release notes show that Gemini 4 Neon became generally available on November 18, 2026 for complex software engineering, knowledge work and science workflows, with four thinking levels, a unified 1M-token context window and long decode support up to 1M tokens.
At high effort, Artificial Analysis recorded a score of 60 and 14.8 minutes per Intelligence Index task for Neon. That setting matched Carbon (high) on intelligence at less than half Carbon’s cost per task under standard pricing. Neon is a proprietary reasoning model that accepts text, image, speech and video input, produces text and uses high as its default effort.
Coding and Knowledge Work Results
Using Antigravity CLI, Artificial Analysis measured a 59 on its Coding Agent Index for Neon (xhigh), against 55 for Carbon (high) and 58 for both Claude Fable 5.2 (xhigh) and Claude Opus 5.5 (max). Neon’s average cost was $6.86 per coding-agent task, compared with $12.61 for Opus and $15.47 for Fable.
On GDP.pdf, a professional-document reasoning evaluation, Artificial Analysis measured a 33.4% all-pass score for Neon (high), versus 28.4% for Carbon (high) and 28.0% for Neon (xhigh). The highest effort setting therefore did not produce the best result on this test. All-pass is the share of 500 attempts per model, across 100 tasks in 10 professional domains, where every criterion passes.
On AA-Briefcase, Artificial Analysis measured an overall rating of 1700 Elo for Neon (xhigh), on a comparative scale for agentic knowledge work, and 1771 Elo for Presentation Quality. The presentation result was 45 points above the previous leader, Claude Sonnet 5.5 (max) at 1726.
The separate Multimodal Knowledge Agent results put Neon’s Climb WSR at 56%, two points above Carbon and near Claude Fable 5.2 (max) at 57%. Climb WSR is the mean weighted satisfaction rate on harder criteria in Google DeepMind’s evaluation of agent work on visually dense project files. MKA is a standalone evaluation rather than an Intelligence Index component.
Pricing and Availability
Intelligence Index Scope
The Intelligence Index methodology, which Artificial Analysis says may not apply directly to every use case, combines nine evaluations across knowledge work, coding, science and general knowledge, with each category weighted at 25%. The suite is primarily English-language and text-based, although image-capable models receive images in some evaluations. Visual reasoning, speech input and multilingual performance are evaluated separately.
Neon’s results therefore describe a modest composite intelligence advance whose larger documented advantage is operational: faster output and lower measured task costs under standard pricing. Performance still varies by effort level and workload, while Carbon’s temporary discount changes the immediate token-price comparison. For enterprises assessing the model, the central questions are which workloads it will handle, which effort level they require and what completed tasks cost under current pricing.
AI Trends