News

Artificial Analysis puts Gemini 4 Neon one point above Carbon at lower standard-price cost

Artificial Analysis put Gemini 4 Neon one point above Gemini 4 Carbon on its Intelligence Index, with lower estimated cost and faster completion in a comparison that used different effort levels and Carbon's standard price.

D
Nov 20, 2026 · 2 min read

Artificial Analysis scored Google’s Gemini 4 Neon at 61 on its Intelligence Index v5.0.1, one point above Gemini 4 Carbon. It also estimated that Neon completed the benchmark’s average task faster and at a lower cost under the comparison’s stated assumptions. The November 18 measurement report placed Neon third at launch, behind GPT-6.2 Astra and Claude Fable 5.2. By November 20, the live leaderboard had moved Neon to sixth without changing its score.

The result was not a like-for-like effort comparison: Neon ran at xhigh effort and Carbon at high. Artificial Analysis estimated a $6.44 cost per Intelligence Index task for Neon, compared with $7.76 for Carbon using Carbon’s standard, non-discounted rates—a difference of about 17%. Carbon was temporarily available at lower promotional prices through November 30, so the reported saving did not reflect the price customers could obtain for Carbon on publication day.

Lower cost did not mean shorter answers. Artificial Analysis said Neon used about 15% more output tokens, but its listed API prices of $3 per million input tokens and $15 per million output tokens were 25% below Carbon’s standard $4 and $20 rates. Google’s launch post separately listed Neon’s prices, including $0.15 per million cached input tokens.

Artificial Analysis estimated about 21 minutes per task for Neon and about 36 minutes for Carbon, while measuring output speeds of roughly 110 and 57 tokens per second, respectively. Its benchmarking methodology says time per task is derived from weighted output-token use and representative endpoint speed, not wall-clock timing of the benchmark runs. The speed figure averages measured workloads over the preceding five days.

On the publisher’s Coding Agent Index, Neon xhigh scored 59 using Antigravity CLI, compared with 55 for Carbon high. Neon also edged Claude Fable 5.2 xhigh and Claude Opus 5.5 max, which each scored 58 using Claude Code. Average task cost was $6.86 for Neon, $15.47 for Fable and $12.61 for Opus. Because the Gemini and Anthropic models ran in different coding harnesses, the published results do not isolate model performance from harness effects.

Other measurements also favored Neon over Carbon: 33.4% versus 28.4% on GDP.pdf. Neon xhigh reached 1,700 Elo overall and 1,771 Elo for presentation quality on AA-Briefcase. On Google’s MKA evaluation, Artificial Analysis listed Neon at a 56% Climb win-success rate, compared with 54% for Carbon and 57% for Claude Fable 5.2 max.

Google’s launch material separately promoted Neon as third on the Artificial Analysis index, second on the Vals Index at 70.9%, and first on the Coding Agent Index at 58.6. Those are benchmark claims selected by Google for its announcement, distinct from Artificial Analysis’s measurement article. The Gemini 4 Neon model card says the model is adapted from Carbon, accepts text, image, audio and video inputs, and supports a combined input-output context window of up to one million tokens with text output. Google began rolling Gemini 4 Neon out through the Gemini API, AI Studio, business products, Antigravity and the Gemini app.

More news