Google Ranks Gemini 4 Neon Third on Artificial Analysis; the Site's Own Page Says Sixth
A.I. / news
Google Ranks Gemini 4 Neon Third on Artificial Analysis; the Site's Own Page Says Sixth
Google's Nov. 18 launch post sells Neon as fast, but Artificial Analysis clocks 21.5 minutes per task against a 12.1-minute median.

Google released Gemini 4 Neon on Nov. 18, pricing it at $3 per million input tokens and $15 per million output tokens, and described it as a model with "frontier-level intelligence with fast end-to-end task completion." Artificial Analysis, the independent benchmarking firm Google cites, lists the same model as one of the slowest and most verbose it has measured.
Koray Kavukcuoglu, senior vice president of Google DeepMind and Chief AI Architect at Google, wrote the launch post. The model has a 1 million token context window, four thinking levels and a default setting of high. Google did not disclose its size.
Where Neon is available
The model is in Google AI Studio and the paid Gemini API tier, in Android Studio, and in the Gemini app for Pro and Ultra subscribers. Google Antigravity gives it to AI Pro and Ultra subscribers. Cached input costs $0.15 per million tokens, which Google calls a 95 percent discount.
A separate variant, Gemini 4 Neon for cyber, goes only to trusted defenders in Google's Fairwind Program over the following days. Requests go through a Cloud account team.
Neon sits in the middle of the Gemini 4 line. Google says Carbon remains the pick for the hardest long-horizon work, and Hydrogen, which Google put behind Search on Nov. 12, handles fast everyday tasks.
Third or sixth
Google's post says Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh setting, which it describes as third overall, "just behind larger, slower models." Artificial Analysis's own model page gives the same score of 61 and ranks Neon sixth of 96 models. The page lists 60 at high, 55 at medium and 44 at low. The two sources do not explain the gap in rank.
The page also gives the cost of the index run at $6.44 and ranks Neon 79th of 95 on cost. It reports about 110 tokens per second of output and a time to first answer token of about 56.5 seconds.
- Gemini 4 Neon283 M tokens
- Median, comparable models65.1 M tokens
Source: Artificial Analysis model page for Gemini 4 Neon at xhigh effort, accessed 2026-11-19
Artificial Analysis says Neon needs 21.5 minutes per index task, against a median of 12.1 minutes. Its summary calls the model strong on intelligence but expensive for its price tier, slow per task and very verbose.
Google's cost claims use a different setting. The post says that at high, cost per task is 32 percent lower than Carbon and 50 to 80 percent lower than top competing models. At xhigh it says costs are at least 36 percent lower than competitors at their highest settings. Artificial Analysis's figures are for xhigh, so the two sets of numbers do not contradict each other directly, and neither source gives the high-setting cost on the other's terms.
Benchmarks Google supplied
Every figure below comes from Google's post and has not been reproduced independently. Google attributes most of them to the named third party that ran them.
| Benchmark | Neon result | Run by |
|---|---|---|
| Vals Index | 70.9%, second | Vals AI |
| Terminal-Bench 4.1 | 65.2% | Terminal-Bench |
| FrontierCode 2.0 (high) | 46.2% | Cognition |
| AutomationBench | 56.3%, first | Zapier |
| Vending-Bench 2 | $16,843, first | Andon Labs |
| SEC-bench Pro | 90.5% | SEC-bench |
Google also reports that Neon took shortcuts on 2.45 percent of coding tasks, the lowest rate in the Gemini 4 line, and a 0.09 percent misaligned action rate on its agentic safety suite. It describes its fabrication rate as "our lowest fabrication rate yet in internal Google Antigravity testing", which is an internal measure.

Google did not say how it reconciles the Intelligence Index rank in its post with the one on Artificial Analysis's page. The cyber variant begins reaching Fairwind Program members in the days after Nov. 18.
Sources
More in A.I.
- 01Tencent Hy4-flash Card Says Weights Are Out; Launch Coverage Says PlannedThe Hugging Face page lists 220B parameters and a custom licence, while launch coverage gives 224B and a future release, at $0.070 per million input tokens.
- 02Anthropic Says Claude Closed a Black Hole Bound Left Between 8.503 and 12The company's post names one model for the result; the paper's own record names two others, and no outside physicist has reviewed it publicly.
- 03Four Students Sue Omnilert and Baltimore County Over the Chip-Bag Gun Alert at Kenwood HighThe Nov. 17 suit targets a $2.6 million school contract, and the district's own timeline shows staff cancelled the alert 1 to 2 minutes after it fired.
- 04Gemini 4 Carbon Takes 36 Minutes a Task; Kernel Counts 7x the Steps of AstraArtificial Analysis puts Carbon at 36.0 minutes per Intelligence Index task against a 12.4-minute median, and a Kernel study traces most of the gap to extra steps.