sysnote
Analysis

Gemini 4 Neon matches Carbon in less than half the time; its model card skips the two models that score higher

Google released Gemini 4 Neon on November 18, in a post signed by Koray Kavukcuoglu, Google DeepMind's SVP and chief AI architect. In the API it is gemini-4-neon, paid tier only, at $3 per million input tokens and $15 per million output tokens, with the same 1M-token shared window as the rest of the Gemini 4 family. Its model card says it is "adapted from Gemini 4 Carbon," and Google places it between Carbon and Hydrogen: Carbon for the hardest work, Hydrogen for everyday tasks, Neon for "complex, high-frequency workflows."

The word Google keeps repeating is "fast," and against Carbon the independent numbers agree. What Neon costs relative to Carbon depends on whether you read the price list before or after November 30, and the comparison table in the model card leaves out the two models that score above Neon.

Neon against Carbon

Artificial Analysis ran Neon through its Intelligence Index at all four thinking levels. These are its figures as of November 19, at list prices. For Carbon the site uses $4/$20, the price that applies from December 1.

Model (setting) Intelligence Index Cost per task Minutes per task Output tokens per task
Gemini 4 Neon (low) 44.1 $0.75 3.7 21.7K
Gemini 4 Neon (medium) 54.9 $1.90 8.3 50.0K
Gemini 4 Neon (high, default) 59.6 $3.74 14.8 90.2K
Gemini 4 Neon (xhigh) 61.4 $6.44 21.5 141.6K
Gemini 4 Carbon (medium) 58.3 $6.59 29.6 103.0K
Gemini 4 Carbon (high) 60.0 $7.76 37.1 122.8K

At their high settings the two models score the same within half a point, and Neon finishes a task in 14.8 minutes against Carbon's 37.1. Part of that is raw speed: Artificial Analysis measures Neon at about 101 output tokens per second and Carbon at 55. The rest is that Neon writes fewer tokens. One caveat on the time column: the site's speed figures are rolling averages, and Carbon's own time per task was 25.6 minutes when we covered its general release on November 10. Some of the gap is Carbon's endpoint getting slower since then.

At December prices Neon (high) costs 52% less per task than Carbon (high). Two things produce that. Neon's tokens are cheaper, and it uses fewer of them: on the full index run it read 4.39 billion input tokens and wrote 123 million, where Carbon read 7.62 billion and wrote 205 million. If I reprice Neon's run at Carbon's $4/$20 rates (Artificial Analysis publishes the per-token breakdown, so this is just a multiplication), it comes to $4.99, still 36% below Carbon. So most of the saving comes from token use, not from the price list.

Until November 30, though, Carbon is on introductory pricing at $2/$10. At that rate Carbon (high) costs about $3.88 per task, and Neon (high) at $3.74 is only about 4% cheaper. Per token, Neon is 50% more expensive than Carbon today. Google's Antigravity post says Neon at high shows "a 32% lower cost per task than Carbon." It doesn't say on which benchmark, at which Carbon setting, or at which Carbon price, and neither 52% nor 4% matches it.

For someone running Carbon at high now, Neon gives the same Intelligence Index score in less than half the time, at roughly the same cost per task until the end of the month and about half the cost after.

Where the 61 comes from

Google's announcement leads with "#3 overall with a score of 61 on the Artificial Analysis Intelligence Index," "landing just behind larger, slower models." The 61 is Neon at xhigh, a level Carbon doesn't offer. It costs $6.44 per task and takes 21.5 minutes, because at that setting Neon writes about 142,000 output tokens per task. Artificial Analysis' own page calls it "notably slow per task and very verbose" compared with models in its price range.

Counted by model, #3 is right: Claude Fable 5.2 scores 64.6 at max and GPT-6.2 Astra 62.1 at max. Counted by setting, Neon (xhigh) is sixth, behind three Fable 5.2 settings and two Astra settings.

"Slower" fits Fable 5.2, which takes 41.8 minutes per task at max. It fits Astra only if speed means tokens per second. Astra generates about 49 per second against Neon's 110, but it writes far fewer of them: 41,000 per task at max, where it needs 13.9 minutes, and 22,000 at xhigh, where it scores 61.6 in 7.5 minutes for $2.52. Per finished task, Astra (xhigh) matches Neon's headline score in about a third of the time. "Larger" can't be checked: Google gives no parameter count for Neon or for the models it calls larger.

The comparison table

The model card's results table puts Neon next to Claude Opus 5.5 and GPT-6.2 Sol. Neither is its maker's top model. Fable 5.2 and GPT-6.2 Astra, the two that score above Neon on the index Google quotes in the same table, aren't in it, and neither is Carbon.

Against those two, Neon leads most rows. Opus 5.5 is ahead on Terminal-bench 4.1 (65.8% against 65.2%) and FrontierCode 2.0 (47.3% against 46.2%), and Sol on Chartography (75.1% against 73.5%). The widest gap in the table is SWE-sweep: 39.0% for Neon against 10.1% for Opus 5.5 and 8.3% for Sol, in the "provider harness" setting.

The evaluation methodology page explains that row. Each model ran in its own vendor's agent: Antigravity-CLI 1.5.9 for Neon, Claude Code for Opus 5.5, Codex for Sol. Google then adds that "the public Antigravity-CLI harness includes a default tool that allows for sub-agents that fan out across the codebase, which is particularly suited to bug finding tasks like SWE-sweep." The table also has a second SWE-sweep row, all three models in the same mini-swe-agent harness. There the scores are 12.8% for Neon, 3.6% for Opus 5.5 and 11.6% for Sol. In the same harness, Neon's lead over Sol is 1.2 points, not 30.7. The 39.0% is a result for Neon inside Antigravity, and Google says so in the methodology, just not in the table.

Two more methodology details matter when reading the rows. Neon's Terminal-Bench 4.1 score comes from Artificial Analysis using Antigravity-CLI, while the competitors' come from the official leaderboard. And "all the results for non-Gemini models are sourced from providers' self-reported numbers unless otherwise mentioned," so many rows compare Google's runs with other labs' runs.

No new safety assessment

The model card's safety section follows the route Carbon took. Google ran a Material Capability Change assessment against Gemini 4 Argon, the last model that went through a full critical capability assessment. Neon "showed an increase in general capability... however this falls below our MCC threshold," so Argon's determinations carry forward: Google still can't rule out the biological uplift and cyber uplift critical capability levels, or the stealth and situational awareness tracked level, and it keeps the mitigations that come with them.

The comparison is with Argon, not with Carbon, which is the stricter choice: each derived model is measured against the last fully assessed one, so a series of small steps can't add up unnoticed. What the card doesn't say is how far above Argon Neon lands on Google's own measure. On Artificial Analysis, Argon (high) scores 50.4 and Neon 59.6 at high and 61.4 at xhigh. Google's MCC test uses its internal General Capability Index, where the bar for Carbon was about four months of progress, so the external gap doesn't translate directly. The Neon Safety & Security Report, which should contain the index values, "will be available in the coming days."

Before switching

From the What's new in Gemini 4 Neon guide: Neon defaults to high and supports xhigh, while Carbon and Argon default to medium and stop at high, so a straight model-ID swap from Carbon changes the thinking level too. minimal returns an error. temperature, top_p and top_k are ignored when in range and rejected with a 400 when out of range. Explicit context caching isn't supported; implicit caching applies to prompts of at least 4,096 tokens, with cached input at $0.15 per million. Google recommends HTTP timeouts of at least 15 minutes. Artificial Analysis measured 40.6 seconds to the first answer token at high and 56.5 seconds at xhigh.

Google's guide advises running evaluations at all four thinking levels and picking the lowest one that meets your bar. Artificial Analysis' numbers show why: Neon at medium costs half as much per task as at high and takes a little over half the time, for 4.7 fewer index points.

Comments

No comments yet. Be the first.

Write a comment

Comments are published after review.