Gemini 4 Neon: Third on the Artificial Analysis Index, Matching or Beating Carbon Almost Everywhere

Gemini 4 Neon review

Google released Gemini 4 Neon, the mid-tier model of the Gemini 4 family. It sits between the fast Hydrogen and the flagship Carbon, and is built for agentic coding, for working with spreadsheets, PDFs, audio and video, and for scientific tasks.

At its highest reasoning level, Neon scored 61 points and took third place on the Artificial Analysis Intelligence Index, behind only Claude Fable 5.2 and GPT-6.2 Astra. Neon is available through the Gemini API at $3 per million input tokens and $15 per million output tokens. Until the end of November, that is 1.5 times the price of Carbon. From December 1, when Carbon’s discount ends, Neon will be 25% cheaper on both input and output. On Zapier’s AutomationBench, which tests business process automation, Neon scored 56.3%, ahead of Gemini 4 Carbon at 53.9% and GPT-6.2 Astra at 52.4%.

What Gemini 4 Neon can do

Neon takes text, images, PDFs, audio and video as input and responds with text. The prompt, internal reasoning and response share a single budget of 1,048,576 tokens. In standard mode, the model can spend up to 65,536 tokens on reasoning and output in one pass. For longer outputs, a background long decode mode keeps generation going as long as room remains in the shared window. Details are in the model card and the API documentation.

In Google Antigravity, Neon can either run a project by itself or split the work up. For example, it drafts a plan, hands parts of the task to several Hydrogen agents and then reviews what they produce. The tools available to the model have a big effect on code quality. On SWE-sweep, the same Neon solved 12.8% of tasks in the basic mini-SWE harness and 39% in Antigravity.

For security teams, Google prepared a separate version, Gemini 4 Neon for cyber. According to the company, it scored 90.5% on SEC-bench Pro, a test of finding vulnerabilities in code. Google said members of the Fairwind Program will start getting access in the days following the release.

How Neon compares with other models

The Artificial Analysis Intelligence Index combines nine evaluations covering coding, document analysis, science and other tasks. Neon has four reasoning levels: low, medium, high and xhigh. At high, the default, Neon scored 60 points, the same as Carbon. Switching to xhigh added one more point. As noted by Kingy.ai, Neon’s row was not yet in the public leaderboard on release day, so at first the 61 was Google’s figure; Artificial Analysis later published its own results with the same score. The closest model above Neon is GPT-6.2 Astra, which OpenAI released shortly before Neon.

Model Intelligence Index, points Input, $ per 1M tokens Output, $ per 1M tokens
Claude Fable 5.2 65 8 40
GPT-6.2 Astra 62 8 40
Gemini 4 Neon 61 3 15
Gemini 4 Carbon 60 2 → 4 10 → 20
Claude Opus 5.5 59 4 20

Intelligence Index data from Artificial Analysis; Neon at xhigh, Carbon at high, all other models at their maximum level. Prices are standard rates; for Carbon, the current rate and the rate from December 1.

On Text Arena, where users compare model responses, Neon at high scored 1,518 points. That put it third, after Carbon (1,526) and Argon (1,521). Given the statistical uncertainty, Neon could rank anywhere from first to fifth.

Agent Arena rates agents based on real user sessions. In the first snapshot, Neon had 1,336 points across 4,234 sessions, just one point ahead of Opus 5.5. In the November 18 update, the two swapped places: Neon dropped to 1,333 and third place, while Opus had 1,336 and second. The median cost per task on this arena was about $2.60 for Neon and $1.34 for Carbon.

Where the highest reasoning level pays off

In CursorBench 4.1, agents get tasks from real coding sessions that require working across several files at once. At high, Neon scored 63.7% at an average cost of $6.54 per task. Carbon at the same level scored 61.5% at $8.52. Moving Neon to xhigh raised its score to 65.4%. That is a gain of 1.7 percentage points for a 63% higher cost, $10.67 per task.

Zapier’s results show a similar picture. In AutomationBench, agents carry out work tasks using business apps. Neon at xhigh scored 56.3%, compared with 53.9% for Carbon at high and 52.4% for GPT-6.2 Astra at max. A single Neon run at xhigh cost $1.69. At high, the score dropped to 54.3%, but the cost was almost halved, to $0.88. For comparison, an Astra run cost $0.83 and a Carbon run $2.12. All four Neon modes are covered in Zapier’s report.

In GraphWalks, the model receives a long list of connections and has to work out which nodes can be reached in a given number of steps. On tasks from 256K to 1M tokens, Neon scored 95.2% F1, a metric that penalizes both missed and extra answers. GPT-6.2 Sol scored 80.1% and Claude Opus 5.5 scored 66.8%. Google described the setup in its evaluation report.

On other tests, Neon trailed competitors. In the Google DeepMind table, Opus 5.5 is slightly ahead on terminal tasks (65.8% vs 65.2%) and on FrontierCode 2.0, which grades code edits (47.3% vs 46.2% for Neon at high). GPT-6.2 Sol read charts better on Chartography, 75.1% vs 73.5%. On WebDev Arena, where users pick the better website, Neon started with 1,764 points and seventh place; the leader, Fable 5.2, had 1,824. In the first Agent Arena snapshot, Neon also ranked 49th for reliability of Bash commands, the text commands used to work in a terminal.

What Neon costs to run

Google said a task from the Artificial Analysis suite costs 32% less on Neon at the default high level than on Carbon. For xhigh, the company claimed savings of 36% or more compared with competing models at their maximum settings. Artificial Analysis published its own calculation: Neon spent an average of $3.74 per task at high and $6.44 at xhigh. For Carbon at high, the authors used the December rate and arrived at $7.76, which means Neon at xhigh saves about 17%. If Carbon’s November discount is applied to the same token usage, a task costs about $3.88 – almost the same as Neon at high and noticeably less than Neon at xhigh.

Neon has no free tier in the Gemini API. Tokens the model spends on reasoning are billed at the output rate. When the API recognizes a repeated prompt prefix and serves it from cache, that input costs $0.15 per million tokens, 95% less than the standard rate. Caching for Neon is automatic. In Batch and Flex modes, a million tokens costs $1.50 for input and $7.50 for output; in Priority mode, $5.40 and $27 respectively.

What one session costs (editor’s estimate). Suppose that over several requests an agent processed 2 million input tokens, 80% of them from cache, and generated 200,000 output tokens, including reasoning. On Neon, that comes to $4.44: $1.20 for uncached input, $0.24 for cached input and $3 for output. With the same usage and the same cache share, Hydrogen would cost $2.22, and Carbon $2.96 until the end of November and $5.92 from December 1. In real tasks, models may use different numbers of tokens.

On OpenRouter, the effective input price on the route that carries most of Neon’s traffic was about $0.31 per million tokens when we checked on November 19, 2026, at 10:37 UTC. This is the average across the platform’s user requests, including caching and discounts. Cache hit rates differed even between the two routes of Neon’s providers on the platform, Google AI Studio and Google Vertex: about 95% on one and 85% on the other.

Developers can use Neon in Google AI Studio and Android Studio through the paid tier of the Gemini API. The model ID is gemini-4-neon. Google AI Pro and Ultra subscribers can select it in the Gemini app and in Google Antigravity. For companies, the model is available in Gemini for business.

What the first tests showed

In Google’s demos, Neon built a browser-based ski simulator using real terrain data for Mont Blanc. The model wrote 10,200 lines of rendering code and added avalanches; the simulator updates the skier’s motion 120 times per second. In another example, it used archival photos to recreate Alan Turing’s 1936 study in Cambridge, complete with a working virtual Turing machine on the desk. In a third, several agents turned a single image into a Google Sheets spreadsheet, where more than 500,000 formulas detected object edges, reconstructed scene depth and relit the image.

Blogger Chubby showed 17 interactive demos built with Neon in Antigravity in a single day. In one of them, Neon added three building types and three tools: in the game scene, a cannon knocked whole chunks out of a wall, while a chisel chipped off a few blocks at a time. A Solar System model contained a factual mistake – Neon gave an outdated number of moons and corrected it after the author pointed it out.

Greg Isenberg described the place Neon has taken in their work: it cuts their spending roughly in half, though the hardest tasks still often go to Claude.

According to data from Glean, Neon handled tasks at least as well as Carbon, but three times faster and 40% cheaper. In a Priors measurement at xhigh, the first token arrived after 57 seconds, after which Neon generated about 110 tokens per second – almost twice as fast as Carbon.

When Carbon still makes sense

Neon is at least as good as Carbon almost everywhere and noticeably faster, so it makes sense to start new projects with it. Where Carbon is formally ahead, the gap mostly falls within the margin of error: 8 points on Text Arena, 5 on WebDev Arena (1,769 vs 1,764) and one point on GraphWalks with long documents (96.2% vs 95.2%). Real reasons to choose Carbon remain in three cases.

The first is long autonomous tasks. In Agents’ Last Exam, agents spend hours on professional assignments with verifiable results, and Carbon scores 77.4% there, first place on the leaderboard. Neon has 75.8% in Google’s table. Google itself positions Carbon as the model for the hardest multi-step projects: in internal tests, it worked for days without abandoning a task when it hit a dead end.

The second is factual questions without search. AA-Omniscience deducts points for wrong answers but not for an honest “I don’t know”. Artificial Analysis found that Carbon at high makes up an answer in 11.4% of cases when it does not know the correct one, while Neon does so in 13.9%, with nearly identical accuracy of about 58–59%. Where a wrong answer costs more than an honest “no data”, that is an argument for Carbon.

The third is large batch runs before the end of November. While Carbon’s discount lasts, Batch and Flex modes cost $1 per million input tokens and $5 per million output tokens, versus $1.50 and $7.50 for Neon. At high, Carbon takes more than twice as long as Neon to complete a task from the Artificial Analysis suite, but for overnight runs speed matters less. From December 1, this argument disappears, and Neon becomes cheaper on every tier.


bnr2mob
Subscribe
Notify of
guest

0 Comments
Oldest
Newest Most Voted