
Gemini 4 Neon costs $3 and $15 per million tokens, and its rank depends on who counts
By Zain·
Google launched Gemini 4 Neon on November 18, 2026 and says it ranks third on the Artificial Analysis index. Artificial Analysis's own page lists it sixth of 96.
Google released Gemini 4 Neon on November 18, 2026, positioning it between the expensive, long-horizon Carbon and the fast, cheap Hydrogen. The pitch is "frontier-level intelligence with fast end-to-end task completion." The most useful fact about the launch is a small disagreement: Google's post says Neon ranks third on the Artificial Analysis Intelligence Index, while Artificial Analysis's own model page, read on November 19, lists it sixth of 96. The score is the same in both places. Only the rank differs.
What Google is selling
Neon has a 1 million token context window and costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15, a 95 percent discount. Google says that at its high thinking level, Neon's cost per task is 32 percent lower than Carbon's. Our earlier piece on Carbon opening to paying developers covers the price rise for that model after November 30.
Availability is broad. Developers get Neon on the paid tier of the Gemini API in Google AI Studio and Android Studio. Enterprises get it through Gemini for business and Google Antigravity. Consumers on Google AI Pro and Ultra can pick it in the Gemini app. A cyber variant is rolling out to trusted defenders in the Fairwind Program over the coming days.
The numbers, with the owner of each
Everything in this list is from Google's launch post unless stated. Treat it as vendor-reported.
- Artificial Analysis Intelligence Index: 61 at xhigh thinking, which Google calls third overall.
- Vals Index: 70.9 percent, second.
- Agent Arena Index: 1336, second.
- Zapier AutomationBench: 56.3 percent, first.
- Terminal-Bench 4.1: 65.2 percent. FrontierCode 2.0: 46.2 percent at high.
- Artificial Analysis Coding Agent Index: 58.6, which Google calls first.

What the independent page adds
Artificial Analysis lists four effort settings for Neon: 61 at xhigh, 60 at high, 55 at medium and 44 at low. That spread matters for buyers. The headline score needs the slowest, most expensive setting. At xhigh the page shows about 110 tokens a second, 56.51 seconds before the first answer token, and about 21.5 minutes to complete a full index run. It estimates a full run at about $6.44, and a blended price of about $2.21 per million tokens.
“frontier-level intelligence with fast end-to-end task completion”
So the independent view is that Neon is a very strong model with a steep price in time at the top setting. The page does not say Neon is the best at anything. Google's post claims firsts on several agent and coding indexes. Those firsts rest on indexes Google cites without publishing the full leaderboards in the post, and some are run by partners, such as Zapier and Vals. We could not verify them against the original leaderboards today.
Safety numbers
Google reports that Neon refused 1.98 percent of tasks on Terminal-Bench 4.0, cheated on 0.0 percent of those tasks, and took shortcuts on 2.45 percent of tasks in a separate coding evaluation, the lowest of any Gemini 4 model. It reports a 0.09 percent misaligned action rate on an agentic safety suite and 99.2 percent on an agent honesty evaluation. On Gray Swan's indirect prompt injection benchmark it reports a 0.0 percent attack success rate at the highest thinking level.
These are the lab's own evaluations. A 0.0 percent figure on a single benchmark says the benchmark was passed, not that the model cannot be attacked. Gray Swan and other red teams keep updating their attack sets, and a static score ages quickly.
Where Neon fits in the race
The release continues a pattern: each lab now ships a ladder of models, each priced against the one above and below. OpenAI's cut to GPT-6.2 Astra's price, covered in our report on the GPT-6.2 launch, put Astra at $8 and $40. Neon undercuts it by more than half on output price, though the two sit at different points on the quality curve.
How Neon compares with its siblings
Google's own ladder makes the pricing logic visible. Carbon, the top model, costs $2 input and $10 output as an introductory price and rises to $4 and $20 after November 30, 2026. Neon at $3 and $15 sits between those two Carbon prices, and Google says it costs 32 percent less per task than Carbon at high thinking. For the same launch week, Google also reported Carbon's per-attempt breach rate at 0.8 percent, down from 3.1 percent on Argon, and a 92 percent pass rate on a 375-task agentic honesty benchmark against 70 percent for Argon. Those are Carbon figures, but they set the safety baseline the Neon numbers are meant to match or beat.
For a team choosing between them, the sensible test is narrow. Take a fixed set of tasks, run Neon at high and at xhigh, run Carbon on the same set, and record accepted results and total cost. If Neon at high thinking clears the same tasks, the 60 versus 61 index gap does not justify the longer waits that Artificial Analysis measured at the top setting.
Our take
Neon looks like a real step for everyday engineering and analysis work, and its pricing is competitive. The honest summary is a score of 61, an independent rank of sixth, a vendor claim of third, and a set of agent leaderboard wins that need independent confirmation. If you buy on cost per finished task, run your own tests at high thinking before paying for xhigh.
Frequently asked questions
How much does Gemini 4 Neon cost?
Google lists $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. The context window is 1 million tokens. Google says cost per task at high thinking is 32 percent below Carbon's.
What is Gemini 4 Neon's Intelligence Index score?
Both Google and Artificial Analysis report 61 at xhigh thinking. Artificial Analysis also lists 60 at high, 55 at medium and 44 at low effort.
Is Gemini 4 Neon the third best model?
Google says it ranks third on the Artificial Analysis Intelligence Index. Artificial Analysis's own model page, read November 19, 2026, ranks it sixth of 96, so the rank depends on how entries are counted.
Who can use Gemini 4 Neon?
Developers on the paid Gemini API tier, enterprises through Gemini for business and Antigravity, and Google AI Pro and Ultra subscribers in the Gemini app. A cyber variant goes to Fairwind Program defenders.
How fast is Gemini 4 Neon?
Artificial Analysis lists about 110 tokens a second at xhigh thinking with 56.51 seconds to the first answer token. A full Intelligence Index run takes about 21.5 minutes at that setting.
Sources
What each one is, and whose it is.
- 1
Introducing Gemini 4 Neon, Google (November 18, 2026)
Vendor announcement - 2
Gemini 4 Neon model page, Artificial Analysis (November 18, 2026)
BenchmarkIndependent of the vendor - 3
Gemini 4 Carbon now available, Google (November 10, 2026)
Vendor announcement - 4
Everything that happened in AI today, Thursday, November 19, The Neuron (November 19, 2026)
Press reportIndependent of the vendor