Skip to main content
TokenCost logoTokenCost
Model ReleaseNovember 19, 2026·7 min read

Gemini 4 Neon costs $3/$15, 50% more than Carbon through November 30. On December 1 it becomes the cheaper of the two, and Argon users get 10 points for less money.

Neon's price was set for December, not for this week. Whether it saves you money depends on which Gemini 4 you run now and how much you let it think.

Glowing blue neon tube curving in loops across a dark background

Photo by Max Kleinen on Unsplash

Google made gemini-4-neon generally available on November 18, paid tier only. It bills $3.00 per million input tokens, $0.15 cached and $15.00 output, thinking included, at one rate across the whole 1M window, per the Gemini API pricing page. The DeepMind model card says Neon is adapted from Gemini 4 Carbon.

Today that makes it the expensive sibling. Carbon and Argon still sit on $2 / $10 through November 30. On December 1 they jump to $4 / $20 and Neon, which has no dated change on the page, ends up 25% cheaper per token than either. Per token, Google is launching Neon at a price that only looks good 12 days from now. Per task, at the default setting, it already wins. If you're on Argon, that's good news: Neon scores 10 points higher on Artificial Analysis and will cost less.

The four Gemini 4 price cards side by side

Per 1M tokensNeonCarbon / Argon to Nov 30Carbon / Argon from Dec 1Hydrogen
Input$3.00$2.00$4.00$1.50
Cached input$0.15$0.10$0.20$0.075
Output (incl. thinking)$15.00$10.00$20.00$7.50
Batch / Flex$1.50 / $7.50$1.00 / $5.00$2.00 / $10.00$0.75 / $3.75
Priority$5.40 / $27.00$3.60 / $18.00$7.20 / $36.00$2.70 / $13.50
Free tierNoNoNoYes

Google pricing and changelog, read November 19, 2026. Batch, Flex and Priority show input / output. None of the four has a long-context tier.

Neon's Batch rate, $1.50 / $7.50, is exactly Hydrogen's Standard rate. That's a useful line to remember if your jobs can wait: the bigger model at the smaller model's price.

Caching works the way it does on the rest of Gemini 4. It's implicit only, with no explicit cache, no write fee and no storage fee. Google's Cloud guide puts the minimum at 4,096 tokens. You can't force a hit, so budget for some misses on prompts that change a lot.

What one agent turn costs

Take a 50,000-token prompt where 90% hits the cache, and 2,000 tokens out. Each bar is list price for that one call.

Gemini 4 Hydrogen
$0.0259
Gemini 4 Carbon (to Nov 30)
$0.0345
GPT-6.2 Sol
$0.0345
Gemini 4 Neon
$0.0517
Gemini 4 Carbon / Argon (Dec 1)
$0.0690
Claude Opus 5.5
$0.0690

Our arithmetic from the Google, OpenAI and Anthropic price pages. Opus 5.5 is $4 / $0.20 cached / $20.

Neon costs $0.0518 for that turn. Carbon costs $0.0345 now and $0.069 in December. Per token, then, Neon goes from 50% dearer to 25% cheaper overnight. GPT-6.2 Sol stays at $0.0345 on both dates, since OpenAI didn't touch its card.

Long documents flip the Sol comparison. Say you drop a 600,000-token contract bundle into one request, nothing cached, and ask for a 4,000-token summary. Sol passes the 272K line, so the whole request bills at $4 / $15: $2.46. Neon has no such step and charges $1.86. Carbon at the December card would be $2.48.

Per task, Neon is a heavy writer

Token prices only tell half of it. Artificial Analysis ran Neon at all four thinking levels, and at xhigh it wrote 283 million output tokens over the Intelligence Index, 4.3 times the 65.1 million median on its model page. That matters more than the $15 rate. Neon rows are highlighted.

Model and effortIndex$ / taskOutput tokens (run)
GPT-6.2 Astra xhigh61.6$2.5227.1M
Claude Fable 5.2 high61.8$9.0885.3M
Gemini 4 Neon xhigh61$6.44283M
Gemini 4 Neon high (default)60not published123M
Gemini 4 Carbon high60$7.76 ($3.88 today)205M
Claude Opus 5.5 max59.1$10.04218M
GPT-6.2 Sol max57.3$1.0160.9M
Gemini 4 Neon medium55not published62.9M
Gemini 4 Hydrogen xhigh54$3.16304M
Gemini 4 Argon high50$4.97 (~$9.94 Dec 1)113M

Artificial Analysis Intelligence Index v5.0.1 (Neon, Carbon). Neon and Carbon read November 19, the rest November 18. AA prices Carbon at $4 / $20, so the $3.88 figure is ours, halved for the current card. AA publishes a cost only for Neon xhigh. AA prices Argon at $2 / $10; the December figure is ours, doubled. Gemini and Claude token counts are AA estimates.

Against Carbon, Neon xhigh is a fair deal from December: one more point, $6.44 against $7.76, and 21.5 minutes a task against 37.1. That's the 17% saving in AA's write-up. Until then Carbon high costs about $3.88 a task, and Neon xhigh is 66% more.

Against OpenAI, it's not close. GPT-6.2 Astra at xhigh scores 61.6 for $2.52, 39% of Neon's bill, even though Astra's tokens cost $8 / $40. It wrote 27 million output tokens where Neon wrote 283 million. If you pick models by the Intelligence Index and the bill, Neon xhigh isn't the one to pick.

The default, high, is the better buy. It scores 60 on 123 million output tokens, less than half of xhigh. Google's launch post says Neon high costs 32% less per task than Carbon. AA goes further: its article says Neon high matches Carbon high at less than half the cost per task, with Carbon priced at $4 / $20. That puts Neon high under $3.88, which is below what Carbon high costs even on today's discounted card. AA didn't print the exact figure, so we haven't either.

Coding agents are the exception

AA also ran its Coding Agent Index with each lab's own harness. Neon at xhigh inside Antigravity scored 59 for $6.86 a task. Claude Code running Fable 5.2 at xhigh scored 58 for $15.47, and Opus 5.5 at max 58 for $12.61. Neon finishes that work for 44% of Fable's cost and 54% of Opus's. Carbon high scored 55, with no cost listed.

GPT-6.2 Astra and Sol weren't in that table, so we can't say whether the Astra gap above carries over to agent work. Two caveats. The Neon number comes from Google's own harness, and Antigravity runs on paid AI Pro and Ultra plans, not API billing. Through the API you pay token prices and bring your own loop.

Before you switch the model string

Most of the friction is in parameters. Thinking runs low, medium, high (the default) or xhigh, and per the Cloud guide a thinking_budget or MINIMAL gets you a 400, so anything still written for Gemini 3.x budgets breaks. Temperature, top_p and top_k are silently ignored. Penalties and a candidate_count above 1 are rejected outright.

Output stops at 65,536 tokens unless you use long decode, which stretches it to about 1M according to the model page. Neon reads text, images, video, audio and PDFs, writes text only, and has no Live API. Its knowledge cutoff is March 2026, same as Hydrogen. Search and Maps grounding draw on the shared Gemini 3+ allowance of 5,000 free a month, then $14 per 1,000.

One annoyance: there's no free API tier, so testing means paying from the first call. Hydrogen still has one. AI Pro and Ultra subscribers can try Neon in AI Studio, per Google's launch post.

Who should move, and when

Argon users should plan to move before December 1. Argon scores 50 and goes to $4 / $20 that day. Neon high scores 60 for $3 / $15. We can't think of a reason to stay on Argon past November 30 besides contracts or testing lead time.

Carbon users at the default setting don't need to wait. Per token Carbon is cheaper for 12 more days, but by AA's numbers Neon high reaches the same score for less per task already, and finishes in 14.8 minutes against 37.1. The exception is xhigh: there Neon writes more than Carbon does, so stay on Carbon until December 1 if you need that extra point.

If you're not on Gemini already, Neon isn't the obvious pick for general work. GPT-6.2 Sol is cheaper per token and per task, and Astra gets a higher score for less. Neon makes a stronger case for coding agents and very long prompts.

We covered the rest of the family in Gemini 4 Carbon, Gemini 4 Hydrogen and the November price changes, and OpenAI's side in GPT-6.2 Astra and Sol. The full card is on the Gemini 4 Neon page, or run your own numbers in the calculator.

Sources