Gemini 4 Neon Is 25% Cheaper Than Carbon, but Carbon Is Half Price Until 30 November
Google's Gemini 4 Neon cuts token prices 25% against Carbon's list rate, but Carbon is half price until 30 November and Neon has no free tier.

Google released Gemini 4 Neon on 18 November, a model it says brings frontier reasoning to fast, high volume work at $3 per million input tokens and $15 per million output tokens. It sits between Gemini 4 Carbon, the company’s largest reasoning model, and Gemini 4 Hydrogen, its cheaper everyday model.
The headline number is a 25% cut against Carbon’s standard price of $4 and $20. The catch is the calendar. Carbon is currently selling at $2 and $10, a 50% discount that Google’s pricing page says ends on 30 November. Until then, Neon costs 50% more per token than the model it is supposed to undercut.
Google also says Neon ranks third overall on the Artificial Analysis Intelligence Index with a score of 61, one point above Carbon, and first on the Artificial Analysis Coding Agent Index. Independent testing confirms the 61 but finds a narrower gain than the marketing implies, and no evidence for the three times faster claim that circulated with the launch.
Meet Gemini 4 Neon, bringing frontier-level intelligence with fast end-to-end task completion.
Built for complex coding, knowledge work, and science workflows, it ranks #3 overall on the @ArtificialAnlys Intelligence Index at xhigh (ahead of comparable models) and #2 on the Vals… pic.twitter.com/5pTgQeTFz5
— Google (@Google) November 18, 2026
Quick answer: Neon is a real model with a real price cut, but the cut applies to a list price that few teams are paying this month. Anyone on Carbon’s promotional rate pays more by moving before 1 December. After that date Neon is 25% cheaper. The decision is a calendar problem first and a capability problem second.
What Google actually shipped
Neon accepts text, images, video, audio and PDFs and returns text. It has a 1,048,576 token context window and a default maximum output of 65,536 tokens, which can stretch to the full million with long decode. It supports low, medium, high and xhigh thinking levels.
The API price is $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens. Batch jobs cost $1.50 and $7.50. There is no free tier, and explicit context caching is not supported, only implicit caching.
Google is selling Neon as the middle child: deeper than Hydrogen, faster and cheaper than Carbon at list price, built for coding agents, spreadsheet work, legal and tax analysis, and long tool heavy sessions. It is available in the paid Gemini API and AI Studio, in Gemini for business, in Antigravity for paid Google AI Pro and Ultra subscribers, and in the Gemini app.
On Google’s own benchmark sheet, Neon scores 70.9% on the Vals Index for second place, takes first on the Legal Research Bench and Tax Agent Bench, reaches 65.2% on Terminal-Bench 4.1 and 58.6 on the Artificial Analysis Coding Agent Index. Google also reports a 0.09% misaligned action rate on its agentic safety suite and a 99.2% pass rate on its agent honesty evaluation.
The price cut that arrives on 1 December
The comparison Google wants readers to make is $3 against $4 on input and $15 against $20 on output. That is a genuine 25% reduction, and it is the number in most launch coverage. It is also the number that applies after Carbon’s promotional pricing expires.
Carbon costs $2 and $10 through 30 November, then returns to $4 and $20. So the same Neon launch is either a 50% price increase or a 25% price decrease depending on the deadline. Google has published no commitment that Neon’s price will hold after the promotion ends.
The gap shows up in per task economics. Artificial Analysis measured Neon at $6.44 per task and Carbon at $7.76 using list price, with Neon consuming about 15% more output tokens. Apply Carbon’s current 50% discount and its cost per task falls to roughly $3.88, a calculation from published figures rather than a measured result. In that window Neon costs about 66% more per completed task.
Batch pricing keeps the same order. A workload that burns 10 million input tokens and 2 million output tokens costs $60 on Neon at standard rates, $40 on Carbon at the promotional rate and $80 on Carbon at list. On 1 December those totals become $60 for Neon and $80 for Carbon.
What independent tests show
Artificial Analysis placed Neon at xhigh on 61 points, third overall behind GPT-6.2 Astra and Claude Fable 5.2, both at max effort. The one point gain over Carbon at high effort comes with small improvements in agentic knowledge work, visual reasoning and coding, not a step change in raw intelligence.
Google continues to accelerate its release cadence with Gemini 4 Neon, delivering a small Intelligence gain over Gemini 4 Carbon at a ~17% lower Cost per Task@GoogleDeepMind has released Gemini 4 Neon, its fourth model in just seven weeks. Neon (xhigh) scores 61 on the… pic.twitter.com/L76oKwyxXa
— Artificial Analysis (@ArtificialAnlys) November 18, 2026
On the Coding Agent Index, Neon scores 59 in the Antigravity command line harness, three points above Carbon and level with Claude Code on Claude Fable 5.2 and Opus 5.5. The difference is cost: $6.86 per task for Neon against $12.61 for Opus 5.5 and $15.47 for Fable 5.2.
Visual work improved more visibly. Neon scored 33.4% on GDP.pdf at high effort, five points above Carbon, and reached 1771 Elo on presentation quality in AA-Briefcase. Those gains matter for consulting decks, charts and dense PDF review, which is where Indian service firms do a lot of billable work.
The speed claim that got rounded up
At least one launch headline, indexed by Techmeme, described Neon as three times faster than Carbon. The independent measurement does not support that. Artificial Analysis recorded about 110 output tokens per second for Neon against about 57 for Carbon, close to 1.9 times, and about 21 minutes per task against 36.
Faster is still the right word. Nearly double the output speed matters in interactive coding, where a stalled generation costs attention. It is not a threefold jump, and teams sizing infrastructure around the larger number will overestimate the throughput they gain.
Specs and prices side by side
| Item | Gemini 4 Neon | Gemini 4 Carbon | Gemini 4 Hydrogen |
|---|---|---|---|
| Input price per 1M tokens | $3.00 | $2.00 until 30 Nov, then $4.00 | $1.50 |
| Output price per 1M tokens | $15.00 | $10.00 until 30 Nov, then $20.00 | $7.50 |
| Cached input per 1M tokens | $0.15, implicit only | $0.10 until 30 Nov, then $0.20 | $0.075, implicit only |
| Batch input and output | $1.50 and $7.50 | $1.00 and $5.00 until 30 Nov, then $2.00 and $10.00 | $0.75 and $3.75 |
| Context window | 1,048,576 tokens | 1M tokens | 1M tokens |
| Default maximum output | 65,536 tokens | 65,536 tokens | 65,536 tokens |
| Thinking levels | Low, medium, high, xhigh | Low, medium, high, xhigh | Low, medium, high, xhigh |
| Free tier | No | No | Yes |
The table makes the positioning plain. Neon is not the cheap model in the family. Hydrogen costs less on every token and still offers a free tier. Neon is the mid priced model meant to pull agentic coding away from larger competitors, not from Google’s own budget option.
Where the model is genuinely good
The strongest evidence is not the intelligence score. It is the combination of near frontier coding agent performance and a cost per task below the Claude models it matches. For teams running many parallel agents, the gap between $6.86 and $15.47 per task is the difference between a pilot and a production budget.
Google’s agentic examples are unusually specific: a 10,200 line WebGL2 ski simulator, a 100,000 cell spreadsheet built from one image, and a 3D recreation of Alan Turing’s 1936 study with a working Universal Turing Machine. These are demos, not independent audits, but they show the target workload: long, tool heavy tasks that would previously have gone to a larger and slower model.
The cybersecurity claims need a caveat. The numbers for AutoBaxBench, WebExploitBench and SEC-bench Pro describe Gemini 4 Neon for cyber, a separate variant rolling out to trusted defenders through the Fairwind Program. The general API model is not the cyber model.
Confirmed versus not confirmed
| Claim | Status |
|---|---|
| Gemini 4 Neon launched on 18 November 2026 | Confirmed by Google’s blog and model docs |
| Standard price is $3 input and $15 output per 1M tokens | Confirmed on Google’s pricing page |
| Carbon costs $2 and $10 until 30 November, then $4 and $20 | Confirmed on Google’s pricing page |
| Neon has a 1M token context and no free tier | Confirmed in Google’s model documentation |
| Neon scores 61 on the Artificial Analysis Intelligence Index | Confirmed independently by Artificial Analysis |
| Neon is three times faster than Carbon | Not confirmed, independent measurement shows about 1.9 times output speed |
| Neon’s price will stay at $3 and $15 after 30 November | Not confirmed, Google has published no commitment |
| Google’s first place coding agent claim holds on every leaderboard | Not confirmed, Artificial Analysis scores it 59 and level with top Claude models |
What it means for Indian developers
Indian teams bill in rupees and pay Google in dollars, so every basis point of token price lands on margin. The practical move is not to migrate on launch day. A team already on Carbon’s promotional rate should keep it until 30 November and revisit on 1 December, when Neon becomes cheaper for the same work.
Teams starting fresh have a different calculation. Neon beats Carbon at list price, but Hydrogen is cheaper for testing whether a workload works at all and has a free tier. The sensible sequence is Hydrogen for prototypes, Neon for production agents, and Carbon for the hardest jobs where the extra point changes the outcome.
Caching deserves a closer look. The $0.15 cached input rate is a 95% discount, which matters for agents that resend a long system prompt on every turn. But Neon supports implicit caching only, so a team cannot pin a cache with an explicit TTL. The discount depends on Google’s automatic cache hits, not on the application’s design.
Procurement should change too. Indian IT services firms negotiating multi quarter AI contracts should tie token pricing to list rates and add review dates, because Google is shipping a new Gemini 4 model roughly every one to four weeks. Neon is the fourth in seven weeks, after Argon on 30 September, Carbon, and Hydrogen on 12 November. Each launch resets benchmark comparisons, and the naming gives no clue about size, speed or price without the docs.
Frequently asked questions
How much does Gemini 4 Neon cost?
It costs $3 per million input tokens and $15 per million output tokens. Cached input is $0.15 per million tokens, and batch jobs cost $1.50 and $7.50. There is no free tier.
Is Neon cheaper than Gemini 4 Carbon?
Against Carbon’s standard rate of $4 and $20, yes, by 25%. Until 30 November, Carbon is discounted to $2 and $10, so Neon actually costs 50% more per token during that window.
How fast is Gemini 4 Neon?
Artificial Analysis measured about 110 output tokens per second against about 57 for Carbon, and about 21 minutes per task against 36. That is roughly 1.7 to 1.9 times faster, not three times.
What is the context window and output limit?
The context window is 1,048,576 tokens shared between input and output. The default maximum output is 65,536 tokens, and long decode allows output up to the full context limit.
Can I try Gemini 4 Neon for free?
No. Neon is not on the free tier. Developers need a paid Gemini API account or a paid Google AI Pro or Ultra subscription through AI Studio, Antigravity or the Gemini app.
Sources
- Google, Introducing Gemini 4 Neon, published 18 November 2026, accessed 19 November 2026.
- Google, Gemini API pricing, accessed 19 November 2026.
- Google, Gemini 4 Neon model documentation, accessed 19 November 2026.
- Artificial Analysis, Google continues to accelerate its release cadence with Gemini 4 Neon, published 18 November 2026, accessed 19 November 2026.
- Techmeme, launch coverage index, accessed 19 November 2026, used only to identify the three times faster headline claim.