Skip to content
TEN Brief Ten verified stories a day 2026.11.20 KO

이 기사는 한국어로도 읽을 수 있습니다 →

Tech · 13 min read · Explainer

Gemini 4 Neon pricing — $3 per 1M, under Carbon from Dec 1

Gemini 4 Neon, released by Google on November 18, 2026 (US time), costs $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. The same rate applies across the full 1-million-token context and there is no free tier. Until November 30, 2026 it is 50% more expensive than Gemini 4 Carbon and Argon at $2 and $10, but those two rise to $4 and $20 on December 1, leaving Neon 25% cheaper. Google's model card shows Neon ahead of Claude Opus 5.5 and GPT-6.2 Sol on most tests and gives no Carbon comparison. Independent data puts Neon at 61 on the Artificial Analysis index at $6.44 per task, 36% below Opus 5.5 at $10.04 and more than six times GPT-6.2 Sol at $1.01. In the Gemini app it is open to Google AI Pro and Ultra subscribers

Two developers seen from behind looking at a laptop on a tall wooden table in a sunlit office

The three lines

  • Price — $3 input, $0.15 cached input and $15 output per million tokens, half that in batch. No free tier. Twice Hydrogen, and 75% of Carbon from December 1
  • Performance — Google's model card has Neon ahead of Opus 5.5 and GPT-6.2 Sol on most of 19 tests, with no Carbon scores. The independent index gives 61 against Carbon's 60
  • Cost — independently measured cost per task is 30-64% of Opus 5.5 and 2.3 to 6.4 times GPT-6.2 Sol. Neon was absent from the Vertex AI model list on November 20, 2026

Key questions

Gemini 4 Neon pricing
**$3 per million input tokens, $0.15 for cached input and $15 per million output tokens. It is not an introductory rate with an end date, and there is no free tier.** | Tier (per 1M tokens) | Input | Cached input | Output (includes thinking tokens) | |---|---|---|---| | Standard | $3.00 | $0.15 | $15.00 | | Batch | $1.50 | $0.075 | $7.50 | | Flex | $1.50 | $0.075 | $7.50 | | Priority | $5.40 | $0.27 | $27.00 | | Free tier | Not offered | — | — | Source: Google AI Studio price page, viewed November 20, 2026.
Gemini 4 Neon vs Carbon vs Hydrogen
**Neon sits between Hydrogen and Carbon. It costs twice as much as Hydrogen and, from December 1, 2026, 75% of Carbon.** | Item | Hydrogen | Neon | Carbon | |---|---|---|---| | Input / output (per 1M tokens) | $1.50 / $7.50 | $3 / $15 | $2 / $10, then $4 / $20 from December 1 | | Cached input | $0.075 | $0.15 | $0.10, then $0.20 | | API free tier | Yes | No | No | | Gemini app | Everyone (free users via auto selection) | Pro and Ultra subscribers | Ultra subscribers | | Artificial Analysis index (top setting) | 54 | 61 | 60 | | Cost per task at that setting | $3.16 | $6.44 | $7.76 | | Time per task at that setting | 15.3 min | 21.5 min | 37.5 min |
Gemini 4 Neon vs Opus 5.5 vs GPT-6.2 Sol
**On token price the order is Sol, Neon, Opus 5.5 from cheapest. Neon has the highest independent index score and Sol the lowest cost per task.** | Item | Gemini 4 Neon | Claude Opus 5.5 | GPT-6.2 Sol | |---|---|---|---| | Input / output (per 1M tokens) | $3 / $15 | $4 / $20 | $2 / $10 | | Cached input | $0.15 | $0.20 | $0.10 | | Artificial Analysis index | 61 | 59 | 57 | | Cost per task on that index | $6.44 | $10.04 | $1.01 | | Vals Index | 70.87% | 70.14% | 66.06% | | Vals cost per test | $7.40 | $24.77 | $3.18 |

Gemini 4 Neon costs $3 per million input tokens and $15 per million output tokens through the API. Google released Gemini 4 Neon, the fourth model in the Gemini 4 family, on November 18, 2026 (US time). The announcement was signed by Koray Kavukcuoglu, senior vice president at Google DeepMind. In Google's description Carbon takes the hardest long-horizon work, Hydrogen handles fast everyday execution, and Neon balances depth and efficiency. Artificial Analysis noted that this is Google's fourth model in seven weeks, as SiliconANGLE reported.

On the numbers alone Neon sits in an odd place. Until November 30, 2026 it is 50% more expensive than Carbon, the model one step above it, which is priced at $2 and $10. On December 1 Carbon and Argon rise to $4 and $20, the order rights itself and Neon becomes 25% cheaper. Performance needs the same care. Google's comparison table sets Neon against Claude Opus 5.5 and GPT-6.2 Sol, and Google has published no table with Neon and Carbon side by side. This article covers the price list, where the price falls within the family and against rivals, the gap between Google's claims and independent measurements, where the model can be used, and what remains open.

1. The price list — $3 input, $0.15 cached, $15 output

These figures come from Google AI Studio's price page, viewed November 20, 2026. The model ID is gemini-4-neon. Google's blog, 9to5Google and SiliconANGLE give the same numbers.

Tier (per 1M tokens)InputCached inputOutput (includes thinking tokens)Free tier
Standard$3.00$0.15$15.00None
Batch$1.50$0.075$7.50None
Flex$1.50$0.075$7.50None
Priority$5.40$0.27$27.00None

Five points matter when reading it.

First, there is no free tier. The price page says Carbon, Argon and Neon are not available on the free tier. Hydrogen is the only Gemini 4 model that is. Calling Neon through the API requires a project with billing attached. The page states that content on paid tiers is not used to improve Google's products.

Second, this is not an introductory price with an end date. The Carbon and Argon rows are marked with a change on December 1, 2026. The Neon and Hydrogen rows carry no date.

Third, there is no long-context surcharge. Google lists one rate for the 1-million-token context. OpenAI's GPT-6.2 Sol moves to $4 input and $15 output once input exceeds 272,000 tokens, according to OpenAI's price page. In that range Neon's $3 input is below Sol's and the output price is the same. For work that loads long documents or a large codebase in one request, the default-tier price is not the whole comparison.

Fourth, output charges include thinking tokens. Neon has four thinking levels: low, medium, high (the default) and xhigh. Higher levels spend more tokens before the answer is written, and those tokens are billed. We explained why output costs five times input in the August 15, 2026 article "What is AI token pricing."

Fifth, caching is automatic only. The price page says implicit caching is supported and explicit caching is not. Cached input is $0.15, or 5% of the input price. Because developers cannot create and pin a cache themselves, the hit rate depends on keeping the start of the prompt identical from call to call. The mechanism is covered in the September 5, 2026 article "What is prompt caching."

The specifications are in Google's model card. Inputs are text, images, audio and video. The context window is 1 million tokens shared between input and output; synchronous output is up to 64K tokens by default, with longer output available in background mode. The knowledge cutoff is March 2026. The model card says Neon was adapted from Carbon.

2. Price comparison — in the middle of the family, and in the middle of its rivals

Start with the Gemini 4 family. These are standard-tier prices from Google AI Studio's price page.

ModelInputCached inputOutputAPI free tierNote
Gemini 4 Hydrogen$1.50$0.075$7.50YesReleased November 12, 2026
Gemini 4 Neon$3$0.15$15NoReleased November 18, 2026
Gemini 4 Carbon (until November 30)$2$0.10$10NoIntroductory
Gemini 4 Carbon (from December 1)$4$0.20$20NoList
Gemini 4 Argon$2, then $4$0.10, then $0.20$10, then $20NoRises on the same day as Carbon

Neon is exactly twice Hydrogen on all three lines. Its relationship to Carbon flips with the calendar. Until November 30, 2026 Carbon is one-third cheaper than Neon. From December 1 Neon is 75% of Carbon. Carbon's introductory price and the increase schedule match what we reported in the November 12, 2026 article "Gemini 4 Carbon pricing and general availability."

Now the rivals. Google, OpenAI and Anthropic prices were rechecked on each company's price page on November 20, 2026.

ModelInputCached inputOutputInput price relative to Neon
Claude Fable 5.2$8$0.20$402.67x
GPT-6.2 Astra$8$0.40$402.67x
Claude Opus 5.5$4$0.20$201.33x
Gemini 4 Carbon and Argon (from December 1)$4$0.20$201.33x
Gemini 4 Neon$3$0.15$15Baseline
GPT-6.2 Sol$2$0.10$100.67x
Claude Sonnet 5.5$2$0.10$100.67x
Gemini 4 Hydrogen$1.50$0.075$7.500.5x
GPT-6.1 Terra$1$0.05$50.33x
MAI Code 2 (Copilot promotion)$0.50$0.025$2.500.17x

GPT models cost more above 272,000 input tokens ($16 and $60 for Astra, $4 and $15 for Sol). MAI Code 2's rate is a promotional price that applies only inside GitHub Copilot, so it does not sit comfortably in a table of API prices.

Neon's price lands exactly between the two models Google chose to compare it with. It is 25% cheaper than Opus 5.5 and 50% more expensive than GPT-6.2 Sol.

For a small service that uses 10 million input tokens and 2 million output tokens a month, the arithmetic at list price looks like this.

ModelInput chargeOutput chargeMonthly total
Gemini 4 Hydrogen$15$15$30
GPT-6.2 Sol$20$20$40
Gemini 4 Carbon (until November 30)$20$20$40
Gemini 4 Neon$30$30$60
Claude Opus 5.5$40$40$80
Gemini 4 Carbon (from December 1)$40$40$80
Claude Fable 5.2 and GPT-6.2 Astra$80$80$160

Agent workloads turn on caching. Assume 40 calls, 60,000 input tokens per call of which 80% is read from cache, and 3,000 output tokens per call.

ItemTokensRate (per 1M)Charge
Cached input1.92M$0.15$0.288
New input0.48M$3$1.44
Output0.12M$15$1.80
Total$3.528
With no cache hits2.4M input, 0.12M output$9.00

With caching the same session falls from $9 to about $3.53, a 61% reduction. The same assumptions gave $1.764 for Hydrogen in our November 16, 2026 article "Gemini 4 Hydrogen pricing." Double the rate, double the total. In practice models spend different numbers of tokens on the same job, so that ratio will not hold. Seeing that difference requires cost per task.

3. Performance — separating Google's claims from independent measurement

What Google published. Google's model card sets Neon, Claude Opus 5.5 and GPT-6.2 Sol against each other on 19 tests. Counting the top score in each row gives Neon 16, Opus 5.5 two and Sol one. DataCamp says 17 of 19, a difference of one row. Google's blog post gives no tally.

Test (Google model card)NeonOpus 5.5GPT-6.2 SolHighest
AutomationBench (work automation)56.3%42.5%41.3%Neon
Agents' Last Exam v1.175.8%71.6%73.0%Neon
GraphWalks (256K to 1M tokens)95.2%66.8%80.1%Neon
CursorBench 4.165.4%58.1%Not listedNeon
HLE-Diamond (reasoning)71.3%67.9%64.0%Neon
Terminal-bench 4.165.2%65.8%61.5%Opus 5.5
FrontierCode 2.046.2%47.3%43.8%Opus 5.5
Chartography (reading professional charts)73.5%66.3%75.1%GPT-6.2 Sol

Two of the three rows Neon loses are coding tests, by margins of 0.6 and 1.1 percentage points. Its leads on long-context GraphWalks and on work automation are much wider. Every number in the table is Google's own, and we could not confirm the settings used for the rival models.

The most important thing about the table is the missing column. Carbon is not in it. The model card contains no Carbon scores, and DataCamp also notes that Google published no Neon-versus-Carbon results. Google's blog post makes three Carbon comparisons: a 32% lower cost per task at the default high setting, similar performance on the security test ExploitGym with about half the tokens, and a 3.4% uplift on the patching test CWE-bench v2. All three are Google's own measurements.

SiliconANGLE reported Google as saying Neon completes most tasks about three times as quickly as Carbon and costs about 60% less. We did not find either figure in Google's blog post as viewed on November 20, 2026. They are confirmed in one place only.

What independent trackers published. This is the Artificial Analysis leaderboard as viewed on November 20, 2026.

Model (setting)Intelligence IndexCost per taskTime per task
Claude Fable 5.2 (max)65$26.5441.7 min
GPT-6.2 Astra (max)62$4.0014.1 min
Gemini 4 Neon (xhigh)61$6.4421.5 min
Gemini 4 Carbon (high)60$7.7637.5 min
Claude Opus 5.5 (max)59$10.0425.2 min
GPT-6.2 Sol (max)57$1.019.1 min
Gemini 4 Hydrogen (xhigh)54$3.1615.3 min
Gemini 4 Argon (high)50$4.9729.2 min

Against Carbon, Neon scores one point higher, costs 17% less per task and takes 43% less time. That is a smaller gap than Google's 32% at the default setting, but it points the same way. The closest independent figure to the three-times-faster line is a default-to-default comparison: Neon at high takes 14.8 minutes per task, about 2.5 times faster than Carbon's 37.5.

The same page shows how the thinking level changes the result.

Neon thinking levelIntelligence IndexTime per task
Low443.7 min
Medium558.3 min
High (default)6014.8 min
Xhigh6121.5 min

Moving from high to xhigh adds one point and 45% more time. Cost per task is published only for xhigh, so the default setting's cost is unknown. Neon at low (44) scores below Hydrogen at its top setting (54). For simple jobs, Hydrogen at half the price may be the better choice than Neon turned down.

Cost per task reads two ways. On the Vals Index Neon scores 70.87%, second of 42 models behind Fable 5.2 at 71.31%, at $7.40 per test. Vals ranks Neon first on its legal research and tax agent tests and seventh of 69 on Terminal-Bench 4.0.

ComparisonArtificial Analysis (per task)Vals (per test)
Neon$6.44$7.40
Opus 5.5$10.04$24.77
GPT-6.2 Sol$1.01$3.18
Neon as a share of Opus 5.564%30%
Neon as a multiple of Sol6.4x2.3x

Against Opus 5.5, Neon scores slightly higher at between one-third and two-thirds of the cost, a far bigger gap than the 25% difference in token price. Google's blog statement that Neon at its highest setting costs at least 36% less per task than rivals matches the Artificial Analysis numbers. Against GPT-6.2 Sol the picture reverses. Neon's list price is 1.5 times Sol's, but its cost per task is 2.3 to 6.4 times higher, for four to five more points. Whether that is worth it depends on the work. GPT-6.2 Astra, one tier up, scores one point above Neon at a lower $4 per task.

On safety, Google's model card says Neon is broadly comparable to Hydrogen and that unjustified refusals, meaning refusals of harmless requests, fell by 3.00 percentage points. For cyber and biological risk it carries forward earlier determinations that a critical level cannot be ruled out. The full safety report is listed as coming soon.

4. Where to use it — paid API only, Pro and above in the app, not in the Vertex AI list

RouteConditionSource
Gemini APIPaid tier. Model ID gemini-4-neonGoogle price page
Google AI StudioPaid-tier developers. Pro and Ultra subscribers can also try itGoogle blog
Gemini appGoogle AI Pro and Ultra subscribers, via the model pickerGoogle blog, subscription page
Google AntigravityPaid Google AI Pro and Ultra subscribersGoogle blog
Gemini for businessThrough the API or the model dropdownGoogle blog
Vertex AINot in the model list on November 20, 2026Google Cloud documentation
Cyber defendersFairwind Program, access starting within daysGoogle blog

The entry point is one step lower than Carbon's. Carbon is limited to Ultra subscribers in the Gemini app; Neon starts at Pro. According to 9to5Google, the app's model picker places Hydrogen under everyday help, Neon under deeper analysis and Carbon under code and complex work. The same outlet says Neon replaces Argon and Gemini 3.1 Pro in the app. Google's blog post does not address replacement. On the API side, Google's deprecation schedule shows no shutdown date for gemini-4-argon. From December 1, 2026 Argon will cost more than Neon while scoring 11 points lower on the independent index. Teams using Argon through the API have a reason to run the same work on Neon before that date.

There is a gap on the enterprise side. Google Cloud's Vertex AI model list shows Hydrogen, Carbon and Argon as generally available and does not show Neon. Beri.net reported the same as of November 19, 2026. Google's blog post describes the enterprise route only as Gemini for business and does not mention Vertex AI. Organizations that can adopt models only through Vertex AI region pinning or existing contracts will have to wait for a date.

Availability by country. Google has published no country-by-country schedule. The API region list is the firmest evidence: it names supported countries and carries no Neon-specific restriction. As one example, South Korea, where this publication is based, is on that list, and Google's Korean subscription page shows Neon under the Pro and Ultra plans (Ultra is 119,000 or 300,000 won a month; the Pro price was not displayed on the page we viewed). The free plan and the lower-priced AI Plus plan show Hydrogen only. For individuals the shortest route to Neon is a Google AI Pro subscription; for developers it is a Gemini API project with billing enabled. There is no free way to try it. The subscription page says app usage limits refresh every five hours until a weekly limit is reached and gives no counts. We found no published measurement of quality in Korean or other non-English languages.

5. What remains open and unconfirmed

Open questionStatus as of November 20, 2026
Neon versus Carbon scoresGoogle published no comparison table. The independent index reads 61 to 60
Three times faster and 60% cheaperSiliconANGLE only. Google's blog post has only the 32% cost-per-task figure
How many of the 19 tests Neon leads16 by our count of the model card table; DataCamp says 17
Cost per task at the default settingArtificial Analysis lists only the xhigh figure ($6.44)
Index rankGoogle says third; the Artificial Analysis model page says sixth of 95
Vertex AINot in the model list. No date announced
The future of ArgonOne report says Neon replaces it in the app. No API shutdown date
Neon's price after December 1No change marked on the price page, and no commitment that it will hold
Country rollout and non-English qualityNo schedule published. No public non-English measurement
Safety reportListed as coming soon on the model card

Neon's price has to be read against two dates. Until November 30, 2026 Carbon is cheaper and one point behind on the independent index. From December 1 Neon is 25% cheaper and faster. Google's table points at Opus 5.5 and Sol, but what actually changes for Google's customers is where to move work that ran on Carbon and Argon. Google did not publish the Carbon comparison that question needs. The remaining option is to run your own workload at the default setting and measure the cost per task.

For the model below Neon, see the November 16, 2026 article "Gemini 4 Hydrogen pricing." For the model above it and the December 1 increase, see the November 12 article "Gemini 4 Carbon pricing and general availability." The model at the top of the independent index is covered in the November 11 article "Claude Fable 5.2 released." The August 15 article "What is AI token pricing" and the September 5 article "What is prompt caching" explain why list price and the final bill diverge.

Sources

  1. Google Blog — Introducing Gemini 4 Neon (2026-11-18)
  2. Google AI Studio — Gemini API pricing (gemini-4-neon, gemini-4-carbon, gemini-4-argon, gemini-4-hydrogen, viewed 2026-11-20)
  3. Google DeepMind — Gemini 4 Neon model card (2026-11-18)
  4. Google AI Studio — Gemini API deprecations (viewed 2026-11-20)
  5. Google AI Studio — Available regions for the Gemini API and AI Studio (viewed 2026-11-20)
  6. Google Cloud — Vertex AI model list (viewed 2026-11-20)
  7. Google — Gemini subscriptions (Korean page, viewed 2026-11-20)
  8. Artificial Analysis — Gemini 4 Neon model page (viewed 2026-11-20)
  9. Artificial Analysis — LLM Leaderboard (viewed 2026-11-20)
  10. Vals AI — Gemini 4 Neon model page (updated 2026-11-18)
  11. SiliconANGLE — Google's latest midtier model Gemini 4 Neon is three times faster than Carbon and wins on some metrics (2026-11-18)
  12. 9to5Google — Gemini 4 Neon announcement (2026-11-18)
  13. DataCamp — Gemini 4 Neon (2026-11-19)
  14. Beri.net — Gemini 4 Neon GA: $3/$15 pricing, 1M context, thinking levels vs Opus 5.5 and GPT-6.2 Sol (2026-11-19)
  15. OpenAI — API pricing (gpt-6.2-astra, gpt-6.2-sol, gpt-6.1-terra, viewed 2026-11-20)
  16. Anthropic — Claude API pricing (Fable 5.2, Opus 5.5, Sonnet 5.5, viewed 2026-11-20)

Verification

Published
Last modified
Cross-check
Checked against 16 independent sources.
Unverified
  • The statement that Neon is about three times faster and about 60% cheaper than Carbon appears only in SiliconANGLE. Google's blog post as viewed on November 20, 2026 does not contain those figures; it says only that Neon at its default high setting has a 32% lower cost per task than Carbon. We could not establish what the 60% is measured against. Google's 32% is its own measurement and has not been reproduced independently.
  • Counting the top score in each of the 19 rows of Google's model card table gives Neon 16 (Opus 5.5 leads on Terminal-bench 4.1 and FrontierCode 2.0, GPT-6.2 Sol on Chartography). DataCamp says 17 of 19. Google's blog post gives no tally. This article uses our count of 16 and flags the difference. We could not confirm the settings Google used for the rival models.
  • The Artificial Analysis rank differs by source: Google's blog and SiliconANGLE say third, while the Artificial Analysis model page says sixth of 95. Read model by model, the leaderboard puts Neon third behind Fable 5.2 (65) and GPT-6.2 Astra (62). Sixth appears to count other settings of the same models as separate rows, but we could not confirm that.
  • Neon's cost per task is published only for the xhigh setting ($6.44). The default high setting has no cost figure on the Artificial Analysis page. Carbon's $7.76 per task and 37.5 minutes are the values viewed on November 20, 2026. They differ from the values we viewed for our November 16 article ($5.48 and 33.8 minutes), and we could not determine why.
  • The Vals Index figures for Opus 5.5 (70.14%, $24.77 per test) and GPT-6.2 Sol (66.06%, $3.18) are relayed by Beri.net. On Vals's Neon page we confirmed only Neon's own values (70.87%, $7.405, second of 42). The scores match Google's model card table (70.1% and 66.1%).
  • That Neon replaces Argon and Gemini 3.1 Pro in the Gemini app, the labels in the model picker and the make-up of auto selection for paid users come from 9to5Google alone. Google's blog post does not address replacement. We confirmed only that Argon no longer appears on Google's Korean subscription page. No shutdown date has been announced for gemini-4-argon in the API.
  • Neon's absence from the Vertex AI model list was confirmed on the Google Cloud documentation page viewed November 20, 2026 and in Beri.net's November 19 account. Google describes the enterprise route only as Gemini for business. We found no Google statement on Vertex AI timing. Availability on OpenRouter comes from DataCamp alone.
  • We could not confirm when Neon reached accounts in any given country. For South Korea we confirmed only that Google's Korean subscription page shows Neon under the Pro and Ultra plans and that the country is on the API region list. The page did not display a won price for Google AI Pro. We found no published measurement of Korean-language or other non-English quality.
  • Maximum output is given as 64K tokens by default in Google's model card and as 65,536 tokens by DataCamp, which appear to be two notations of the same value. The DataCamp author's chess page test (Neon succeeded in 3 of 3 runs at $0.83 to $1.06 each, Carbon in 1 of 3) is one person's three runs per model.
  • The safety figures in Google's blog post (99.2% on the agent honesty evaluation, a 0.09% misaligned action rate, a 0.0% indirect prompt injection success rate) and the cyber benchmark scores are Google's own evaluations or tests it commissioned. The full safety report is listed as coming soon on the model card.
  • MAI Code 2's price ($0.50 and $2.50) is a promotional rate inside GitHub Copilot, confirmed in our November 19, 2026 reporting. We did not reopen GitHub's price page. The cost examples in the body multiply list prices by assumed token counts and will differ from real bills.
Authoring
Reviewed by a person before publication. The full process is described in the Editorial.

Ten stories, once each morning

We send the three-line summaries only; the full pieces stay on the site. One-click unsubscribe, any time.

Related