Google Releases Gemini 4 Neon at 40% of Flagship Pricing

Original title: 谷歌发布 Gemini 4 Neon:独立评测排第三,价格不到 GPT、Claude 旗舰的四成,Pro 和 Ultra 订阅用户今天起能用
78AI Score

Google released Gemini 4 Neon, which scores 61 on Artificial Analysis' composite intelligence index, ranking third behind GPT-6.2 Astra and Claude Fable 5.2 while costing under 40% of those flagships. Positioned between Carbon and Hydrogen, Neon is priced at $3/$15 per million input/output tokens with cached input at $0.15, accepts text, image, video and audio input, and is available today to Google AI Pro and Ultra subscribers.

Full Report

宝玉 (@dotey) · 2026-11-19 00

Google releases Gemini 4 Neon: ranked third in independent evaluations, priced at less than 40% of the GPT and Claude flagships, available to Pro and Ultra subscribers starting today

Google released the large model Gemini 4 Neon on November 18. On the composite intelligence index of the independent evaluation firm Artificial Analysis, Neon scores 61 points, ranking third overall, behind only OpenAI's GPT-6.2 Astra and Anthropic's Claude Fable 5.2, while its price is less than 40% of these two flagship models. Users subscribed to Google AI Pro or Ultra can select it directly in the Gemini App starting today, counted as part of the subscription with no extra charge.

Developers use Neon through the API. The API is the interface through which developers integrate the model into their own products, billed by usage, with the unit of measurement being tokens; one Chinese character roughly counts as one to two tokens. Neon's price is $3 per 1 million input tokens and $15 per 1 million output tokens; repeatedly used input content goes through caching and is charged only $0.15, which is 5% of the original price. Neon can read text, images, video, and speech, but outputs only text.

[What is its relationship with Argon, Carbon, and Hydrogen]

Gemini 4 has released four models within seven weeks. The earliest was Argon, released on September 30, which Google at the time called a "frontier model." Then came the stronger Carbon (opened to everyone on November 10), the smaller and faster Hydrogen (November 12), and now Neon.

The family comparison chart Google provided this time lists only three: Carbon is the strongest, used for the hardest and most time-consuming projects; Hydrogen is the fastest and cheapest, handling everyday tasks and available even to free users; Neon sits in the middle. The chart does not include Argon, and Google has not said whether Argon will be discontinued; developers can still call it through the API. When Google introduced Carbon on November 10, it also said Argon "remains the ideal choice for most software development and knowledge work," which largely overlaps with Neon's current positioning.

Comparing with Artificial Analysis data: on the composite intelligence index, Argon scores 50 points, lower even than Hydrogen (54 points), and Neon is 11 points higher than it; on output speed, Argon is about 66 tokens per second, and Neon about 110. On price, Argon is the same as Carbon, at a promotional price of 2/2/10 until November 30, after which it returns to 4/4/20, at which point Neon is 25% cheaper than them; Hydrogen is 1.5/1.5/7.5.

Neon and Carbon have almost the same intelligence index score (61 vs. 60), but Neon is cheaper and faster. Artificial Analysis uses the same set of questions to test all models and calculates the average cost and time to solve one question. Neon at its highest setting (xhigh) costs $6.44 per question and takes about 21 minutes; Carbon at its normal price of 4/4/20 costs $7.76 and takes about 36 minutes. Neon actually uses about 15% more output tokens; the savings come entirely from its lower unit price, and the time savings come from its faster output (about 110 tokens per second, versus about 57 for Carbon). If Neon uses the default high setting, its score matches Carbon's, and the cost per question is less than half of Carbon's.

[Comparison with GPT and Claude]

Neon is not in the most expensive flagship tier. GPT-6.2 Astra (released November 17) and Claude Fable 5.2 (released November 9) are both $8 input and $40 output, and Neon's price is 37.5% of theirs. One tier down, Claude Opus 5.5 is 4/4/20, one-third more expensive than Neon; GPT-6.2 Sol is 2/2/10, one-third cheaper than Neon.

In its release article, Google compares Neon with the next tier down, Opus 5.5 and GPT-6.2 Sol. In Google's table, Neon leads in most tests, for example scoring 56.3% on AutomationBench, which has AI automatically handle office workflows, while the other two are between 41% and 43%. There are also areas where it trails: on the coding tests Terminal-Bench 4.1 and FrontierCode 2.0, Opus 5.5 is slightly higher (65.8% vs. 65.2%, 47.3% vs. 46.2%); on Chartography, which reads professional charts, GPT-6.2 Sol is slightly higher (75.1% vs. 73.5%).

On coding, Artificial Analysis compared it directly with the flagships. Its coding agent index measures how a model performs when paired with its own coding tools: Neon paired with Google's Antigravity command-line tool scores 59 points, while Claude Fable 5.2 and Opus 5.5 paired with Claude Code both score 58 points. The average cost per task is $6.86 for Neon, $12.61 for Opus 5.5, and $15.47 for Fable 5.2.

In terms of specifications, Neon can read up to 1 million tokens at a time, equivalent to thousands of pages of documents; GPT-6.2 and Claude's current models are also around 1 million. Neon's default maximum output per single response is 65,000 tokens, while GPT-6.2 Sol, GPT-6.2 Astra, and Claude's current models are all 128,000, so Neon is half of them; Neon can enable Google's "long output" feature to continue writing in segments, with input plus output totaling up to about 1.05 million tokens.

[What it can be used for]

Google highlights three main categories of use:

  1. Writing code. Neon is suited for "agentic" work, where the AI performs multiple steps in sequence on its own—for example, reading code, modifying code, running tests, and then continuing to make changes based on the results. In Google's AI coding tool Antigravity, Neon can also act as the orchestrator, breaking work apart and distributing it to multiple Hydrogens to do simultaneously. Google says it will read duplicate files less, and when working for very long periods it can properly compress earlier content and keep going.

  2. Processing office materials. Spreadsheets, PDFs, charts, audio, and video can all be handed to it together and organized into a report. In the AA-Briefcase test, which evaluates the results of office-type tasks by Artificial Analysis, Neon ranked first in the "presentation quality" category (1771 points), surpassing the previous first-place Claude Sonnet 5.5 (1726 points).

  3. Everyday tasks. The "Tasks" feature in the Gemini App can now be run using Neon. Google's examples are analyzing bills to find out why utility costs suddenly rose, and checking multiple receipts to apply for a refund.

Neon has four thinking levels: low, medium, high, and xhigh. The higher the level, the longer the model thinks before answering, producing better results and using more tokens. The default is high. The highest xhigh level was only added to Gemini last week when Google released Hydrogen; Carbon does not have this level.

Google has also made a cybersecurity-specific version, Neon for cyber, used to find code vulnerabilities and write patches. In the coming days, it will first be opened through Google's Fairwind program to vetted security defense personnel.

[How to use it]

  1. Regular users: Google AI Pro and Ultra subscribers can select Neon from the model menu in the Gemini App. Free users cannot use it; what free users can use is Hydrogen.

  2. Developers: The paid tier of the Gemini API in Google AI Studio, as well as Android Studio. Pro and Ultra subscribers can also use it in Antigravity.

  3. Enterprises: Gemini for business, either via the API or by selecting it from the model menu.

The test scores in Google's table, including the scores for Opus 5.5 and GPT-6.2 Sol, were all published by Google; the rankings, cost per question, and speed data come from Artificial Analysis's evaluation on the day of release. Competitors' prices and output limits come from the official model pages of OpenAI and Anthropic.

Original link: https://t.co/q2qeH9jIRX Source link: Google blog "Introducing Gemini 4 Neon" (2026-11-18) https://t.co/q2qeH9jIRX

↳ Quoting @sundarpichai (Sundar Pichai): Gemini 4 Neon: frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows.

Begins rolling out today in the Gemini API in @GoogleAIStudio, Gemini for business, for paid Google AI Pro and Ultra subscribers in Google @Antigravity, and for Google AI Pro and Ultra subscribers in the @GeminiApp.

Tweet image