Google has launched Gemini 4 Neon, a new artificial intelligence model built for fast reasoning across complex coding, knowledge work and science, priced at $3 per million input tokens and $15 per million output tokens. Google DeepMind published the model’s technical card on Tuesday 18 November, and the model began rolling out the same day through Google’s developer tools and paid subscription products. Neon is the fourth model in the Gemini 4 family to arrive in just over five weeks, and Google positions it in the middle of the range: more capable than the fast, low-cost Hydrogen, but cheaper and quicker than the flagship Carbon, which is designed for the hardest long-running projects.
Key Takeaways
- Google launched Gemini 4 Neon on 18 November 2026, a fast reasoning model adapted from the flagship Gemini 4 Carbon and aimed at complex coding, analytical reasoning and enterprise work.
- Neon costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million — double the price of Hydrogen, but below Carbon’s standard $4/$20 rate from 1 December.
- It has a unified context window of 1,048,576 tokens and can produce up to 64,000 tokens by default, or up to 1 million tokens through a background long-decode mode.
- The model accepts text, image, audio and video inputs and returns text; its knowledge cutoff is March 2026.
- Google’s own model-card table gives Neon 61 on the Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5 at 59 and GPT-6.2 Sol at 57 — but no independent third-party evaluation of Neon has been published yet.
- Neon is available through the Gemini API, AI Studio, Android Studio, Gemini for business, Antigravity and the Gemini app for Pro and Ultra subscribers, and is generally available in GitHub Copilot’s higher tiers.
Where Neon sits in the Gemini 4 family
Google now has four Gemini 4 models, each aimed at a different kind of work. Argon became generally available on 13 October, Carbon was announced on 10 November, and Hydrogen, the fast low-cost model for everyday tasks, followed on 12 November. Neon arrived on 18 November, completing a rapid run of releases. Neon is adapted from Carbon, according to the model card published by Google DeepMind, which means it is built on the flagship model’s foundations rather than trained from scratch as a separate system. Google describes the difference in terms of trade-offs: Carbon tackles the hardest, long-horizon projects; Hydrogen handles fast everyday execution; Neon balances depth and efficiency. In Google’s own words, it delivers frontier-level reasoning for complex, high-frequency workflows, faster and more cost efficiently than the larger model.
Price, and the December turning point
Neon’s headline price is $3 per million input tokens and $15 per million output tokens. Cached input — repeated prompts or documents the model has already processed — costs $0.15 per million tokens, a 95 percent discount on the standard input rate. The price only makes sense in the context of the rest of the family. Hydrogen costs $1.50 per million input tokens and $7.50 for output, exactly half of Neon’s rates. Carbon and Argon are currently on an introductory rate of $2 for input and $10 for output, which makes them cheaper per token than Neon today — but that rate ends on 30 November. From 1 December, Carbon and Argon rise to $4 per million input tokens and $20 for output, and from that point Neon undercuts them by 25 percent. Google also argues that the price per token is not the full story. At its default high thinking level, the company says Neon costs 32 percent less per task than Carbon, because it uses fewer tokens to reach a similar result. Those are Google’s own figures, measured on its own test workloads, and developers will need to test them against their own work before deciding whether the savings hold up in practice.
A million-token window, built for long jobs
Neon’s most important technical feature is the size of its working memory. It has a unified context window of 1,048,576 tokens — roughly a million tokens shared between what goes in and what comes out. That is enough room for very large codebases, long documents, spreadsheets, PDFs, charts, audio and video to be handled in a single session. By default, a single response can run to 64,000 tokens. A background mode called long decode can extend output up to 1 million tokens, for jobs that need extremely long responses. The model accepts text, images, audio and video as inputs and returns text. It does not generate images or audio, according to Google’s documentation — it is a workhorse for agents and long tasks, not a media model. Neon also offers four thinking levels, settings that trade answer depth against token spend. High is the default. The highest level, xhigh, adds deeper multi-step reasoning for the hardest workloads, and it is at this level that Neon records its best benchmark scores. Google says the model can act as an orchestrator in its Antigravity coding tool, directing several Hydrogen subagents working in parallel. Google’s model card lists a knowledge cutoff of March 2026, and warns that in some domains the model’s knowledge may be less up to date than that.
What Google’s own benchmark table claims
All of the performance figures published so far come from Google itself. No independent third-party evaluation of Neon has been published yet, and the scores below should be read as Google’s account of its own model’s performance, not as settled fact. On the Artificial Analysis Intelligence Index, Google reports that Neon scores 61 at its highest thinking level, ranking third overall behind larger, slower models. The model card’s comparison table puts Claude Opus 5.5 at 59 and OpenAI’s GPT-6.2 Sol at 57 on the same index. On the Vals Index, which measures professional knowledge work, Neon scores 70.9 percent and ranks second. In coding, Neon ranks first overall on the Artificial Analysis Coding Agent Index with 58.6, scores 65.2 percent on Terminal-Bench 4.1, and 46.2 percent on FrontierCode 2.0 at its default high setting. On agent tasks, it scores 56.3 percent on Zapier’s AutomationBench and tops Andon Labs’ Vending-Bench 2, a long-horizon simulation of running a business — both first-place results, according to Google. In science and mathematics, Google reports Neon as best in class on Terminal-Bench Science at 70.5 percent, RiemannBench at 84.0 percent, and HLE-Diamond at 71.3 percent. Its long-context score stands out: 95.2 percent on GraphWalks, a test of reasoning across up to 1 million tokens, against 66.8 percent for Claude Opus 5.5 and 80.1 percent for GPT-6.2 Sol. There is also a dedicated version for cybersecurity work. Neon for cyber produces more secure code than Carbon across easy, medium and hard challenges on AutoBaxBench, according to Google, and reaches 90.5 percent on SEC-bench Pro for finding code vulnerabilities. That version is rolling out to vetted defenders through Google’s Fairwind Program over the coming days.
Where you can use it
Neon began rolling out on 18 November in the Gemini API paid tier through Google AI Studio and Android Studio, and in Gemini for business through the API and model dropdown. Paid Google AI Pro and Ultra subscribers can use it in Google Antigravity and select it from the model dropdown in the Gemini app. Third-party tools moved quickly. GitHub announced the same day that Gemini 4 Neon is generally available in Copilot Pro+, Max, Business and Enterprise tiers, appearing across VS Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, the Copilot command-line tool, the coding agent, github.com and GitHub Mobile. Copilot Free and Pro are not on that list, and GitHub says the rollout is gradual, so some users may not see the model yet.
Safety, and what is still missing
Google says Neon builds on the safety protections developed for Carbon, including safeguards against cyber and biosecurity misuse comparable to Carbon’s. In internal testing, Google reports that Neon recorded its lowest fabrication rate yet in Antigravity and achieved a 99.2 percent pass rate on its agent honesty evaluation. The model card is open about the limits. It lists hallucinations as a known limitation and says Google is still working to improve resistance to jailbreaks — attempts to get a model to bypass its safeguards. The full Safety and Security Report for Neon has not yet been published; the model card marks it as coming soon. Google also carried forward its earlier Frontier Safety Framework results, under which it says it cannot rule out a critical capability level for biological uplift or for cybersecurity uplift, and says it is deploying mitigations in both areas.
Genie 4 arrives alongside Neon
Neon was not the only launch on 18 November. Google DeepMind also began rolling out Genie 4, its latest interactive world model, inside the browser-based Project Genie platform. Genie 4 is described as Google’s most realistic world model yet, with native audio and the ability for users to change the simulated world with text instructions while a session is running. Access is rolling out first to eligible Google AI Ultra subscribers, with a waitlist open for Google AI Pro members.
What to watch
Three things will settle how important Neon turns out to be. First, independent benchmarks: the scores published so far are Google’s own, and outside testing in the coming weeks will show how close Neon really is to Carbon and to its rivals. Second, the 1 December price change, when Carbon and Argon rise to $4 and $20 per million tokens and Neon’s mid-tier pricing starts to look like a discount rather than a premium. Third, whether developers adopt it in practice — in the Gemini API, in Antigravity and in GitHub Copilot — or stick with the models they already use. The pace of releases is itself part of the story: Google has now shipped four Gemini 4 models in just over five weeks. With OpenAI having launched its GPT-6.2 models the day before Neon arrived, and with Apple’s iCloud outage earlier this week, it has been an unusually crowded week in technology — and Neon is Google’s bet that the middle of the market is where the next round of competition will be won.
Frequently Asked Questions
What is Gemini 4 Neon and how much does it cost?
Gemini 4 Neon is Google’s new fast reasoning AI model, launched on 18 November 2026 and adapted from the flagship Gemini 4 Carbon. It costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million. It sits between Hydrogen ($1.50/$7.50) and Carbon ($4/$20 from 1 December) in Google’s Gemini 4 range.
Where can I use Gemini 4 Neon?
Neon is rolling out in the Gemini API through Google AI Studio and Android Studio, in Gemini for business, and in Google Antigravity and the Gemini app for paid Google AI Pro and Ultra subscribers. It is also generally available in GitHub Copilot Pro+, Max, Business and Enterprise tiers. Free Gemini app users are not listed for access.
How good is Gemini 4 Neon compared with rival models?
According to Google’s own model card, Neon scores 61 on the Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5 at 59 and GPT-6.2 Sol at 57, and ranks first on the Artificial Analysis Coding Agent Index. No independent third-party evaluation has been published yet, so these figures are Google’s own claims and should be treated as such until outside testing confirms them.
Sources
- Google Blog, “Introducing Gemini 4 Neon”, 18 Nov 2026
- Google DeepMind, “Gemini 4 Neon” model card, published 18 Nov 2026
- DailyTechNow, “Gemini 4 Neon: Price, Access and What Changes”, 19 Nov 2026
- AI Frontier Post, “Google releases Gemini 4 Neon, the mid-tier model its billing catalog leaked this morning”, 18 Nov 2026
- SXF Brief, “AI News Nov 19, 2026: Gemini 4 Neon & OpenAI Chip Design”, 19 Nov 2026
