Overview
Gemini 4 Neon is Google DeepMind's model released on November 18, 2026, with rollout starting November 19. It is the fourth new Gemini model in a single month. Hydrogen to Neon took six days; Argon to Carbon took twenty-eight. The cadence itself is a signal: this generation is not finished, and the next point release is already queued.
Google positions it as frontier intelligence paired with fast end-to-end task completion, aimed at complex software engineering, knowledge work and scientific workflows. The family map makes the placement clearer. Carbon handles the hardest long-horizon projects, Hydrogen handles fast everyday execution and can also run as a parallel sub-agent under Neon, and Neon sits in the middle, able to complete a whole task alone or act as an orchestrator with enough judgement to delegate well.
Pricing is 3 dollars per million input tokens and 15 dollars per million output tokens, with cached input at 0.15 dollars, which Google describes as a 95 percent discount. On November 19 that reads as expensive: Carbon and Argon are 2 and 10 dollars, Hydrogen is 1.50 and 7.50. But Carbon and Argon double to 4 and 20 dollars on December 1, and from then on Neon is the cheaper option in that tier. The ordering flips, which is easy to get wrong in a long-run cost model.
The Artificial Analysis Intelligence Index score is 61 at the xhigh effort level, one point above Carbon at high, third overall behind Claude Fable 5.2 and GPT-6.2 Astra. Vals Index puts it second at 70.9 percent on the quality-to-price Pareto frontier. The new Agent Arena Index gives it 1336 points and second place, measured across more than 4,200 real agent sessions.
Cost is reported three ways. At the default high level a task costs 32 percent less than the same task on Carbon. At xhigh it costs more than 36 percent less than the highest tier of competing frontier models. And at high it sits on the Vals and Agent Arena Pareto frontiers at roughly 50 to 80 percent below the leading frontier models per task. On the Agent Arena table its median cost per task is 2.82 dollars, scoring above Claude Opus 5.5 at Max while costing 53 percent less.
Coding is the strongest category here. Artificial Analysis Coding Agent Index places it first at 58.6, Terminal-Bench 4.1 comes in at 65.2 percent, and FrontierCode 2.0 at the default level is 46.2 percent. On a single SWE-sweep benchmark it scores 12.8 percent under the mini-SWE harness and 39.0 percent under Antigravity, which Google calls state of the art and more than triple the score. Cursor reported 63.7 percent on CursorBench 4.1 at default strength against 57.3 percent for Gemini 4 Hydrogen, a new cost-performance frontier on that benchmark.
The context window is one million tokens with four adjustable thinking levels, high by default and xhigh reserved for the hardest engineering, analysis and scientific work. Enterprise and analytical benchmarks also land near the top. Legal Research Bench and Tax Agent Bench are both first, Surge AI's finance, accounting, spreadsheet modeling and constraint benchmarks are mostly first, Chartography is 73.5 percent, Agents' Last Exam v1.1 is 75.8 percent, Zapier AutomationBench is 56.3 percent for first place, and Andon Labs Vending-Bench 2 finishes first at 16,843 dollars. On the research side, Terminal-Bench Science is 70.5 percent, RiemannBench 84.0 percent and HLE-Diamond 71.3 percent.
Cybersecurity ships as a separate variant called Neon for cyber, distributed through the Fairwind program. Its numbers should be read on their own: 76 percent on AutoBaxBench hard, 51.26 percent on WebExploitBench v1.2, 90.5 percent on SEC-bench Pro, and 70.7 percent on Collinear CWE-bench v2, 3.4 points above Carbon, using half the tokens Carbon needs on ExploitGym. These belong to a different variant on a different distribution channel, so they should not be folded into general model capability claims.
Steerability and safety have numbers too. Agent Arena ranks steerability second, meaning the ability to be corrected mid-task. On Terminal-Bench 4.0 the refusal rate is 1.98 percent, zero on the Science variant, the cheating rate is 0.0 percent, shortcut-taking on coding evals is 2.45 percent, the lowest in the Gemini 4 family, agentic safety misalignment is 0.09 percent, the agent honesty evaluation passes at 99.2 percent, Gray Swan prompt-injection attacks are fully blocked at xhigh with real-time protection, under-service drops 43 to 56 percent and overreach drops 15 to 35 percent.
Availability is broad. It is live in the paid Gemini API tier, Google AI Studio, Android Studio and Antigravity for paid Pro and Ultra users, Gemini for business, and the Gemini app for Pro and Ultra subscribers. Fairwind eligible members get access within days. OpenRouter and Cursor both launched it the same day, with OpenRouter routing through Google Cloud.
Three things are not on the official page. Maximum output tokens are not listed. Speed is described only qualitatively; the claim of roughly three times Carbon comes from an individual Google employee post, while the official blog says only faster and cheaper. Country-level availability is not listed either, only the rolling rollout language. One community signal is worth noting: some developers argue Google's biggest current problem is pairing its strongest models with a previous-generation harness, so the Antigravity experience lags the benchmark numbers. That is not an official conclusion, but it is worth a trial in your own workflow before committing.
Key Features
- Four thinking levels and a 1M context: Four adjustable effort levels, high by default, xhigh for the hardest engineering, analysis and science work. The context window is one million tokens, enough for multi-file repositories, long reports and cross-document retrieval.
- First on the coding agent index: 58.6 on the Artificial Analysis Coding Agent Index for first place. Terminal-Bench 4.1 at 65.2 percent, FrontierCode 2.0 at 46.2 percent, and SWE-sweep rising from 12.8 to 39.0 percent when moving from the mini-SWE harness to Antigravity.
- Enterprise analysis benchmarks cluster at the top: First on Legal Research Bench and Tax Agent Bench, first on Zapier AutomationBench at 56.3 percent, first on Vending-Bench 2 at 16,843 dollars, 73.5 percent on Chartography and 75.8 percent on Agents' Last Exam v1.1.
- Neon for cyber as a separate variant: A security-focused variant distributed through the Fairwind program. 76 percent on AutoBaxBench hard, 90.5 percent on SEC-bench Pro, 51.26 percent on WebExploitBench v1.2, and 3.4 points above Carbon on Collinear CWE-bench v2.
- Frontier capability at a lower task cost: 32 percent cheaper per task than Carbon at the default high level, more than 36 percent below the top competing tier at xhigh, and roughly 50 to 80 percent below leading frontier models on the Vals and Agent Arena Pareto frontiers. Cached input is billed at 0.15 dollars, a 95 percent discount.
- Wide same-day availability: Live in the paid Gemini API, AI Studio, Android Studio, Antigravity, Gemini for business and the Gemini app for Pro and Ultra subscribers, with OpenRouter and Cursor integrations shipping the same day.
- Steerability and low fabrication: Second on Agent Arena steerability, so it can be redirected mid-task. Google says internal Antigravity testing produced its lowest fabrication rate to date, with a 1.98 percent refusal rate and 0.0 percent cheating on Terminal-Bench 4.0.
Use Cases
- Debugging, refactoring and cross-module changes across a multi-file repository, including running the terminal test suite afterwards
- Long-chain enterprise work such as legal research, tax and financial modeling where conclusions need to be traceable
- Turning dense research material, charts and PDFs into structured reports and then interpreting the figures
- Acting as an orchestrator that dispatches several Hydrogen sub-agents in parallel while keeping judgement and synthesis
- Vulnerability validation and exploit-surface assessment for security teams, through the Fairwind channel for the cyber variant
- Generating front-end pages in bulk and comparing models on real-scenario benchmarks such as CursorBench
Pros
- First on the coding agent index, and the SWE-sweep jump across harnesses shows the capability is not tied to one scaffold
- Task cost runs more than 30 percent below models at the same score band, and the 2.82 dollar median on Agent Arena is usable for budgeting
- A one-million-token window with four effort levels lets long and short tasks share one model instead of swapping between models
- Enterprise benchmarks concentrate on legal, tax, spreadsheet modeling and automated workflows, which is where agents actually land
- Safety and steerability data is unusually specific, with refusal, cheating, shortcut and prompt-injection numbers all published
- OpenRouter, Cursor, AI Studio and Antigravity all shipped on day one, keeping switching costs low
Pricing
The paid Gemini API tier charges 3 dollars per million input tokens and 15 dollars per million output tokens, with cached input at 0.15 dollars, which Google describes as a 95 percent discount. That is expensive inside the Gemini 4 family today: Carbon and Argon are 2 and 10 dollars, Hydrogen is 1.50 and 7.50. But Carbon and Argon move to 4 and 20 dollars on December 1, 2026, at which point Neon becomes the cheaper choice in that tier. Note that Google's own comparisons are per-task cost rather than per-token price, and both headline numbers, 32 percent below Carbon at high and more than 36 percent below the top competing tier at xhigh, already account for token consumption. On the subscription side, Gemini app Pro and Ultra users can select 4 Neon from the model dropdown, labeled Deeper analysis. Check the official Google pricing page for billing details.
Summary
Neon fills the gap between Carbon and Hydrogen: frontier-level judgement with the speed to finish a task end to end. Its strong numbers are in coding and agents, first on the Coding Agent Index, second on Agent Arena at 1336, and a 2.82 dollar median cost per task. At 3 and 15 dollars it currently costs more than Carbon, but that flips when Carbon and Argon double on December 1, so run the comparison at the date you will actually be paying. It suits multi-file refactors, long-chain enterprise analysis and workflows that orchestrate several sub-agents. For short tasks where the lowest unit price matters, Hydrogen is still cheaper. Security work should look at Neon for cyber, a separate variant distributed through Fairwind.
Version History
- Gemini 4 Neon released and opened up (2026-11-18): The Gemini API changelog confirmed gemini-4-neon as generally available, with rollout from November 19 to Gemini app Pro and Ultra, AI Studio, Android Studio, Antigravity, Gemini for business and the Fairwind program. Positioned between Carbon and Hydrogen with a one-million-token context window and four thinking levels, priced at 3 and 15 dollars per million tokens with cached input at 0.15 dollars. Artificial Analysis ranks it third on the Intelligence Index at 61, first on the Coding Agent Index at 58.6 and second on the Agent Arena Index at 1336.