Google places Neon between its two other Gemini 4 releases. Carbon is meant for the hardest, long-horizon projects and Hydrogen for fast everyday execution, while Neon aims to balance depth with efficiency in high-frequency enterprise work. Its rates are double the $1.50 and $7.50 per million tokens Google charges for Gemini 4 Hydrogen, which began rolling out on November 12. Carbon costs an introductory $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 once that period expires on November 30, 2026. Cached input on Neon costs $0.15 per million tokens, a 95% discount, and its context window holds 1 million tokens.
Users can choose among four thinking levels, deciding how much reasoning, and how many tokens, a given job deserves. At the default high level, Google said Neon costs 32% less per task than Carbon and runs roughly 50% to 80% cheaper per task than top competing frontier models on the Vals Index and Agent Arena. At xhigh, the deepest setting, it posts its best intelligence score while still costing at least 36% less per task than rival models at their top settings, according to the company. Once Carbon moves to its full price, Neon's lower per-token rates could make it the more economical choice for teams whose work does not need Google's largest model.
On the third-party leaderboards Google cited, Neon scored 61 on the Artificial Analysis Intelligence Index at xhigh, ranking #3 overall and ahead of comparable models, just behind larger, slower systems. It is #2 on the Vals Index with 70.9%, sitting on the Pareto frontier, and debuted #2 globally on Arena's new Agent Arena Index with a score of 1336 across more than 4,200 real-world agent sessions. It also placed #2 there on steerability, a measure of how cleanly a model adjusts when a user corrects it mid-task, and Google said it fabricated less than any of its earlier models in internal Antigravity tests.
Several results bear directly on finance. Neon is #1 on both the Vals Legal Research Bench and Tax Agent Bench, and Google said it tops most Surge AI benchmarks covering finance, accounting, spreadsheet modeling and complex constraint following. It scored 73.5% on Chartography, which tests the reading of dense professional charts, and 75.8% on Agents' Last Exam v1.1, level with leading models. It also came #1 overall on Zapier's AutomationBench (56.3%) and on Andon Labs' Vending-Bench 2, a long-horizon business simulation, with $16,843. The model can combine spreadsheets, PDFs, charts, audio and video in one context window and produce structured outputs from them, the company said.
Thai Tran, AI Product Lead of Glean, said: “Gemini 4 Neon helps turn scattered information into clear, grounded answers. It delivers a compelling balance of quality, cost, and speed to day-to-day workplace tasks. In our early evaluations, it matched or exceeded Gemini 4 Carbon on overall quality at 3x the speed and 60% of the cost. We’re excited to bring these capabilities to Glean customers tackling multi-step work that demands accuracy, detail, and speed.”
In scientific and mathematical reasoning, including quantitative analysis and technical research problems, Google rated Neon top of its class with 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond. Brandon Valafar, Strategic AI Lead of Micro1, said: “Across our evaluations, we observed Gemini 4 Neon performing at the top of the leaderboard on our Medical Index and internal life sciences benchmark, while maintaining strong cost-efficiency and meaningful speed.”
Software work is the other focus. Neon can run long, tool-heavy agent jobs alone or act as a fast Orchestrator directing parallel Hydrogen subagents in Antigravity, and Google said it makes fewer redundant file reads, keeps users updated as it goes and compacts its context dependably. It ranks #1 overall on the Artificial Analysis Coding Agent Index (58.6), scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting, and leads SWE-sweep on the mini-SWE harness at 12.8%, a score that more than triples to 39.0% inside Antigravity. In Google's demonstrations, a swarm of Neon agents converted one image into a Google Sheet of 100,000 cells powered by more than 500,000 live formulas, and the model wrote a ski simulator from scratch with 10,200 lines of custom shaders and built interactive 3D recreations of a lab bench and Alan Turing's 1936 study at King's College, Cambridge.
Karina Nguyen, Founder of Thoughtful Lab, said: “Neon is significantly more thorough and careful than other Gemini models we have tested. Neon does well on exploring large code bases and getting an understanding of complex architectures quickly. In our cyber evals, Neon was able to find several subtle vulnerabilities other models missed while using a fraction of the tokens. We think Neon is a very good model for cost-efficient exploration and research tasks.”
Trenton Chang, Chief Scientist of Quadrillion, added: “A novice researcher runs whatever experiment you give it. A great researcher keeps asking why. Gemini 4 Neon is the first model that feels like the latter. It keeps exploring: hunting for new data and imagining new signals to extract to train better models. It's the first model we've seen that pushes itself to do better without a human watching: instead of giving up in the face of bad results, Neon redesigns its modeling strategies to achieve better results.”