AIGenerated
EN · IT · ES

Google launches cheaper Gemini 4 Neon that matches top rivals on coding, law and science

A new lower-cost Google AI model is competitive with pricier leaders on hard tasks, and you can start using it today.

Gemini 4 Neon, by Google.
Picture: from the announcement video, deepmind.google

What you could do with it

  • A business analyst could feed in spreadsheets, PDFs, charts, audio and video together and get back structured deliverables from the mix 1.
  • A developer could write multi-file code, debug it, and refactor entire repositories with surgical edits through Google Antigravity 1.
  • A cybersecurity team could test its own defenses by having Neon generate secure code and probe systems for vulnerabilities at half the cost of larger competing models 1.

What the model does and how fast it scores

Google DeepMind says the new model delivers "frontier-level intelligence" for complex software engineering, knowledge work and science workflows 1. At its xhigh thinking level, it ranks #3 overall on the Artificial Analysis Intelligence Index with a score of 61, placing it just behind larger, slower models . On the independent Vals benchmark list, Neon ranks #2 with a score of 70.9 percent and sits on the so-called Pareto frontier — the group of models that give the best performance for their price .

Neon also ranks #2 on steerability on the new Agent Arena Index, meaning it adapts cleanly when a user corrects it mid-task across 4,200-plus real-world agent sessions, debuting at #2 globally with a score of 1,336 1. Internally, Google says it has its lowest fabrication rate yet in Google Antigravity tests .

Where it leads — with numbers

For enterprise and analytical workloads, Google states that Neon ranks first on both the Legal Research Bench and the Tax Agent Bench, and leads Surge AI categories related to finance, accounting, spreadsheet modeling, and complex constraint following. It scored 73.5 percent on Chartography, a test of reading dense professional charts, and matched leading models on Agents' Last Exam v1.1 with 75.8 percent 1. It also topped Zapier's AutomationBench at 56.3 percent and Andon Labs' Vending-Bench 2, a long-horizon simulation worth $16,843 in reported value .

In science and math, Neon ranked first on Terminal-Bench Science with 70.5 percent, achieved 84.0 percent on RiemannBench and scored 71.3 percent on HLE-Diamond 1. For coding, Google says it is #1 overall on the Artificial Analysis Coding Agent Index with 58.6, reached 65.2 percent on Terminal-Bench 4.1 and 46.2 percent on FrontierCode 2.0 . On SWE-sweep, a multi-file debugging and refactoring test, Neon recorded 12.8 percent on the mini harness and 39.0 percent after running through Google Antigravity .

For cyber defense, Neon reportedly produces more secure code than the earlier Gemini 4 Carbon model across easy, medium and hard challenges on AutoBaxBench, achieving 76 percent on the hard tier, according to Google 1. It also hit state-of-the-art scores on WebExploitBench v1.2 at 51.26 percent and SEC-bench Pro at 90.5 percent, while costing about half as much per challenge as some large rival models .

Concrete demos from the launch

Google shared several live demonstrations built inside Google Antigravity to show what Neon can handle end-to-end. In one showcase, a tool named Blender MCP with /swarm was employed to create a photorealistic 3D lab bench that showed model names beside elemental accuracy, matching Bunsen flames for Gemini 4 Carbon and Hydrogen alongside high-voltage gas discharge tubes for Gemini 4 Argon and Neon. A second project was created entirely from scratch as a first-person WebGL2 freeride ski simulator built on real Mont Blanc and Chamonix terrain. It merged 10,200 lines of custom shaders with 120 Hz ski physics spanning six alpine lines, featuring dynamic carved tracks, slab avalanches, and spatial Web Audio—all without any pre-made assets. A third demo recreated Alan Turing's 1936 King's College Cambridge study in interactive 3D, complete with a working Universal Turing Machine on his desk 1. Google also demonstrated a Neon agent swarm turning a single image into a 100,000-cell Google Sheet using over 500,000 live formulas to perform edge detection, map 3D depth and relight the scene across six modes.

Price and access

The pricing Google announced is $3 per million input tokens, with a 95 percent discount bringing cached input tokens down to $0.15 per million, and $15 per million output tokens 1. Neon supports a 1 million-token context window and offers four steerable thinking levels that let teams balance intelligence against token spend . Google says that at its default high thinking level, Neon delivers a 32 percent lower cost per task than Gemini 4 Carbon, and sits on the Pareto frontier with roughly 50 percent to 80 percent lower cost per task than top competing frontier models; at xhigh thinking, it still costs 36 percent or more per task less than competitors at their highest setting .

Neon became available starting November 18, 2026. It can be accessed via the Gemini API within Google AI Studio, through Gemini for business, and by paid Google AI Pro and Ultra subscribers using Google Antigravity or the Gemini app.

Key facts

  • Price: $3 per million input tokens and $15 per million output tokens
  • Ranks #2 on Vals Index and #3 overall on the Artificial Analysis Intelligence Index (xhigh thinking)
  • 1 million-token context window with four steerable thinking levels
  • Lifted to 39.0 percent on SWE-sweep when run in Google Antigravity
  • Available from November 18, 2026 via API, AI Studio and Gemini apps

Sources

  1. Introducing Gemini 4 Neon

If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work

Keep reading