Google DeepMind launched Gemini 4 Neon on 18 November 2026, adding a model for software engineering, knowledge work and science to its Gemini 4 line. The company says Neon is rolling out through the Gemini API in Google AI Studio, Gemini for business, Google Antigravity for paid Google AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra subscribers.
Key points
- Neon costs $3 per million input tokens and $15 per million output tokens, with a 1 million-token context window.
- Google reports a score of 61 and third place on the Artificial Analysis Intelligence Index at Neon’s xhigh thinking level.
- Crypto Briefing reports estimates of roughly 17% lower cost per task than Gemini 4 Carbon, or approximately 32% at the default high thinking level.
Neon’s $3 input and $15 output rates
Google prices Neon at $3 per million input tokens and $15 per million output tokens. Cached input costs $0.15 per million tokens, a 95% discount on the standard input rate. Those prices give developers separate rates for material sent to the model and material it generates, rather than a fixed charge for each completed task.
The model has a 1 million-token context window: the space available for material supplied to it in one request. Google says Neon can work across spreadsheets, PDFs, charts, audio and video within that window and produce structured outputs from them. The window describes how much material can be put before the model, rather than a score for the quality of its answer.
Reviewing paperwork could bring a long PDF, its charts and an audio recording into one request, provided the material fits the stated window. The requested output could be a structured account of those inputs, the kind of work Google says Neon performs.
Crypto Briefing reported that a post on X estimated Neon’s cost per task at about 17% below Gemini 4 Carbon’s, rising to approximately 32% below Carbon’s at the default high thinking level. Thinking level governs how much reasoning the model does before answering, so the setting affects the amount of work billed for a task.
Neon is the fourth Gemini 4 model, after Argon, Carbon and Hydrogen. AI Affairs previously reported Carbon’s wider release and a preview of agent swarms, then Hydrogen’s launch and model switching in Antigravity. Google is presenting Neon as an option for complex work in which both the result and the cost of repeated calls count.
Neon scores 61 on Artificial Analysis
Google says Neon scored 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, placing third overall behind models it describes as larger and slower. That is a result at a specified reasoning setting, whereas the reported per-task saving of approximately 32% against Carbon uses Neon’s default high setting.
On the Vals Index, Google reports a score of 70.9% and second place. It also reports a score of 1336 and second place on the Agent Arena Index, based on more than 4,200 real-world agent sessions. Google says Neon ranked second there for steerability, which it describes as adapting to corrections made partway through a task. The session count belongs to that agent evaluation, rather than to the other rankings.
Google reports first-place results on Legal Research Bench and Tax Agent Bench. In scientific and mathematical testing, it reports 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond.
Antigravity lifts Neon’s SWE-sweep score to 39.0%
For coding, Google says Neon can work through long tasks involving tools and multiple files, either on its own or directing parallel Hydrogen subagents in Google Antigravity. Its description of the model’s behaviour includes fewer repeated file reads, progress updates during a task and the ability to compress earlier context in long sessions.
Neon scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting. On SWE-sweep, it reports 12.8% using the mini-SWE harness and 39.0% in Google Antigravity. A harness is the software environment in which a model attempts the test, so the Antigravity result includes the way that environment lets Neon carry out the work. Google also places Neon first on the Artificial Analysis Coding Agent Index, with a score of 58.6.
Google says Neon can make targeted changes across repositories for debugging and refactoring. It also reports its lowest fabrication rate yet in internal Google Antigravity testing, an assessment conducted within the company rather than one of the named external coding evaluations.