Google Releases Gemini 4 Neon: 65.2% on Terminal-Bench 4.1 at $3 per Million Tokens
Google has released Gemini 4 Neon, a model that sits between the lightweight Hydrogen and the heavyweight Carbon. On Terminal-Bench 4.1, which tests whether a model can solve engineering tasks in a terminal, it scored 65.2%. A million input tokens cost $3, and a million output tokens cost $15. On SWE-sweep, the results depended on the environment: 12.8% with mini-SWE and 39% in Antigravity. That is why you should compare Neon with your current model inside the same harness.
AI-processed from TG: @neurogen_news (Neurogen); edited by Hamidun News
Google has added a Neon model to the Gemini 4 lineup. Until now the choice was simple. Hydrogen is a fast, cheap model for everyday tasks, something like a flash drive you always carry with you.
Carbon is a heavyweight model for large projects. Neon sits in the middle. Google says it handles complex code, agents, and analytics, but it is built with a focus on speed and cost.
The key numbers are these. On Terminal-Bench 4.1, a test where a model solves engineering tasks directly in the command line, the way a programming agent would, Neon scored 65.
2%. Through the API, Google charges $3 per million input tokens (a token is the unit of text used for billing) and $15 per million output tokens, including the model's reasoning. Below we look at what sits behind these numbers and how to check them on your own tasks.
What Neon Does and What It Costs
The model works with text, images, audio, video, and documents. Its context window is 1,048,576 tokens, and this is a shared limit: your request, the model's internal reasoning, and the final answer all count against it. If thinking mode is enabled, reasoning takes up part of the window, which matters when you work with very long documents.
Pricing:
- input: $3 per million tokens;
- output: $15 per million tokens, with reasoning billed as output;
- cached input: $0.15 per million tokens.
The last line matters more than it seems. Agents almost always reprocess the same long context: a project description, instructions, and the history of previous steps. If this repeated block is served from the cache, it costs $0.15 instead of $3, which is twenty times cheaper. In long agentic sessions, caching often determines the final bill.
Neon is available on Google AI Pro and Ultra subscriptions. In the Antigravity environment it can act as an orchestrator: it plans the work and hands out tasks to several agents running on Hydrogen. The logic is clear. The stronger model thinks and checks, while the cheaper executors handle routine work. Google also says the model will come to NeuroAPI soon.
What the Benchmarks Show
Besides the 65.2% on Terminal-Bench 4.1, there is an interesting result on SWE-sweep. It depended heavily on the environment the model ran in: 12.8% with mini-SWE and 39% in Antigravity. That is roughly a threefold difference for the same model. This is a useful lesson for anyone choosing a model from tables: the outcome depends on both the model itself and the harness around it, including which tools it can use, how retries work, and how the context is assembled.
In the KingBench 4 test from MindStudio, Neon took second place out of 18 with 37.9 points out of 60. The leader, Claude Fable 5.2, scored 42 points. A caveat: the models were tested in different agentic environments, so the comparison cannot be treated as direct. The 4.1-point gap may reflect a real advantage for Claude, or it may come from different launch conditions.
What to Do With This in Practice
First, do not treat these numbers as a final verdict. The data comes mostly from Google and from third-party tests run in different environments, and there are still few independent comparisons on an identical harness. The most reliable check is to take 20 to 30 of your typical tasks (bug fixes, refactoring, log analysis) and run them through Neon and your current model in the same agentic framework.
Second, count the cost of a session, not of a single token. At $15 per million output tokens, long reasoning chains can get expensive. The $0.15 cached rate will smooth the bill if you structure prompts so the unchanging part comes first.
Third, if you already work in Antigravity, try the pairing of a Neon orchestrator with Hydrogen executors. It is the most direct way to find a balance between price and quality within a single ecosystem.
Bottom line: Neon does not aim for a record. Its idea is adequate quality at moderate cost. It covers code and agents more cheaply than the heavyweight models, and second place in KingBench 4 suggests it stays close to the leader on quality. How far that will go for your project only your own tests will show.
- Nov 21, 2026Apple Is Building an AI Agent and Making It Harder for Outsiders to Get Access: Permission Now Takes 14 Steps, Not 4
- Nov 21, 2026Gemini 4 Neon shrank the TS Rust binary from 47.44 to 7.97 MB in under 10 hours
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms
Key Questions Answered
What Neon Does and What It Costs?
The model works with text, images, audio, video, and documents. Its context window is 1,048,576 tokens, and this is a shared limit: your request, the model's internal reasoning, and the final answer all count against it. If thinking mode is enabled, reasoning takes up part of the window, which matters when you work with very long documents. Pricing: input: $3 per million tokens; output: $15 per million tokens, with reasoning billed as output; cached input: $0.15 per million tokens.
What the Benchmarks Show?
Besides the 65.2% on Terminal-Bench 4.1, there is an interesting result on SWE-sweep. It depended heavily on the environment the model ran in: 12.8% with mini-SWE and 39% in Antigravity. That is roughly a threefold difference for the same model. This is a useful lesson for anyone choosing a model from tables: the outcome depends on both the model itself and the harness around it, including which tools it can use, how retries work, and how the context is assembled.
What to Do With This in Practice?
First, do not treat these numbers as a final verdict. The data comes mostly from Google and from third-party tests run in different environments, and there are still few independent comparisons on an identical harness. The most reliable check is to take 20 to 30 of your typical tasks (bug fixes, refactoring, log analysis) and run them through Neon and your current model in the same agentic framework. Second, count the cost of a session, not of a single token.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.