Хочете будувати кар’єру в Data Science та AI? Vodafone Big Data Lab відкриває новий набір —>
Nataliia KhandusenkoAI Eng
19 November 2026, 12:02
2026-11-19
Google Gemini 4 Neon triples its coding test results thanks to Antigravity environment — without changing the model itself
Google's new Gemini 4 Neon AI model solved 12.8% of the problems in the SWE-sweep programming benchmark when run in the base mini-SWE shell, but its performance increased to 39.0% when the same model was run inside the Antigravity environment.
Google's new Gemini 4 Neon AI model solved 12.8% of the problems in the SWE-sweep programming benchmark when run in the base mini-SWE shell, but its performance increased to 39.0% when the same model was run inside the Antigravity environment.
The more than threefold performance boost that Google announced Wednesday during the Neon release demonstrates how the effectiveness of AI agents for development depends not only on the underlying neural network, but also on the software framework that manages its tools, context, and code execution, writes The New Stack. A day earlier, research by FrontierHarness also showed that running the same open model can cost five times more depending on the chosen framework.
One model — different results
In the Gemini 4 Neon datasheet, the company compared three leading models in the SWE-sweep test, running them in the base mini-SWE environment and in the proprietary shells of the respective developers:
Gemini 4 Neon — growth from 12.8% (in mini-SWE) to 39.0% (in Antigravity);
Claude Opus 5.5 — a jump from 3.6% (in mini-SWE) to 10.1% (in Claude Code);
GPT-6.2 Sol: reverse dynamics - drop from 11.6% (in mini-SWE) to 8.3% (in Codex).
Google tested each model in two configurations, although the comparison does not specify what changes caused the difference, or whether these results will hold up in other benchmarks and repositories.
How Antigravity works and role distribution with Hydrogen
The Antigravity environment was trained in conjunction with the Gemini models. Neon itself also received updates that make it do fewer redundant file reads and better manage context during long programming sessions. Antigravity is also able to run subagents in separate Git working trees and check in their changes. However, Google has not tested each of these improvements individually or compared Neon with third-party shells other than mini-SWE.
Neon also joins the model switching system introduced with Gemini 4 Hydrogen:
Neon is responsible for planning and the most complex programming tasks;
Hydrogen takes over the routine work — Neon can run several such subagents in parallel using the /swarm function, introduced in the Carbon model.
The company did not specify how exactly such a connection affects the final SWE-sweep score and the final cost of completing the task.
Other benchmarks, price and availability
According to Google, at a standard high level of reasoning, the Neon model scored 65.2% in the Terminal-Bench 4.1 test and 46.2% in FrontierCode 2.0, and also topped the Artificial Analysis Coding Agent Index with a score of 58.6.
Google says that running a single task using Neon is 32% cheaper than using Gemini 4 Carbon (according to the Artificial Analysis Intelligence Index). The cost of API access is $3 per million input tokens, $15 per million output tokens, and $0.15 per million cached input tokens for a context window of 1 million tokens.
The model is already available through the Gemini API, Google AI Studio, and Gemini for Business, and Google AI Pro and Ultra subscribers can work with it in the Antigravity environment.