Google has released Gemini 4 Neon inside its Google Antigravity coding tool. Neon is meant to sit between the more powerful Gemini 4 Carbon and the faster Gemini 4 Hydrogen, and it is aimed at everyday coding and agent work.
Gemini 4 Neon is now available inside Google Antigravity.
Neon is built for coding tasks like editing repositories, debugging, and refactoring, and it reportedly makes fewer made-up answers than earlier models in Google's internal tests.
The new Orchestrator feature lets Neon plan a task while Hydrogen handles routine steps, so users don't have to switch models by hand.
Access is for paid Google AI Pro and Google AI Ultra subscribers in the desktop app, CLI, and IDE extensions, and enterprises can use it through Gemini for business.
Google Research reported results from Antigravity's /teamwork, a system where multiple AI agents work together, on long-open problems in math, chemistry, systems software, and machine learning engineering. The work was done with human experts, and the tools will be released to outside users.
Using Gemini 4 models and /teamwork, teams produced new results in information theory, formal verification, molecular structures, and kernel optimization.
/teamwork runs several agents in parallel and automatically checks each answer with tools such as proof checkers and test suites.
Each agent receives only the files and tools it needs, which helps keep details intact over multi-day tasks.
Some features are already available in Antigravity, and more will reach outside users in the coming weeks.
Google has added the faster Gemini 4 Hydrogen model to its Antigravity coding tool. It also introduced Orchestrator, an option that uses Hydrogen together with the larger Gemini 4 Carbon model.
Orchestrator switches between the two models during a task, and Carbon revises the plan when a task gets hard.
Google says Orchestrator reached near-Carbon success rates while using 24% less quota and finishing 19% faster.
Hydrogen can also be used alone, with four thinking levels from low to the highest setting.
Developers can use the Antigravity SDK to assign each model a role in an automated system.
Google has released Gemini 4 Carbon, its strongest AI model, and a preview of /swarm, a feature where several AI agents work as a team inside Antigravity. The agents plan, build, and verify tasks at the same time.
/swarm has a coordinator agent that starts several worker agents to split work in parallel.
Each worker runs in its own separate workspace, so their file edits do not collide.
A verifier agent runs tests and checks the changes.
A live screen shows the agents' messages and work flow in real time.
Google Antigravity, its AI work tool, now has a new screen mode that is easier for non-developers to use. Users can switch between it and the original developer screen. Both use the same AI agent, which can handle multi-step tasks on its own, so the features stay the same.
More people now use it for knowledge work such as research, data analysis, documents, and presentations.
The Knowledge Work screen hides the Git workspace setup prompts and code diff views.
You can switch in the Interface Layout settings, and the change applies without a restart.
When you first sign in and pick your job, the matching screen and suggested features are set up automatically.
Google has added a voice mode to its Antigravity AI coding tool. Users can give tasks by speaking and change direction while the agent is working. The voice mode runs on Gemini Live, Google's real-time voice AI model.
You can turn on voice mode in the app and talk with the agent like on a call.
While a task runs, you can say "stop" to halt it or change its direction.
You can approve or deny permission requests by voice.
Risky actions such as deleting files cannot be approved by voice and need a click on screen.
Google has opened a Plugin Marketplace in its AI coding tool Antigravity. Users can browse plugins, bundles that add features, made by Google and outside companies in the Antigravity app and CLI, a command-line tool, and install them with a single click.
Until now, MCP connections (a standard way to link AI to outside tools), environment variables, skills, and subagents that take on parts of the work had to be set up by hand in each repository.
The marketplace packages these settings into verified bundles that can be browsed by category and installed right away.
Plugins from partners in fields such as development, design, productivity, and security launched alongside it, including Figma, Canva, Notion, Vercel, and Stripe.
Each plugin is defined by a plugin.json file and bundles up to five components that install, update, and switch on or off together.
Google has added its newest AI model, Gemini 4 Argon, to its AI coding tool Google Antigravity. Google AI Pro and Ultra subscribers can use it in the Antigravity desktop app and CLI, a command-line tool, while businesses can reach it through the Gemini platform API, a gateway for connecting programs, or eligible Gemini Enterprise plans.
Gemini 4 Argon aims for top-tier results in actual software development, business tasks such as legal and finance work, and cyber defense.
Across Alphabet, Google's parent company, more than 100,000 people use Argon each month in an internal version of Antigravity.
Over four weeks, code written by AI agents rose 35%, while changes rolled back after submission fell 30%.
It halved the CPU use of YouTube's comments service and cleared 97 stale or duplicate reports from a queue of 134 bugs in under three hours.
Teams across Google have started adding custom agents, AI helpers built for specific jobs, to the official Antigravity plugins offered through Build with Google. Three agents were presented, covering accessibility checks that help everyone use an app, security rules, and pre-release app reviews.
The Flutter accessibility agent spots missing semantic labels for screen readers, touch targets smaller than 48x48 dp, and color contrast problems, then fixes the code.
The Firebase Security Rules agent writes Firestore security rules from your data structure and app needs, and catches overly loose read and write access ahead of time.
The Firebase agent comes inside the official Firebase plugin and can be switched on from settings.
Developers can also build a read-only agent with Android skills that checks Google Play submission requirements and run it from the terminal.
Google has made Gemini 3.8 Flash available in its AI coding tool Google Antigravity starting today. It is the third Flash model in six weeks and is designed as an everyday workhorse.
Pricing stays at the usual Flash level of $0.75 per million input tokens, the units AI uses to process text, and $3.75 per million output tokens.
It does better than 3.7 Flash on coding work done with AI agents and often comes close to pricier top-tier models.
Users can set the thinking level for each task to balance the depth of reasoning, how the AI works out an answer, against response speed.
3.7 Flash can still be used for jobs where saving compute is the priority.
Google has updated Teamwork in Antigravity. In Teamwork, several AI agents suggest ideas, critique them, and improve one another's work on their own for hours or even days, and it can be used now as /teamwork-preview on all paid plans. The update builds on the launch of Gemini 3.7 Flash.
It solved seven open problems in math and theoretical computer science, and Knuth's Cycles Conjecture was confirmed with the Lean proof checker.
It scored 71% on TCSBench, a theoretical computer science benchmark, the highest result in Google's internal testing.
It built a RISC-V CPU simulator that boots an operating system from scratch and contributed speedups to open-source libraries such as Eigen.
Gemini analyzes each request and automatically picks a fitting team setup, called a pattern.
Google introduced Generative UI, which lets AI build screen elements on the spot, along with visual tools in its AI coding tool Google Antigravity. AI agents can now show complex structures through graphics, dashboards, and hands-on interactive screens instead of text alone.
Agents can build visual elements, live dashboards, hands-on results, and design drafts, then show and tweak them right away.
In one example, a project's dependencies, the other libraries its code relies on, appeared as a graph whose nodes can be dragged around and clicked.
Google also showed a quadtree explorer, a structure that splits space into four parts, where users place points, plus a solar system simulation of planetary orbits.
Google says visuals help people understand data and structures that text alone struggles to convey.
Google has added Git version control tools, which track code change history, and a terminal tab, a window for typing commands, to the right-side panel of its AI coding tool Antigravity 2.0. Users can now review changes and commit, or save, them in one place without switching apps.
The side panel now shows not only edits by AI agents but also changes made by scripts or other editors, measured against the current branch, or line of work.
Users can switch between agent edits, uncommitted changes, and full branch changes, and pick how diffs are displayed.
Files can be staged, meaning picked for the next commit, and committed with a message straight from the panel.
A built-in terminal tab runs commands like git status without leaving the app.