Trust, but verify

Gemini 4 Argon cuts Google's post-submit rollbacks 30%

Google is putting its new Gemini 4 Argon model to work inside Antigravity, and the internal results go well beyond writing code. From untangling decades-old research code to hunting security holes that other frontier models missed, Argon is being tested on the messy, unglamorous work engineers actually face every day.

Oct 14, 2026 | Felipe Santos | Software

Google is rolling out Gemini 4 Argon inside Google Antigravity, its agent-first development app, for Google AI Ultra subscribers. Argon was introduced earlier this month as Google's latest model, with frontier performance in complex workflows across real-world software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. The Antigravity rollout puts it directly into the desktop app and the Antigravity CLI, where agents can work through multi-step engineering problems.

Google says it built Antigravity for an agent-first future, with foundational primitives meant to scale alongside model intelligence. Argon is the first model that Google says brings that experience fully to life in the app. The company also says future models may serve as everyday drivers for a broader range of builders, while Argon is aimed at the hardest problems.

Google's internal use provides the main evidence. Over 100K users across Alphabet use Argon inside an internal version of Antigravity each month to solve complex production problems. Over a four-week period, Google reports a 35% increase in agentic code, along with a 30% decrease in post-submit rollbacks, which it describes as an improvement in code quality as well as developer velocity.

The article describes several internal projects. In one, Argon helped the YouTube team cut the CPU footprint of the comments frontend in half. By tracing production profile data, it found that redundant emoji rendering alone accounted for nearly 45% of server overhead, which led to 12 targeted optimizations. These reduced server comment-load latency by 30 ms and client end-to-end latency by 20 ms. In another, Argon reviewed Wing's proprietary drone physics simulator, reasoning through the physics and math rather than only the code structure. It proposed 18 targeted modelling improvements across aerodynamics, propulsion, and structural dynamics.

Google also describes routine work. Argon took an aging queue of 134 open bug reports, cross-referenced commit histories, checked whether fixes reached production, and inspected live telemetry. In under 3 hours it closed 97 stale or duplicate bugs with documented commit proof, a 72.4% queue reduction, and sorted the remaining 37 by owner, leaving no unassigned tickets. In a legacy migration using a custom skill, a team had Argon map and execute a complete rewrite of a Fortran codebase into modern Rust, delivering a 1.8x performance boost and memory safety. In a security audit of access-control changes, Argon traced how permissions flowed at runtime and found four critical security bugs. Google notes that two earlier frontier models had audited the identical code and both concluded there were zero security holes. For production support, Argon batch-processes on-call tickets, sorts them by issue type, traces root causes, proposes fixes, and drafts replies for human approval, freeing 2-3 hours per shift.

Usage is spreading beyond engineering. Google says monthly active usage outside core engineering is soaring, with over half of Cloud sales, almost half of finance, and over a third of business and operations Googlers using Antigravity. Plugins, skills, and MCP integrations connect the agent to specialized data sources and workflows. Examples include localizing hundreds of pages of developer documentation into 17 languages in a matter of hours, resolving a P0 data center dashboard outage that had stalled for over four months, and cutting research time by 90% for an Ads account executive, saving upwards of 10 hours of manual work each week. A finance lead ran end-to-end reconciliations, budget tracking, contract extraction, and daily executive briefings across 80 workflows. Over seven business days, 15.5 hours of human direction produced about 335 human-equivalent hours of validated output, saving more than eight weeks of manual work.

Key Features at a Glance

Agent-first workflows: Argon runs agents that work through multi-step problems in the Antigravity desktop app and CLI, so engineers can hand off complex tasks rather than individual edits.
Runtime-aware auditing: Argon traced permissions as they flowed at runtime, not just in the code diff, which surfaced four critical security bugs that earlier models missed.
Backlog cleanup: 97 stale or duplicate bugs closed in under 3 hours, a 72.4% queue reduction, with commit proof for each.
Legacy modernization: A complete Fortran-to-Rust rewrite delivered a 1.8x performance boost and memory safety, giving scientists a foundation for future work.
Extensibility: Plugins, skills, and MCP integrations connect Argon to specialized data and workflows for non-engineering teams.
Measured code quality: A 35% increase in agentic code over four weeks came with a 30% decrease in post-submit rollbacks.

Availability

Argon is rolling out starting in the Google Antigravity desktop app and CLI for Google AI Ultra subscribers and eligible Gemini Enterprise subscriptions. It is also available via the API in Gemini platform. Google recommends updating to the latest version of Antigravity to try it on the most complex workflows. The article does not state pricing.

Source: antigravity.google