Gemini 4 Argon opens up: Google's capability deep-dive, claim by claim

Gemini 4 Argon, which launched on September 30 behind a vetted cyber-defender rollout, is now available globally: through the Gemini API in AI Studio and the Gemini platform for developers and enterprises, in Google Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode in Search for Ultra and Pro subscribers. The same day, Google published a long capability deep-dive titled "Built to solve your hardest problems," packed with deployment results. Our September 30 post covered the launch mechanics (pricing, the 1M token limit, the launch benchmark table). This one is mostly receipts, and almost all of them are Google's own numbers.

Engineering results from inside Google

The coding results are the most concrete part of the post. Google reports that Argon in Antigravity helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third. On Chrome, Argon helped separate roughly 6 million lines of legacy browser-feature code from Chromium's 40 million-line codebase, finishing a refactor Google says would have taken months in weeks; for users, the payoff is live webpages that can move between tabs and side panels without reloading.

Other in-house examples carry links, which matters. An overnight autonomous pass on the open-source Agent Development Kit cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%, with the changes live in google/adk-python. Researchers made the widely cited xTB quantum chemistry package up to 2.3x faster at predicting molecular structures, and the changes went to the official xTB repository on GitHub. And Google reports a 2.85x end-to-end inference speedup on Gemma via vLLM on an A100 (from about 37,900 to over 108,000 tokens per second on Gemma 4 E2B) with 100% token parity, done in days where the team says it would normally take months.

Business and science workloads

The post leans hard on multimodal document work. One demo synthesizes roughly 50 GB of airborne LiDAR scans, GIS room vectors, Library of Congress CAD drawings, and NPS archives into 3D models of Chaco Canyon. At Isomorphic Labs, Argon synthesizes complex PDFs and patents to extract pertinent information, reducing a process Google says typically takes weeks to hours.

The science section has four items worth noting on their own terms: Argon translated the internal AlphaQubit quantum error-correction decoder into a single optimized hardware kernel, nearly doubling its inference throughput; an Argon-written implementation of the arcsin quantum subroutine, formally verified in Lean, cuts spacetime cost by more than 100x against a published CORDIC baseline; a team with Professor Matt Landreman found a broader class of counterexamples to Grad's conjecture, a 1967-era hypothesis about stellarator magnetic fields, giving the fusion community new benchmarks; and Gamow Labs researchers used Argon on de-identified newborn genomes to pinpoint a chromosomal rearrangement routine software had missed, with Argon scoring 57% on RareBench and 52.2% on LabBench in the lab's private evals. Google is also giving 10,000 scientists six months free at Antigravity's highest usage tier, applications open.

Cyber defense and robotics

The cyber side: the Fairwind Program, through which defenders get Argon without cyber guardrails, is expanding to hundreds of additional partners. Wiz's Scan for Good program has used the model to remediate a gap that risked exposing sensitive government data, and to find and fix a public file-download bypass in healthcare infrastructure that exposed a master key, protecting hundreds of thousands of patient records.

On robotics, Google reports an 87% score on the ERQA embodied-reasoning benchmark, ahead of every other frontier model, and a demonstration where Argon watched a raw video of a human playing cornhole, mapped the spatial coordinates, built the simulation code, and then the Apptronik Apollo humanoid sank seven consecutive shots in a real-world trial. Google calls the Video-to-Sim-to-Real pipeline a way to compress weeks of manual simulation modeling into hours.

What to verify, and the price after the intro period

The safety section adds specifics: external bio red-teaming unlocked under 1% of the tested bio sub-tasks, violations on an adversarial set fell roughly in half versus previous Gemini models, shortcut attempts on coding benchmarks dropped to 2%, an automated misalignment monitor reads the agent's reasoning to block unintended workarounds in real time, and Argon leads Gray Swan's indirect prompt injection benchmark at 0.7% attack success across 15 adaptive attempts.

The practical read for your planning: access was the bottleneck on Argon, and it is no longer. The introductory pricing holds until the period ends on November 30, 2026, after which Google says the price moves to $4 per million input tokens and $20 per million output. Every number above comes from Google's own post, some with links to upstream pull requests that you can audit. The launch table we covered in September and this deep-dive both belong on your "re-check on your own workloads" list, not your settled rankings list.