Google Launches Gemini 4 Argon for Enterprise and Developer Workloads
Google's new frontier model arrives with a temporary introductory API price, long-horizon coding capabilities, and new benchmarks in physical reasoning.
In short
- Google released Gemini 4 Argon on October 13, 2026, making it accessible through the Gemini API, Google Cloud, AI Studio, Antigravity, and subscription tiers.
- The model scored 87% on the Embodied Reasoning benchmark and demonstrated autonomous real-to-sim robotics pipelines as well as overnight repository refactoring.
- Introductory API pricing is set at $2 per million input tokens and $10 per million output tokens through November 30, 2026, before doubling on December 1, 2026.
- Third-party evaluations show strengths in domain generalization and legal reasoning, though composite indexes place it behind rivals Claude Opus 5.5 and Claude Sonnet 5.5.
On October 13, 2026, Google announced the broad availability of its flagship frontier model, Gemini 4 Argon, opening access to enterprise clients, software engineers, and subscribers. The release follows an initial unveiling on September 30, 2026, during which access was restricted to early testers and vetted defensive cybersecurity specialists. According to an announcement from Google, the model is designed to tackle sustained, multi-turn workloads across software development, scientific research, robotics, and corporate analysis.
Under model identifier gemini-4-argon, the system is available across Google AI Studio, Google Cloud, and the Antigravity development environment, as well as the Gemini mobile and desktop applications for Google AI Pro and Ultra subscribers. Google also introduced an introductory API price of $2 per million input tokens and $10 per million output tokens through November 30, 2026, with cached inputs discounted by 95 percent, as reported by VentureBeat. Standard rates of $4 per million input tokens and $20 per million output tokens will take effect on December 1, 2026.
Autonomous Codebase Optimization and Engineering Tasks
Google positioned Gemini 4 Argon as an agentic system capable of reasoning across sprawling, multi-file codebases over long execution horizons. In an engineering report published on the Google for Developers Blog, Google Cloud engineer Bo Yang detailed an 11-hour unattended overnight experiment on the open-source Agent Development Kit (google/adk-python). The repository covers roughly 205,000 lines of Python across nearly 800 files and approximately 17,000 unit tests.
Operating in Antigravity's /goal mode and guided by maintainer guidelines placed in .agents/skills, Argon identified six optimization proposals across 32 files. After maintainer code review, two of the optimizations were accepted into the repository's main branch. The accepted changes cut in-memory session preparation overhead by about 40 percent across long sessions and accelerated cold package imports by 21 percent, while preserving test suites and API compatibility. Google noted that maintainers rejected four other proposals because they relied on fragile internal library attributes or synthetic gains.
Beyond framework tuning, Google reported internal productivity metrics from more than 100,000 employees using Argon inside an internal deployment of Antigravity. According to Google Cloud, engineering teams recorded a 35 percent rise in agentic code submissions alongside a nearly one-third drop in rollbacks over a four-week period. The Chrome team used Argon to decouple browser features from core webpage logic inside Chromium's 40-million-line codebase, an effort Google said compressed several months of manual refactoring into weeks.
Physical Reasoning and Robotics Pipelines
Google demonstrated Argon's visual and spatial reasoning across several physical engineering tasks. On the Embodied Reasoning (ERQA) benchmark, which tests a system's ability to solve spatial and physical problems in 3D settings, Argon scored 87 percent, according to the Google for Developers Robotics post.
The robotics team showcased video-to-simulation-to-real workflows. In one trial, Argon observed a video of a person throwing cornhole bags, mapped 3D scene coordinates, computed whole-body joint positions, and generated physics simulation code that enabled an Apptronik Apollo humanoid robot to land seven consecutive throws. In another test, Argon controlled a Boston Dynamics Spot quadruped equipped with an arm, generating an articulated MuJoCo digital twin of a foot-pedal trash can to rehearse stepping motions and balance before performing can disposal on physical hardware.
Argon also generated physics simulations directly from pen-and-paper drawings. In a demonstration documented by Google, the model took a photo of a hand-drawn sketch of a humanoid throwing a ball at a pyramid of boxes, built a GPU-accelerated mujoco_warp environment, planned a 76-degree-of-freedom trajectory, verified knocking down all six boxes, and produced a raytraced render via Blender Cycles.
On the open-source side, Google published gemini-4-demo-vanilla-3D on GitHub, an asset-free repository written by Argon that includes both a NumPy-based raytracer and a vanilla WebGL2 browser engine that resizes an architectural room in real time.
Scientific Computing and Security Findings
Google highlighted applications across scientific research and cyber defense on its Gemini model portal. In quantum computing, Argon translated an internal version of Google's AlphaQubit quantum error correction decoder into a dedicated kernel that nearly doubled inference throughput, while synthesizing an arcsin quantum subroutine verified in Lean that reduced spacetime cost by more than 100 times over a quantum CORDIC baseline. In plasma physics, Google worked with University of Maryland professor Matt Landreman to discover a broader class of counterexamples to Grad's conjecture regarding stellarator magnetic fields.
In biochemistry, Argon helped researchers design five serine esterase enzymes whose catalytic activity was confirmed in wet-lab tests. For clinical research, Gamow Labs reported that Argon pinpointed a previously undetected chromosomal rearrangement in newborn genomes with rare lung conditions, scoring 57 percent on RareBench and 52.2 percent on LabBench.
In cybersecurity, cloud security provider Wiz used early access to Argon in its "Scan for Good" program to discover and remediate a healthcare software vulnerability that exposed master keys and administrative access, protecting hundreds of thousands of records. On public benchmarks, Argon tied for first place on CWE-Bench v1 with a 68 percent vulnerability remediation score, and recorded a 0.7 percent success rate against 15 adaptive prompt injection attempts on the Gray Swan benchmark.
Benchmark Comparisons and Availability
Argon led several domain-specific evaluations. On the GDP-weighted Vals Index, which tracks performance on finance, tax, legal, and programming tasks, Argon scored 68.90 percent, edging out Claude Sonnet 5.5 (67.04 percent) and Claude Opus 5.5 (66.97 percent), according to VentureBeat. On Harvey's Legal Agent Benchmark, Argon registered 19.6 percent, versus 5.4 percent for OpenAI GPT-6 Astra and 3.8 percent for Claude Opus 5.5. On DeepSWE v1.1, Argon reached 77.9 percent.
However, on general composite rankings, independent measurements diverged from Google's claims. On the Artificial Analysis Intelligence Index as of October 13, 2026, Claude Opus 5.5 led with 58, followed by Claude Sonnet 5.5 and Meta Muse Wonder Full at 56, while Gemini 4 Argon tied with GPT-6 Astra and Claude Fable 5.1 at 53. Sam Witteveen, CEO of Red Dragon AI, noted in a VentureBeat review that while Argon demonstrated strong generalization and impressive visual-to-code capabilities, Antigravity still lagged behind developer tools like Claude Code or Codex.
Gemini 4 Argon is available globally for developers through Google AI Studio and the paid Gemini API, with specialized access granted to cybersecurity defenders through Google's Fairwind Program. Google is also offering 10,000 scientific researchers six months of access to Google Antigravity's highest tier through an application program.
Questions people ask
What is Google Gemini 4 Argon?
Gemini 4 Argon is Google's flagship frontier AI model released on October 13, 2026, focused on long-horizon reasoning, agentic coding, scientific research, and physical robotics tasks.
How much does the Gemini 4 Argon API cost?
Google offers an introductory price of $2 per million input tokens and $10 per million output tokens through November 30, 2026, with cached inputs priced at a 95% discount. Starting December 1, 2026, rates rise to $4 per million input tokens and $20 per million output tokens.
Where can users access Gemini 4 Argon?
The model is available via the Gemini API, Google AI Studio, Google Antigravity, Google Cloud, and Android Studio, as well as the Gemini app and Search AI Mode for Google AI Pro and Ultra subscribers.
How did Gemini 4 Argon perform on robotics and coding tests?
Argon achieved an 87% score on the Embodied Reasoning benchmark, generated physics simulation code for humanoid and quadruped robots, and completed an 11-hour unattended run on Google's ADK repository that cut session preparation latency by about 40%.
Sources
- Gemini 4 Argon for Roboticsdevelopers.googleblog.com, Oct 13, 2026, official source
- Optimizing a 200,000-line Python codebase overnight with Gemini 4 Argondevelopers.googleblog.com, Oct 13, 2026, official source
- google-gemini/gemini-4-demo-vanilla-3DGitHub, Oct 12, 2026, official source
- GeminiGoogle DeepMind, official source
- Gemini 4 Argon: Built to solve your hardest problemsGoogle, Oct 13, 2026, official source
- Gemini 4 Argon available on Google CloudGoogle Cloud, Oct 13, 2026, official source
- Google makes Gemini 4 Argon broadly available with limited-time low API pricing of $2/$10 per 1M tokenventurebeat.com, Oct 13, 2026
How this story was made: the newsroom picked it up from developers.googleblog.com, GitHub and theverge.com, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (40 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 14, 2026 at 00:07 UTC
