Google announced on October 13 that Gemini 4 Argon is now available worldwide. Developers can call it through AI Studio and the paid tier of the Gemini API under the model name gemini-4-argon. In the Gemini app and AI Mode in Search, it is open to Google AI Pro and Ultra subscribers, while the coding tool Antigravity remains limited to Ultra subscribers.
Two weeks ago, when the model first launched, it went only to the cybersecurity defense teams in the Fairwind program. This time Google has widened access in one step to developers, enterprises, researchers and everyday subscribers. The security track stays in place, with applications still approved on a rolling basis for a version without the cybersecurity guardrails.
From security teams to every paying user
Strung together, Argon's milestones show Google going narrow first, then wide. At the September 30 debut, the company defined three areas of work: software engineering, enterprise knowledge work such as legal and finance, and cyber defense. The results published at the time centered on coding and vulnerability benchmarks such as DeepSWE and CWE-bench.
For the October 13 round, Google released two more pieces alongside the main announcement: a developer blog post on robotics, and an engineering retrospective on using Argon to optimize its own open-source codebase. The emphasis has shifted from whether the model can write code to whether it can understand the physical world.
Pricing is unchanged: during the introductory period, $2 per million input tokens and $10 per million output tokens, with cached input discounted a further 95%. Google did not say at launch how long the promotional period would last. This time it gave an end date: from November 30, prices return to $4 for input and $20 for output.
The numbers in the robotics blog
Google reported 87% on the embodied reasoning benchmark ERQA, which tests physical and spatial questions in three-dimensional settings. The company's own wording:
"Using pure reasoning, Argon scores 87%, outperforming every other frontier model."
The demos used Boston Dynamics' Spot quadruped. Argon first built an articulated digital twin in MuJoCo and practiced there repeatedly, then moved on to the physical robot, where it learned to step on a trash can pedal and drop a soda can inside. In another example, it controlled a robotic arm zero-shot to set a cup upright and tuck a napkin into it.
The most practical figures come from simulation generation. Researchers took a single photo of a tabletop, had Argon build a simulated scene from it, and generated more than 4,500 domain-randomized episodes to fine-tune a vision-language-action model. In real-robot tests, success rose from 56% to 75%, and completion time dropped from 35 seconds to 24. In a separate pick-and-place task with an SO-101 arm, three rounds of self-improvement cut the time from 58.9 seconds to 22.9 seconds, then to 17.4.
These are all Google's own experiments. The blog does not say how large the samples were or how many trials were run, and no third-party replication has surfaced so far.
Results reported by other customers
In the main announcement, Google listed a series of internal and partner use cases. The Chrome team processed roughly 6 million lines of legacy code in a 40-million-line codebase, shrinking refactors spanning hundreds of files from months to weeks. Inside Google, agent-authored code submissions rose by more than 35%, and rollbacks fell by nearly a third.
One publicly checkable example is the Python version of the open-source agent framework ADK. Engineers let Argon run unattended on the codebase for 11 hours. It proposed six optimizations, of which two were merged after human review, cutting memory preparation time in long sessions by about 40% and speeding up cold-start imports by 21%. Accepting only two of six says more about how the model is used today than the marketing does.
On safety, Google said that in Gray Swan's indirect prompt injection tests, the attack success rate fell to 0.7% under 15 adaptive attacks, and "shortcut-taking" behavior on coding benchmarks dropped to 2%.
Third-party evaluator Artificial Analysis updated its verdict on the day of the general release: Argon matches GPT-6 Astra on its Intelligence Index at roughly 60% of the cost per task. That cost comparison uses introductory pricing and will need to be redone after November 30.
Sources: Google official blog, Google Developers Blog robotics feature and ADK optimization retrospective, CocoLoop, Google DeepMind model card; Artificial Analysis Intelligence Index and per-task cost methodology.