Gemini 4 Argon is open to developers, at $2 per million input tokens

Google has made Gemini 4 Argon available to developers, calling it its frontier model for long-running work in software engineering, enterprise knowledge tasks and cybersecurity defense. According to the official announcement, access starts on the paid tier of the Gemini API, with $2 per million input tokens as the introductory price.

What is new in Gemini 4 Argon

Google positions Argon for long, multi-step jobs that call for deep reasoning, among them code migrations and frontier research. Access runs through the paid tier of the Gemini API, which covers Google AI Studio and Android Studio. Subscribers on the Google AI Ultra plan get it in AI Studio and in Antigravity, the company’s development platform built around agents. Google also says the model is coming to the Gemini app and AI Mode in Search for AI Ultra and Pro subscribers.

In its wider launch post, Google lists the commercial terms. The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%. From November 30, 2026, the listed rates rise to $4 and $20 per million tokens. The Gemini API offers a unified input and output limit of 1 million tokens.

Google credits the model with several results, most of them reported by the company itself:

• Argon ranks first on the Vals Index, which covers finance, coding, legal and tax work.
• It scores 87% on ERQA, a test of embodied reasoning, which Google says is above every other frontier model it tested.
• In Antigravity, teams using it increased agentic code submissions by more than 35%, and code rollbacks fell by nearly a third.
• Google says a refactor of roughly 6 million lines of legacy code inside the 40 million line Chrome codebase took weeks instead of months.

Koray Kavukcuoglu of Google framed the release this way: “We’re giving developers, enterprises, scientists, and consumers access to Gemini 4 Argon, our latest frontier model.”

Google also published a robotics post. In it, the company says the model learned a cornhole skill from video through a video-to-sim-to-real pipeline, and that Apptronik’s Apollo humanoid sank seven consecutive shots in a real-world trial.

Security and safety measures

Cybersecurity is a headline use case, and the safety material reflects it. Google says Argon solved all 12 end-to-end simulated cyberattack scenarios in its evaluations, up from five for the previous generation. Wiz used the model to find a critical bug in hospital software that earlier models had missed, according to Google, and its own security teams used it to find and patch bugs in internal and open-source code. Some of those flaws are going through responsible disclosure.

Google says its Frontier Safety Framework flagged elevated capabilities in cyber and biology, so it applied its newest safeguards to both. Security defenders can apply to the Fairwind Program for a version without the cyber restrictions. The company describes three real-time layers: input screening by smaller models, probes that read internal signals linked to cyber or biological knowledge, and continuous monitoring that can cut off a response mid-stream.

On misuse, Google reports that violations on an adversarial test set fell nearly twofold compared with earlier Gemini models. External red teamers could unlock less than 1% of the biological sub-tasks tested. On Gray Swan’s indirect prompt injection benchmark, attack success was 0.7% across 15 adaptive attempts. The company says it does not penalize the model’s reasoning text in training, so that an automated monitor can keep reading it. Full details are in the model card and a separate safety and security post.

Why it matters

For teams choosing a model for long-running agent work, the list price is only part of the cost. Computing for Geeks, citing Artificial Analysis figures, reports that Argon uses about 62,000 output tokens per index task, against about 27,000 for GPT-6 Astra. That puts its cost per task at $1.99 at the introductory price and $3.98 after the increase. The same write-up lists Anthropic’s Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens. That is where Argon’s rates land once the introductory pricing ends. Argon’s advantage on price therefore depends on how many tokens a given workload burns, and a team should measure that on its own tasks.

The same source also complicates the reliability case. In its own small test of infrastructure code, with three runs per task, Argon’s generated Kubernetes deployments never reached a ready state with HTTP 200 in either default or high-effort mode (0 of 3 each). Opus 5.5 and GPT-6.1 Sol managed 3 of 3. Linting passed for all four configurations. This is a single, narrow test and not a verdict on the model. Still, it suggests that DevOps and migration teams should run pilots on their real pipelines before relying on the vendor-reported results. The same write-up has Argon scoring 57% on Terminal Bench 4, behind Opus 5.5 at 60%.

Security teams face a different practical change, because access depends on vetting. Coverage by Noze says organizations in the Fairwind Program may share Argon only with internal security, incident response or penetration testing staff. They must also log who uses it, apply access controls and require phishing-resistant multifactor authentication. Defenders hoping for the unrestricted version should expect that kind of paperwork, and everyone else will work with the safeguarded model.