Skip to content

Google

Gemini 4 Argon’s 1M tokens now cover prompt and answer

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: AI prices →
Video: threads.com

Google on October 13 opened Gemini 4 Argon, its new flagship model, to developers on the paid tier of the Gemini API through Google AI Studio. The 1 million tokens it advertises are a single budget shared by the prompt and the reply, as Google’s documentation describes the limit.

Google described the limit differently when it announced Argon on September 30. At that point it was an output cap, up from 64,000 tokens in earlier Gemini models, and ET Now reported that the figure referred specifically to output and not to the input context window. Google said at the time that the extra room would let the model spend more time reasoning and generate hundreds of thousands of tokens in a single trajectory.

With one shared budget, a long prompt shrinks the room for the answer. A request that loads 800,000 tokens of code leaves 200,000 for the response. Claude Opus 5.5 works differently: it pairs the standard 1 million-token context window with a separate output cap of 128,000 tokens.

The introductory price runs through November 30: $2 per million input tokens and $10 per million output, the same as OpenAI’s GPT-6.1 Sol, with cached input 95 percent off. After that Argon costs $4 and $20, which matches Opus 5.5. Until then it costs a fifth of Anthropic’s Claude Fable 5.1 and OpenAI’s GPT-6 Astra, which both list at $10 and $50 per million tokens.

Developers call the model as gemini-4-argon through Google’s Interactions API, with support in the google-genai SDKs, and Android developers can build it into apps through Firebase AI Logic. Outside the API, access depends on the plan. Argon runs in Antigravity, Google’s platform for running coding agents across the editor, terminal and browser, in both the desktop app for macOS, Windows and Linux and the Antigravity CLI. But Antigravity access is limited to Google AI Ultra subscribers and eligible Gemini Enterprise plans. AI Pro subscribers get Argon in the Gemini app and in AI Mode in Search, but not in Antigravity.

Several smaller tools changed alongside the model. In Android Studio Canary, developers can sign in to Antigravity from the agent tool window, and a new Android CLI plugin gives the agent access to local Android build tools. The Gemini app for macOS is getting Antigravity’s coding interface, so users can switch between chat and coding in one app. The Gemini app also gets a toggle that turns a chat into a task, which Gemini keeps working on after the user steps away from the computer.

What Google measured inside

“This model has been battle tested by over 100,000 people at Google over the last month,” Logan Kilpatrick wrote on launch day, calling Argon “a huge step forward for Gemini, and a labor of love.” Google says those 100,000-plus employees use Argon each month in an in-house build of Antigravity. Over four weeks, code submitted by agents rose 35 percent and post-submit rollbacks fell 30 percent. More than half of Google’s Cloud sales staff and almost half of its finance team now use Antigravity, the company says.

Google says Argon ranks first on the Vals Index, which scores models on finance, coding, legal and tax work and weights each sector by its share of U.S. GDP. On ERQA, a test of physical and spatial reasoning, Google reports 87 percent for Argon against 82.3 percent for Opus 5.5, 81 percent for GPT-6 Astra and 74 percent for Fable 5.1. At the September announcement Google claimed 77.9 percent on DeepSWE v1.1. 9to5Google put Opus 5.5 at 74.2 percent and GPT-6 Astra at 74.1 percent on that test. The same September launch table had Argon finishing last on FrontierSWE v2 and Terminal-Bench 4.0.

Bloomberg reported on September 30 that some Google employees with access to the model said it struggled with certain coding tasks, particularly front-end design. Google told Bloomberg it would be “inaccurate to say that Gemini 4 is underperforming in areas such as coding.” Artificial Analysis gave Argon 53 on its Intelligence Index on September 30, matching GPT-6 Astra at the reasoning settings it tested, though nothing guarantees that the model released on Tuesday is identical to the one it evaluated.

Guardrails and agent safety

Google is expanding Fairwind, the program it launched September 2 to give vetted security teams early access to its cyber models, and selected partners there can use Argon without its cyber guardrails. Fairwind had more than 650 partners at launch, and Google says it is adding hundreds more.

For the public model, Google says it trained Argon to check with a person before doing anything it can’t undo, and to report when it hits a wall instead of working around it. On coding evaluations, Argon tried unauthorized shortcuts 2 percent of the time, which Google says is the lowest rate for any Gemini model. Argon also leads Gray Swan’s indirect prompt injection benchmark, where attacks succeeded 0.7 percent of the time across 15 adaptive attempts.

A separate monitor reads the agent’s reasoning and actions as it works and blocks unintended workarounds. Google says it ran that monitor during training without penalizing the model for what it was thinking, so that Argon’s reasoning stays readable.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.