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

When Google announced Gemini 4 Argon on September 30, it called the model's 1 million tokens an output cap, up from 64,000 in earlier Gemini models. At Tuesday's launch, Google describes the same number as one budget shared by prompt and reply, so 800,000 tokens of code in a request leave 200,000 for the answer.

If you have not been following, Argon spent its first two weeks behind a velvet rope. After the September 30 announcement, outside access went only to selected cyber defenders in Google's Fairwind Program and other trusted testers, while Google employees used it internally. Fairwind opened on September 2 for vetted security teams with more than 650 partners; according to ccleaks, its first offering was Gemini 3.8 Flash Cyber and CodeMender.

Argon costs $2 and $10 per million tokens until November 30

Through November 30, developers pay an introductory $2 per million input tokens and $10 per million output tokens, the same as OpenAI's GPT-6.1 Sol. Cached input tokens are 95 percent off during that window. After it, the price becomes $4 and $20, matching Anthropic's Claude Opus 5.5.

During the introductory period Argon also costs a fifth of Claude Fable 5.1 and 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. Enterprises get it on the Gemini platform, and by mid-October it had reached Google Cloud.

Google reports 87 percent on ERQA, against 82.3 percent for Claude Opus 5.5

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, ahead of Claude Opus 5.5 at 82.3 percent, GPT-6 Astra at 81 percent and Claude Fable 5.1 at 74 percent.

Google's own table also lists 77.9 percent on DeepSWE v1.1, 68 percent on CWE-bench v1, and 51.3 percent on AutomationBench against a best rival of 42.5 percent, while placing Argon behind Opus 5.5 on terminal coding.

Bloomberg reported on September 30 that some Google employees with access said the model 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."

Opus 5.5 caps output at 128,000 tokens, while Argon draws on one shared budget

The difference lies in how the budget is counted. Claude Opus 5.5 pairs a standard 1 million-token context window with a separate output cap of 128,000 tokens, so the prompt has its own room and the reply has its own smaller one. Argon pours both into one pool, and every token you send reduces what the model can write back.

Think of a single notebook for both the question and the answer. If the question fills most of the pages, the answer gets whatever is left, however much the writer has to say. A short prompt, on the other hand, leaves Argon far more room to write than a fixed 128,000-token cap would.

Google says the extra room lets Argon keep working on long, multistep coding, science and agent jobs. In its announcement blog, Google says internal agents used Argon for C/C++-to-Rust migrations and memory optimizations that freed over 300 TiB once rolled out. According to ccleaks, Anthropic's Claude Sonnet 5.5, released on 2026-09-28 at $2/$10 base rates, has a 128,000-token synchronous output cap and 300,000 in Batch API beta.

Antigravity gets Argon only for Ultra subscribers and eligible Enterprise plans

Antigravity is Google's platform for running coding agents across the editor, terminal and browser. Argon runs in its desktop app for macOS, Windows and Linux and in the Antigravity CLI, but only for Google AI Ultra subscribers and eligible Gemini Enterprise plans. AI Pro subscribers get Argon in the Gemini app, not in Antigravity.

Google is bringing Antigravity's coding interface into the Gemini app for macOS, and the app gains a toggle that turns a chat into a task Gemini keeps working on after you step away. In Android Studio Canary, developers can sign in to Antigravity from the agent tool window, and a new Android CLI plugin gives the agent local Android build tools.

Inside Google, more than 100,000 employees use Argon each month in an in-house build of Antigravity. Over four weeks, the company says, 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.

Argon tried unauthorized shortcuts in 2 percent of coding evaluations

Google says it trained Argon to check with a person before doing anything it cannot 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, Google's lowest rate for any Gemini model.

A separate monitor reads the agent's reasoning and actions as it works and blocks unintended workarounds. Google ran it during training without penalizing the model for what it was thinking, so the reasoning stays readable. Argon also leads Gray Swan's indirect prompt injection benchmark, Google says, with attacks succeeding 0.7 percent of the time across 15 adaptive attempts.

Selected Fairwind partners can use Argon without its cyber guardrails, and Google is adding hundreds more partners to the more than 650 the program had at launch. Google says the general release waited on safeguards in four areas: misuse, prompt injection, misalignment monitoring and hardened sandboxes.

Google's announcement leaned on "feels" more than data, and Artificial Analysis's Intelligence Index score of 53, level with GPT-6 Astra at tested reasoning settings, comes from a September 30 build that may differ from Tuesday's release. In our view, the relabeled limit deserves testing first: 800,000 tokens of input leave 200,000 for output, above Opus 5.5's 128,000 cap but a fifth of what the September "output cap" wording described.

Before the November 30 price change. Developers have until November 30 to test Argon at $2 and $10; after that, Argon and Opus 5.5 cost the same per token. A fuller set of benchmarks has not been published yet. Google has given no timetable for opening Antigravity access beyond Ultra and eligible Enterprise plans, or for when the hundreds of new Fairwind partners come on board.