A big round number leads Google's pitch for Gemini 4 Argon: one million tokens. Read the launch post, the model card and the pages for the previous Pro model side by side, and it is clear the number covers more than one thing.
What Google announced
Google announced Gemini 4 Argon on 30 September 2026. Its launch post, written by Koray Kavukcuoglu, opens with a summary that Google labels as generated by Google AI, which says the model "features an industry-leading 1 million token limit for deep, multi-step problem solving." Later in the same post, Google says it is "significantly expanding the model's output token limit to an industry-leading 1M tokens, up from the previous 64K tokens."
What the model card says
Google's Gemini 4 Argon model card, published on 13 October 2026, describes the limit differently. It lists inputs of text, images, audio and video, "with a combined input-output token context window of up to 1M." Under outputs it says "Text, with up to a 1M token input + output."
Atlas Cloud, a third-party explainer published on 16 October, says Google documents the limit as a shared budget: 1,048,576 tokens shared by input, thinking and output, not one million input tokens plus another million for the answer. It also says Google lists a 65,536-token default synchronous output cap, with a "long decode" mode that can extend generation up to the remaining shared capacity. That is Atlas Cloud's account of Google's specification.
What is new
A million-token context window is not new for Google's Pro line. Google Cloud's documentation for Gemini 3.1 Pro lists a context window of 1,048,576 tokens and maximum output of 65,536 tokens. Its model card, published on 19 February 2026, lists inputs "with a token context window of up to 1M" and outputs of "Text, with a 64K token output."
So the change Google itself highlights is on the output side, where the launch post compares 1M with the previous 64K. Whether a given request can use that room depends on how much of the shared window the input already takes up, and, on Atlas Cloud's reading, the model's thinking too. Atlas Cloud's own arithmetic makes the point: with 900,000 source tokens, at most 148,576 remain for thinking and output.
Who can use it
Argon did not launch to everyone. The 30 September post said it was rolling out to a set of trusted cyber defenders through Google's Fairwind Program. It also said Google would make it available to developers, enterprises and consumers "as soon as possible." CNBC reported that day that Alphabet launched it "to select cyber partners."
On 13 October, Google's safety post said: "Today, we are beginning to roll out Argon broadly, as well as expanding our Fairwind Program to additional trusted cyber defenders." The Neuron reported on 14 October that Pro and Ultra subscribers can try it in the Gemini app and paid API customers can build with it.
Google's launch post gives an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input price. After the introductory period it says $4 and $20 will apply. The Neuron reports the introductory price runs through 30 November.
The benchmarks are Google's table
Google's launch post says Argon "sets a new state of the art on DeepSWE v1.1 (77.9%)." The model card's results table is also Google's own. In it, Argon scores 55.0% on FrontierSWE v2 against 65.5% for GPT-6 Astra. On CWE-bench v1, Google's launch post says Argon "ties for first place with a top score of 68%"; Google DeepMind's cyber page charts GPT-6 Astra and Grok 4.7 at the same 68.0%.
That is a mixed picture, and it comes from the company that built the model.
Our view
The packaging is ahead of the plain reading. One million tokens is a real figure, but the model card describes it as a combined input-output window. Anyone planning around the headline should check how much of that window their own input will take.
Sources
Atlas Cloud (2026) 'Gemini 4 Argon's 1M Context Window: Why 1M Isn't All Input', Atlas Cloud, 16 October 2026, available at: https://www.atlascloud.ai/blog/tips/gemini-4-argon-1m-context-window (Accessed: 25 October 2026).
CNBC (2026) 'Google rolls out Gemini 4 Argon, its most advanced AI model', CNBC, 30 September 2026, available at: https://www.cnbc.com/2026/09/30/google-gemini-4-argon-ai.html (Accessed: 25 October 2026).
Dragan, A. and Popa, R.A. (2026) 'Developing Gemini 4 Argon responsibly and safely', The Keyword (Google), 13 October 2026, available at: https://blog.google/innovation-and-ai/technology/safety-security/gemini-4-model-safety-security/ (Accessed: 25 October 2026).
Google Cloud (2026) 'Gemini 3.1 Pro', Gemini Enterprise Agent Platform documentation, last updated 22 October 2026, available at: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-1-pro (Accessed: 25 October 2026).
Google DeepMind (2026) 'Gemini 3.1 Pro - Model Card', Google DeepMind, 19 February 2026, available at: https://deepmind.google/models/model-cards/gemini-3-1-pro/ (Accessed: 25 October 2026).
Google DeepMind (2026) 'Gemini 4 Argon - Model Card', Google DeepMind, 13 October 2026, available at: https://deepmind.google/models/model-cards/gemini-4-argon/ (Accessed: 25 October 2026).
Google DeepMind (2026) 'Gemini 4 Argon without cyber guardrails', Google DeepMind, available at: https://deepmind.google/models/gemini/cyber/ (Accessed: 25 October 2026).
Kavukcuoglu, K. (2026) 'Gemini 4 Argon: our next era of frontier intelligence', The Keyword (Google), 30 September 2026, available at: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ (Accessed: 25 October 2026).
Ruiz, E.G. and Harvey, G. (2026) 'Google launched Gemini 4 Argon, tying GPT-6', The Neuron, 14 October 2026, available at: https://www.theneurondaily.com/p/google-launched-gemini-4-argon-tying-gpt-6-b85b (Accessed: 25 October 2026).