Updated 14 October 2026, 08:30 CET: Google opened Argon to paying Gemini API customers and to Google AI Pro and Ultra subscribers on 13 October, confirmed the model ID and set an end date for the launch price. We rewrote the price, access and verdict sections and added the first independent benchmark scores.
Google opened Gemini 4 Argon to every paying Gemini API customer on 13 October 2026, two weeks after a launch that only cyber defenders could use. Google AI Pro and Ultra subscribers now get it in the Gemini app as well. The launch price of $2 per million input tokens and $10 per million output tokens runs until 30 November. From 1 December it doubles to $4 and $20, exactly what Claude Opus 5.5 costs.
From 1 December Argon costs the same as Claude Opus 5.5
At launch the open question was how long the cheap rate would last. Reuters and VentureBeat now report the answer: until 30 November. After that, Argon matches Opus 5.5 on both input and output, and costs twice as much as Claude Sonnet 5.5 or GPT-6.1 Sol. A business case built on $2 and $10 is only valid for seven weeks.
- Model ID: gemini-4-argon, listed as stable in Google’s API documentation
- Developers: the paid tier of the Gemini API and Google AI Studio
- Antigravity, Google’s agent coding tool: Ultra subscribers, or businesses through Gemini Enterprise
- Consumers: the Gemini app and AI Mode in Search, for Google AI Pro and Ultra
- Price: $2 and $10 per million tokens until 30 November, then $4 and $20 (as reported by Reuters and VentureBeat; Google’s own posts do not state it)
- Cached input: 95 percent off, so $0.10 per million tokens at the launch rate
Google’s posts name no regions. Reuters describes the release as worldwide. If you need EU data residency, the route is Vertex AI, where Argon has a Model Garden page, but Google has not said which regions host it.
The 1 million token answer needs a separate mode
Google’s API page now fills a gap from the launch post. Argon has one shared context window, the text it can handle at once, of 1,048,576 tokens. Input and output come out of that same budget, so a long prompt leaves less room for the answer.
A normal request still stops at 65,536 output tokens, the same ceiling as earlier Gemini models. The 1 million figure only applies with long decode, a continuation mode you switch on in the API. Agent builders who expected one call to write a full codebase refactor will need to change their integration first.
Outside tests rank Argon first on business work, not overall
The first independent leaderboards agree with Google on professional tasks and disagree on general ability.
| Model | Vals Index (finance, legal, tax and coding tasks) | Artificial Analysis Intelligence Index (broad mix of hard tests) |
|---|---|---|
| Gemini 4 Argon | 68.90% (ranked 1st) | 53 |
| Claude Sonnet 5.5 | 67.04% | 56 |
| Claude Opus 5.5 | 66.97% | 58 |
| GPT-6 Astra | not in our sources | 53 |
These are third-party scores, not Google’s. The Vals AI figures are from its leaderboard as of 12 October. The Artificial Analysis figures are as reported by VentureBeat on 13 October. Each lab runs its own test setup, so the two do not contradict each other: Vals weights the document-heavy work Google built Argon for, while Artificial Analysis averages a wider set of reasoning, knowledge and coding tests.
Google’s own launch table pointed the same way. Argon led on DeepSWE and business automation, but trailed Opus 5.5 on Terminal-Bench 4.0 by 9 points, 57.4% against 66.4%.
Run your legal and finance tasks on Argon before 30 November
For teams with document-heavy legal, finance or tax work, Argon is now worth a proper trial. It tops the one independent index built on that kind of work, and for the next seven weeks it costs the same as Sonnet 5.5. That is a cheap window to find out whether it beats your current model on your own files.
Start with a single long job, such as reviewing a full contract set, with long decode switched on. Check whether the single long answer holds together better than your current chunked approach. Then calculate cost per finished task at $4 and $20, because that is the price you will pay in production.
Coding teams that rely on command line agents have less reason to switch. Opus 5.5 still leads Google’s own terminal benchmark and the broader Artificial Analysis index. Dutch and other EU organisations with strict data residency rules should wait until Google names the Vertex AI regions that serve Argon.
For more information, visit the official announcement of Gemini 4 Argon on the Google blog.
