Gemini 4 Argon is open to developers. The $2 price ends November 30.
Google's frontier model is now generally available on the paid Gemini API at $2/$10 per million tokens until November 30, then $4/$20, the same as Claude Opus 5.5. Here is how it compares and the four limits that matter when you build on it.
October 13, 2026
Two weeks after a launch almost nobody could use, Google made Gemini 4 Argon generally available on October 13. It is now on the paid tier of the Gemini API as gemini-4-argon, on Google Cloud, in the Gemini app for Google AI Pro and Ultra subscribers, and in Google's Antigravity coding environment for Ultra subscribers.
The model is the same one we covered at launch. What is new is that you can buy it, and the price comes with a deadline.
The $2 price ends on November 30
Google is keeping Argon's introductory pricing until the end of November, then doubling it. On December 1 it lands at exactly the price of Claude Opus 5.5.
| Per 1M tokens (input / output) | Price | Artificial Analysis Intelligence Index |
|---|---|---|
| Gemini 4 Argon, until Nov 30 | $2 / $10 | 53 |
| Gemini 4 Argon, from Dec 1 | $4 / $20 | |
| Claude Opus 5.5 | $4 / $20 | 58 |
| GPT-6 Astra | $10 / $50 | 53 |
| Meta Muse Wonder (new this week) | $4 / $20 (reported) | 52 |
Prices for Argon, Opus 5.5 and Astra are from each provider's official pricing page; Muse Wonder's are from press reports, since Meta does not list them. Batch and Flex processing halve Argon's rates, and cached input costs $0.10 per million tokens during the introductory period and $0.20 after it.
Per-token price is only half the story. Artificial Analysis describes Argon as verbose: it produced 110 million output tokens across its test suite against a median of 85 million. Even so, at the introductory price Argon cost about 60% less per task than GPT-6 Astra. At the December price, The Decoder calculates it will cost about 20% more.
Four details that matter when you build on it
- No free tier. Argon is paid-only in the API, and Google's docs say API access is not included in the Google AI Pro plan, even though Pro subscribers get it in the Gemini app.
- One shared 1M window. Input, thinking and output share a single 1 million token budget, so a long prompt leaves less room for output. Thinking tokens are billed as output.
- Per-request output caps. A single response is limited to 32,768 tokens without streaming (down from 65,536 on Gemini 3.x) and 65,536 with streaming. Google's SDKs continue longer generations across requests automatically, but raw HTTP code needs to handle this itself.
- Automatic caching only. Prompts of 4,096 tokens or more are cached implicitly; explicit cache control is not supported.
What changed on safety
The generally available Argon ships with jailbreak defenses and misuse monitoring. The version without cyber guardrails is now a separate model, gemini-4-argon-cyber-permissive, restricted to Google's vetted Fairwind security partners. Google reports a 0.7% attack success rate on Gray Swan's indirect prompt-injection test.
"a really, really, really good model." - Sam Witteveen, CEO of Red Dragon AI, to VentureBeat, while questioning whether it is the clear frontier leader
Should you switch?
If you are on GPT-6 Astra for long, document-heavy work, Argon is worth testing now: it matches Astra on independent intelligence scores and posts the lowest hallucination rate among top models, and the introductory price makes a six-week trial cheap. If you are on Claude Opus 5.5, the case is weaker. Opus scores higher on Artificial Analysis's index, and from December the two cost the same per token while Argon uses more tokens per task.
Whichever you test, measure cost per completed task on your own workload rather than price per token, and decide before December 1, when the comparison changes. Our AI Pricing Index will record the price step, and the model index has the full specs for every model above.
Some links in this article are affiliate links. Learn more.