On 30 September Google announced Gemini 4 Argon, its new top model. It is going to trusted defenders through the Fairwind programme, and for them and Google's internal teams it will come without cyber safeguards. The broad release is a plan: paid API customers and Google AI Ultra subscribers first.
- The announced introductory price is 2 dollars per million input tokens and 10 per million output.
- The output limit rises to 1 million tokens from the previous 64 thousand.
- Internally, Argon agents found optimisations that free over 300 TiB of memory in Google's data centres once rolled out, per the company.
Over 300 tebibytes of memory. That, per Google, is what the optimisations Argon agents found on their own in the workload profiles free up in its data centres once rolled out.
Usually a new model is shown with benchmarks. Google puts ahead of them what the model already does in its own house. As stagecraft it is strong. It is as verifiable as the benchmarks, which is to say on their word.
Releasing to defenders first is now a line across the industry. Anthropic took the same road with Mythos Preview, now Google too. The logic is one: if the model is strong at attack, the guard gets it first.
The price is the more surprising part. Two and ten dollars per million tokens, the same as Sonnet 5.5 and GPT-6.1 Sol from the same week. Three companies, three different models, one figure on the sticker, even if Google's is introductory. The difference is now in how many tokens one task consumes.
When it will reach everyone, Google does not say. "Soon" is not a date.