Gemini 4 Argon
Google's return to the frontier tier. Index 53 at high, introductory $2 / $10, and a 1M output claim. The catch is access.
Announced 30 Sep 2026 for trusted cyber defenders in Google's Fairwind program. Paid API and Google AI Ultra access were described as next, with no date. Not a general API model as of 1 Oct 2026.
What this workload costs
List price for one call. Reasoning tokens count as output. A cache hit uses the lab’s cache rate when one is published.
Specs
- API id
- Not published here
- Released
- 30 Sep 2026
- Context
- 1M
- Max output
- 1M
- Input / 1M
- $2.00
- Output / 1M
- $10.00
- Cache read / 1M
- $0.10
- Knowledge
- —
- Inputs
- Text
- Weights
- Closed
Introductory list price from Google's launch post. The post says this later becomes $4 input and $20 output, and does not say when. Cached input is 95% off during the intro. The 1M context is what Artificial Analysis lists. Google published a 1M output limit and had not posted a stable API model card on 1 Oct 2026. No public model id yet.
Google's launch post did not publish a stable modality list. This card counts text only.
Use it when
You are inside the Fairwind program, or you are planning for the week it opens. High effort matches GPT-6 Astra's max score at a lower measured task cost, and it leads the agent automation eval Artificial Analysis published with the launch.
Leave it when
Any architecture that needs a public model id today. Use Gemini 3.8 Flash, Sol, or Opus until Argon is generally available.
Other independent scores
- Terminal-Bench 4.0: 57%. High reasoning. Behind Sonnet 5.5 (64%), Opus 5.5 (60%), and Astra (59%). Artificial Analysis, 30 Sep 2026.
- AutomationBench-AA: 78%. High reasoning. The best score in that comparison. Artificial Analysis, 30 Sep 2026.