Google launched Gemini 4 Argon today, its new frontier model, with model ID gemini-4-argon. It is available in the Gemini API (AI Studio), on Google Cloud and in Antigravity, with introductory pricing until November 30, 2026 that then doubles. If you run Gemini in production, it is worth measuring real cost and latency now.
What changes
- Model:
gemini-4-argon, with a 1M-token unified input-output window per Google's post. - Introductory pricing: $2 per million input tokens and $10 per million output tokens, with 95% off cached input tokens.
- Standard pricing: from November 30, 2026, $4 and $20 respectively. The Google Cloud post does not state cached-token pricing after that date.
- Availability: AI Studio and the paid API, Gemini for business, Google Antigravity and the Google Cloud agent platform (
region=globalin the try-it link). Google does not label the model GA or preview. - Vendor-reported results: first on the Vals Index (finance, coding, legal, tax) and on Gray Swan's indirect prompt injection benchmark, with 0.7% attack success across 15 adaptive attempts. These are vendor figures, not independently verified.
- Offensive security: the model blocks tasks such as pentesting and exploitation; vetted defenders can apply for access without those guardrails through the Fairwind Program.
Why it matters in production
Pricing is the key detail: the current tier is a promotional window, so any cost estimate made today will be too low from December. Also, "available" is not GA: in the material I reviewed Google says nothing about SLA or release stage.
The prompt injection benchmark matters if you build agents with tool access, but 0.7% on one benchmark does not replace testing with your own data and tools.
What I would do
- Measure before migrating: run your current evaluation set against
gemini-4-argonand compare quality, latency and cost per task. - Budget with the standard price ($4 / $20), not the introductory one.
- Keep the model ID in configuration, not code, so rollback is cheap.
# Keep the model id configurable so rollback is a config change
export GEMINI_MODEL=gemini-4-argon- If you run tool-using agents, re-run your own prompt injection tests; don't rely only on the vendor benchmark.