Google announced Gemini 4 Argon on October 13, 2026. Its introductory API pricing gives buyers a useful evaluation window, with a specific end date.
The published rate schedule lists $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20 from December 1. These are dated rates, not a promise that future pricing cannot change. Check the current rate card before a purchase.
flowchart TD
A["October 13: Argon announced"] --> B["Through November 30: input 2 USD / output 10 USD per million"]
B --> C["From December 1: input 4 USD / output 20 USD per million"]
The important distinction is between an introductory offer and a durable budget assumption. The earlier “47-day” framing was misleading: October 13 to December 1 spans 49 days. The dates above are the useful planning inputs; the existing URL is retained so links continue to work.
Compare identical token mixes first
As of October 16, the published standard input/output rates provide this illustrative comparison. It holds usage constant at 100,000 input and 20,000 output tokens, with no caching, batch discounts or other charges. It is not a cost-per-completed-task benchmark.
| Model and rate period | Input / output per million | Illustrative token bill |
|---|---|---|
| Gemini 4 Argon, through November 30 | $2 / $10 | $0.40 |
| GPT-6.1 Sol | $2 / $10 | $0.40 |
| Claude Sonnet 5.5 | $2 / $10 | $0.40 |
| Gemini 4 Argon, from December 1 | $4 / $20 | $0.80 |
| Claude Opus 5.5 | $4 / $20 | $0.80 |
| GPT-6 Astra, short context | $10 / $50 | $2.00 |
Sources: Google's rate schedule, OpenAI pricing, Astra's model page, and Anthropic pricing. Context tiers, cache rules and token accounting can change the bill; compare the terms that apply to your request.
Budget for the rate after the promotion
The sample Argon bill doubles from $0.40 to $0.80 when the stated rate change takes effect. That arithmetic holds usage constant. Real workloads can also change their input size, generated reasoning, retries and caching behavior.
Use the introductory period for a paid evaluation, not as the basis of a long-term savings promise. Run representative tasks and measure completion quality, latency and the entire bill. If caching matters, test eligible repeated context and storage charges separately instead of treating a cache discount as universal.
flowchart TD
A["Published rate and applicable tier"] --> B["Measured tokens, cache and retries"]
B --> C["Cost per acceptable result"]
C --> D["Compare against December pricing"]
D --> E["Choose a model and a fallback"]
Route only where the evidence supports it
The table shows a fivefold difference between $2/$10 and $10/$50 for the same token mix. It does not show that the cheaper model will complete the same task with the same number of tokens or the same success rate.
A lower-cost default with a stronger fallback is worth testing where failures can be detected. Measure the fallback rate and the cost of doing work twice. For tasks where a missed error is expensive, the acceptance bar matters more than the headline rate.
Put December 1 in the cost model. Use the evaluation window to learn whether Argon meets your bar, and keep a migration path if either its quality or its future bill does not.
