Gemini 4 Argon pricing, access and token budgets
Gemini 4 Argon's introductory Standard API price is $2 per million input tokens and $10 per million output tokens through November 30, 2026. Google lists $4 input and $20 output starting December 1. A budget that crosses that date needs both rates.
USD list prices for the Gemini Developer API. Worked examples assume stated token counts and exclude taxes and other services. No paid inference benchmark was run for this article.
Standard, Batch, Flex and Priority prices
Google's Argon price table separates processing modes. Each rate cell below shows fresh input / cached input / output, in USD per million tokens. Output billing includes thinking tokens.
| Mode | Through Nov 30, 2026 | From Dec 1, 2026 |
|---|---|---|
| Standard | $2 / $0.10 / $10 | $4 / $0.20 / $20 |
| Batch | $1 / $0.05 / $5 | $2 / $0.10 / $10 |
| Flex | $1 / $0.05 / $5 | $2 / $0.10 / $10 |
| Priority | $3.60 / $0.18 / $18 | $7.20 / $0.36 / $36 |
Record the processing mode alongside every estimate. A Batch price cannot be applied to a Standard request merely because your application sends many requests together. Likewise, do not blend one mode's input rate with another mode's output rate.
The October 13 Google Cloud announcement confirms the introductory period and the later Standard rates. Those dates give you a way to build a forecast now without assuming the launch offer lasts indefinitely.
API access and Gemini App access are separate
The developer price page marks Argon unavailable on the API Free tier. Google's Gemini App announcement, meanwhile, says Pro and Ultra subscribers can access Argon in the app. Read the product name attached to an access claim: an app subscription is not the per-token API offer described here.
For application development, check the API project, billing route and selected model. For interactive app use, check the model offered in your app account. If a plan comparison only says "Gemini 4 access," establish which product it covers before purchasing it for a different workflow.
Google identifies the API model as gemini-4-argon. The current developer guide describes a shared 1,048,576-token window for input, thinking and output, with low, medium and high thinking levels. A larger output reservation leaves less room for input within that shared limit.
What does a 100K-input workload cost?
Take one request with 100,000 fresh input tokens and 10,000 total billable output tokens, including thinking. At introductory Standard rates, input costs $0.20 and output costs $0.10. The total is $0.30, or $300 for 1,000 identical requests. At the published December rates it becomes $0.60 per request and $600 for 1,000.
| Processing mode | Introductory / later cost |
|---|---|
| Standard | $0.30 / $0.60 |
| Batch | $0.15 / $0.30 |
| Flex | $0.15 / $0.30 |
| Priority | $0.54 / $1.08 |
For a forecast with 600 Standard requests before the price change and 400 after it, the token cost is $180 + $240 = $420. Using the promotional rate across the entire period would understate this example by $120. Split your volume by billing period before calculating an average.
A comparison with Muse Wonder's direct API price can use the same token counts as a starting point. Equal-token math does not establish which model completes an assignment more cheaply; include retries and rejected results when testing the actual task.
Implicit caching and thinking change the bill
The migration guide says Argon supports implicit caching, with eligibility beginning at 4,096 prompt tokens, but not explicit cache creation. Eligibility does not guarantee a hit. Use reported cached usage, and do not carry an explicit-cache storage assumption over from another Gemini model.
If 80,000 of our example's input tokens receive cached billing and 20,000 are fresh, introductory Standard input costs $0.008 + $0.04. Add $0.10 for the 10,000 billable output tokens: $0.148 per request. That is $148 for 1,000 requests. The same counts cost $296 at the later published rates.
Track total billable output rather than the length of the visible answer. An illustrative run with 2,000 answer tokens and 8,000 thinking tokens still uses the 10,000-output assumption above. A shorter-looking response can therefore cost more than a longer one with less reasoning.
Long decode needs an explicit budget
Google's guide requires background execution and an output limit above 65,536 for longer responses. The output limit includes thinking. Turning on background execution by itself retains the default 64K limit.
A ceiling of 262,144 billable output tokens would represent $2.62144 in introductory Standard output charges if fully used, before input or tools. It is a ceiling, not a prediction. Choose a limit appropriate to the requested deliverable, then inspect both completion status and usage instead of interpreting a partial answer as a successful low-cost result.
Keep grounding and task quality in the forecast
The price table lists a shared monthly Google Search grounding allowance of 5,000 across Gemini 3 and newer models, then $14 per 1,000 billable search queries. One user request can produce several queries. Do not allocate the whole shared allowance to Argon if other models already consume it.
For a first evaluation, retain the selected mode, price period, fresh and cached input, total output and separate tool charges. Pair that usage record with an acceptance result. A code fix should reproduce and resolve the reported failure; a research answer should have verifiable citations. Review time belongs in your comparison even though it does not appear on Google's invoice.
AI API Prices checked these official sources on October 14, 2026. Examples are independent arithmetic, not measured model throughput. Use the API budgeting guides for other model-specific cost boundaries, and recheck Google's table before committing a production budget beyond the introductory period.