Gemini 4 Argon Pricing: Checklist Before the 2x Hike

Know the introductory rates and what changes on November 30

Ars Technica reports that Gemini 4 Argon’s introductory API rates are $2 per million input tokens and $10 per million output tokens . The report also states that cached tokens receive a 95% discount. On November 30, the reported rates rise to $4 per million input tokens and $20 per million output tokens .

If the discount is applied to the input-token rate, cached input works out to $0.10 per million tokens at the introductory price and $0.20 after the change. Check your API billing details to confirm how cached usage is recorded for your account.

For a simple example, suppose a task uses 1 million input tokens and 200,000 output tokens, with half the input cached. At introductory rates, that is $1 for uncached input, $0.05 for cached input, and $2 for output: $3.05 total . At the announced rates, the same usage costs $6.10 . That is a useful first estimate, but only if your token volume and cache share stay the same.

These are API charges. Treat any subscription price or product access fee as a separate budget line rather than adding it to, or substituting it for, the API rate calculation.

Build a baseline from real API usage

Start with your highest-volume workflows, not an average across every request. For each one, collect at least a representative week of usage, or enough requests to capture normal variation. Separate workloads such as document summarization, code assistance, and customer support; combining them can hide which task drives the bill.

Use a simple measurement record:

Record totals and a typical per-request result, such as the median. The total shows budget impact; the per-request figure helps explain whether a few unusually large jobs are skewing it. Include retries and failed or revised requests if they consume billable tokens.

For each workflow, calculate the current cost using the actual input, output, and cached-token counts in your billing data. Keep the task’s quality measure alongside the cost. A cheap answer that needs repeated correction may be more expensive in practice than a higher-quality first response.

Benchmark representative tasks for token use and quality

Select a fixed sample of real tasks and run them with the same instructions, context, and success criteria. Track input and output tokens for every run, then judge the result against a rubric relevant to the work: factual coverage for summaries, tests passed for code, or reviewer acceptance for drafted replies.

The Verge notes that different models can use different numbers of tokens to complete the same task. So compare cost per accepted result , not only cost per request. For example, if one model produces a usable response in one run while another needs two revisions, include both attempts in the comparison.

Keep the sample stable when testing changes. If you revise the prompt, shorten the context, or alter the output format, note that as a separate version. Otherwise, you will not know whether a cost change came from the model, the prompt, or a different mix of tasks.

Forecast post-hike spend and test caching assumptions

Recalculate measured usage at the announced November 30 rates. If token volume and cache share stay constant, the token bill doubles. If either changes, use the measured numbers rather than applying a blanket multiplier.

A useful forecast for each workflow is:

• Uncached input tokens divided by one million, multiplied by $4.
• Cached input tokens divided by one million, multiplied by the applicable cached rate, if eligible.
• Output tokens divided by one million, multiplied by $20.
• The three amounts added together, then multiplied by expected monthly request volume.

Run at least two scenarios: your observed cache share and a conservative case with little or no caching. Caching can materially reduce eligible input costs, but do not assume every request qualifies. Measure how much repeated context your real workloads contain and confirm cached-token counts in billing data. A discount is only valuable when the workflow can actually use it.

Set a clear decision rule before changing production traffic. For example: optimize a workflow if its projected monthly cost exceeds its budget threshold and a prompt or caching test preserves the quality target; reroute it if another model passes the same rubric at a lower cost per accepted result; keep it if it meets the budget and quality bar, even if another option has a lower token rate. Set an alert at the threshold you choose, such as 80% of the workflow’s monthly allocation, so there is time to respond before the limit is reached.

Compare alternatives by task cost, and separate API from subscription access

The Verge reports that Argon’s introductory per-token rates are about one-fifth of the cited GPT-6 Astra and Fable 5 rates, and that Argon remains cheaper per token after the increase. That comparison is a starting point, not a decision: token use and output quality can differ by model. Benchmark alternatives on the same tasks and compare the cost of results that meet your quality bar.

Also distinguish API pricing from access through Google products. Ars Technica reports Gemini app availability for AI Pro, priced at $20 per month, and AI Ultra, starting at $99 per month. It reports that Antigravity and AI Studio access is limited to AI Ultra. These subscription and access details do not replace the per-token API forecast, so budget them separately if your team uses both.

Before November 30, run this checklist on your highest-volume workflows, recalculate them at the announced rates, and set a budget alert for any workload that no longer meets its cost or quality targets.

Step-by-Step Guide

1. 1

   Select high-volume workflows

   Separate workloads such as support replies, document summaries, and code proposals. Prioritize workflows that account for the most requests or spend.

2. 2

   Collect representative usage data

   Measure a representative week or enough requests to capture normal variation. Record input tokens, output tokens, cached input, retries, failures, revisions, current cost, and a quality result.

3. 3

   Benchmark tasks against a fixed rubric

   Run the same sample tasks with consistent instructions and context. Track token use and score each output against a relevant quality measure, such as reviewer acceptance, facts retained, or tests passed.

4. 4

   Forecast costs at the announced rates

   Recalculate uncached input at $4 per million tokens and output at $20 per million tokens, using the applicable cached-input rate only for eligible usage. Model both your observed cache share and a low-caching scenario.

5. 5

   Apply a workflow decision rule

   Compare projected monthly spend and cost per accepted result with your budget and quality thresholds. Optimize, reroute, or keep each workflow, and set an alert before it reaches its spending limit.

Key Statistics

• Argon's reported API input rate is scheduled to rise from $2 to $4 per million tokens, while its output rate rises from $10 to $20 per million tokens on November 30.Ars Technica reports the introductory and announced post-increase rates. These are API token charges, separate from subscription fees.
• Ars Technica reports a 95% discount for cached tokens; applying that discount to the reported input rate implies $0.10 per million cached input tokens before the increase and $0.20 afterward.The cached rates are calculated from the reported input rates and discount. Confirm cached-token eligibility and billing treatment in your account.
• The Verge reports that Argon's introductory per-token rates are about one-fifth of the cited GPT-6 Astra and Fable 5 rates.This is a reported per-token price comparison, not a comparison of cost per completed or accepted task. Token usage and output quality can differ between models.

Frequently Asked Questions