Gemini 4 Argon Pricing: $2/$10 Now, $4/$20 From December
Gemini 4 Argon costs $2/$10 per million tokens until November 30, then doubles to $4/$20. Here's the full pricing breakdown and when it beats cheaper models.
Gemini 4 Argon launched on October 13, 2026 at $2 per million input tokens and $10 per million output tokens. Both rates double on December 1: input goes to $4 and output to $20. The model is available only on paid API tiers, with no free tier access, and the introductory prices run through November 30, 2026, on Standard, Batch, Flex and Priority alike.
The full price table
Argon has four price modes, and every one of them doubles on December 1.
| Mode | Input now | Output now | Input from Dec 1 | Output from Dec 1 |
|---|---|---|---|---|
| Standard | $2.00 | $10.00 | $4.00 | $20.00 |
| Batch / Flex | $1.00 | $5.00 | $2.00 | $10.00 |
| Priority | $3.60 | $18.00 | $7.20 | $36.00 |
| Implicit cache hit (Standard) | $0.10 | - | $0.20 | - |
| Implicit cache hit (Batch / Flex) | $0.05 | - | $0.10 | - |
| Implicit cache hit (Priority) | $0.18 | - | $0.36 | - |
Prices are per million tokens. Output pricing includes thinking tokens, so the bill covers the model's reasoning as well as the answer you see. Explicit context caching is not supported; implicit caching is on by default for prompts of at least 4,096 tokens. The cache discount is 95%: a token that would cost $2.00 to process fresh costs $0.10 when it hits the implicit cache.
Batch and Flex are both half the Standard price. If your workload can tolerate asynchronous processing, Batch gets you Argon at half the per-token rate on both sides of the December price change.
What the December increase actually costs you
If you run a workload that consumes 2 million input tokens and 500,000 output tokens per day, here is what the two price regimes look like:
- Through November 30 (Standard): (2M × $2) + (0.5M × $10) = $4 + $5 = $9/day, or about $270 for the month
- From December 1 (Standard): (2M × $4) + (0.5M × $20) = $8 + $10 = $18/day, or about $540/month ongoing
That same workload on Batch stays at $4.50/day now and $9.00/day after December. The jump is real, but Batch cuts it in half again if your pipeline can queue.
One planning note: you have about seven weeks at the launch price. That is long enough to run a real evaluation of whether Argon earns its place in your stack, but not long enough to treat the introductory rate as your steady-state budget.
What Argon actually is
Gemini 4 Argon uses a shared 1,048,576-token context window where input, thinking tokens and output all draw from the same pool. By default a response stops at 65,536 tokens; longer outputs, up to 1M tokens minus your input, need background execution (Google calls this long decode). Thinking level is tunable: low, medium (the default) and high. Google says temperature, top_p and top_k are ignored, and recommends a max_output_tokens of at least 32,768 because Argon can spend more of the limit on thinking than earlier models.
Google's pricing page lists Grounding with Google Search for the Standard and Priority tiers, not for Batch or Flex: 5,000 free search requests a month, shared across Gemini 3 and newer models, then $14 per 1,000.
Let each request use the cheapest model that's just as good
PromptUnit checks each request your app sends. When a cheaper model is just as good for that request, it uses it; otherwise your model answers. Free until it has saved you money.
Try it free →When Argon is worth the cost
At current introductory prices, Argon's per-token rate sits well below GPT-6 Astra ($10 input, $50 output per million tokens from the OpenAI pricing page) and below Claude Fable 5.1 ($10 input, $50 output per million tokens from Anthropic's pricing page). After December 1, the gap narrows: Argon at $4/$20 costs the same as Claude Opus 5.5 ($4 input, $20 output).
The question is not just whether Argon is cheaper per token but whether it is cheaper per useful task. Thinking tokens are billed as output, so measure output per task, not per visible answer. A task that reads 200,000 tokens and writes 60,000 thinking and response tokens costs 0.2 × $2 + 0.06 × $10 = $1.00 at today's standard rate, and 0.2 × $4 + 0.06 × $20 = $2.00 from December. The same token counts on Gemini 3.8 Flash at $0.75/$3.75 cost 0.2 × $0.75 + 0.06 × $3.75 = $0.15 + $0.225, about $0.38, less than a fifth of Argon at the December rate. That gap only closes if Argon's answers are materially better for your specific workload.
Argon earns its premium in a few clear cases:
Long-context work. The 1M shared window has one price at any prompt length (Gemini 3.1 Pro Preview charges more above 200K tokens), which makes Argon relatively competitive for tasks that process very large documents in a single pass.
Agentic loops with hard reasoning tasks. Complex multi-step agents on math, scientific analysis or software engineering tasks are where higher thinking effort pays off. For simple classification, summarization or extraction steps, the reasoning overhead is waste rather than value.
Tasks that need 65,000-plus output tokens. With long decode, Argon can produce very long structured outputs, like a full multi-file codebase or a long-form technical report, in a single call without stitching responses together.
For everything else, Gemini 3.8 Flash handles a wide range of production workloads at a fraction of the cost: $0.75/$3.75 through December 31, 2026, then $1.50/$7.50 from January 1, 2027. Gemini 3.1 Flash-Lite goes further: $0.25 input and $1.50 output per million tokens. If your workload fits those models, Argon at the December rate is hard to justify.
This is exactly the kind of routing decision that matters most at the step level of an agent. If a web-research step or a classification step does not need Argon's reasoning depth, running it on Argon anyway costs about 2.7 times Gemini 3.8 Flash's per-token price today ($2 against $0.75 input, $10 against $3.75 output), and about 5.3 times in December, until Gemini 3.8 Flash's own price doubles on January 1. PromptUnit makes that choice per request: it checks each request, and when a cheaper model is just as good for that request it uses that model; otherwise the request goes to the model you asked for. Setup is two settings (your app's AI API base URL and a PromptUnit key), and the dashboard shows what each step of your app costs.
The no-free-tier constraint
Argon is not available on Google's free tier or on the Google AI Pro plan's API access. You need a paid billing account, and your rate limits depend on your project's usage tier (Tier 1 to Tier 3 on Google's rate limits page). Check them before you commit Argon to a production pipeline that needs high throughput.
What to do before December 1
If you are considering Argon for a real workload, the next seven weeks are the right time to run that test, because at $2/$10 you can afford to compare Argon directly against Gemini 3.8 Flash or Claude Opus 5.5 on the same tasks without spending much. Keep in mind that output pricing includes thinking tokens, so measure actual token consumption under your real prompts rather than estimating from average response lengths. At high thinking effort, measured output token counts can run significantly above what the visible text suggests.
After December 1, Argon sits at roughly the same price as Claude Opus 5.5 and noticeably below GPT-6 Astra. Whether it earns that slot depends entirely on the reasoning depth your workload requires. For the tasks where it does, it is a strong option. For everything else, a cheaper model is the right call, and the cost difference compounds fast at any real scale. The multi-agent cost spiral is real: a 10-step agent loop that runs Argon at every step when only two of those steps need it pays Argon prices on eight steps that a cheaper model could handle.
For comparison, the Batch API 50% discount post has worked examples of stacking batch pricing with caching across providers. And if you are evaluating Argon against the Gemini 3.x lineup more broadly, the prior analysis of Gemini 2.5 Flash's long-context routing decision shows the same trade-off on an earlier generation.