Gemini 4 Argon first look: pricing, limits and what to change in your prompts

Gemini 4 Argon costs $2.00 per million input tokens and $10.00 per million output tokens until November 30, 2026, then $4.00 and $20.00 from December 1. It has a 1,048,576-token context window, no Live API, no image or audio generation, and no explicit caching. Those are the facts that decide whether it fits your app, so this post starts there and then shows what a call actually costs.

Everything below was read on Google's own pages on 2026-10-30 (listed at the end). I did not call the model, because no API key was available in this session. Where I computed something, the script is shown and the output is real. Where I could not verify something, it says so.

What is Gemini 4 Argon, and is it generally available?

The model ID is gemini-4-argon. Google's changelog entry for October 13, 2026 says it became generally available and describes it as "our most capable model built for complex coding, agentic, math, and science tasks." The model page lists it under Versions as "Stable: gemini-4-argon".

Inputs are text, image, video, audio and PDF. Output is text only. The page does not state a knowledge cutoff, and it shows no release date beyond the changelog entry, so do not quote either from this post.

What does Gemini 4 Argon not support?

This is the section that saves you a migration afternoon. From the model page:

Two practical consequences. If your product is a voice agent, Argon is not your realtime model; look at the Live-capable Gemini models instead and keep Argon for the text reasoning behind them. And if your cost plan depends on creating a cache object once and reading it for hours, that plan does not exist on this model. You get implicit caching only, which means you cannot reserve a cache, you can only structure prompts so the shared prefix is stable.

What are the token limits?

The context window is 1,048,576 tokens, described as a unified input and output limit. Per Google's thinking page, input, thinking and output tokens share that one window. If input plus max_output_tokens exceeds it, the API caps the output limit to the space left.

Output has a second limit. With synchronous calls the default and maximum max_output_tokens is 65,536, and larger values return a 400 error. To go beyond that, the page says to set background=True with a max_output_tokens above 65,536 (the "long decode" route, up to 1,048,576). I did not test this.

One trap worth knowing: if thinking uses up max_output_tokens, the response status is "incomplete", the visible output may be truncated or empty, and the thinking tokens are still billed. Google's advice on that page is to lower thinking_level to cut cost, not to set a small token cap.

How does the pricing work, and what happens on December 1?

From the pricing page, Standard tier, per 1M tokens:

The model is not available on the Free tier. On the paid tier, Google's page says content is not used to improve its products. Grounding with Google Search gives 5,000 free requests a month shared across Gemini 3 and newer models, then $14 per 1,000 requests, with a daily cap of 1,500 requests per the page.

The intro price is a clock, not a discount you can lock in. Anything you budget in October at $2 and $10 doubles on December 1. If you are writing a cost estimate for a client now, price it at the December numbers.

How does Argon compare per task with Sonnet 5.5 and GPT-6.1 Sol?

List prices alone mislead, so here is a script that turns them into dollars per call. The prices are from the vendor pages read on 2026-10-30 (Google pricing page, OpenAI pricing page, Anthropic's Sonnet 5.5 page). The workload sizes are my assumptions, not measurements, and thinking tokens are counted as output.

PRICES = {  # USD per 1M tokens: input, cached input, output
    "Gemini 4 Argon (to Nov 30)": (2.00, 0.10, 10.00),
    "Gemini 4 Argon (from Dec 1)": (4.00, 0.20, 20.00),
    "Claude Sonnet 5.5": (2.00, 0.10, 10.00),
    "GPT-6.1 Sol": (2.00, 0.10, 10.00),
}
WORKLOADS = {
    "chat reply (3k in, 500 out)": (3_000, 0, 500),
    "doc review (40k in, 3k out)": (40_000, 0, 3_000),
    "agent step (30k in, 24k cached, 1.5k out)": (30_000, 24_000, 1_500),
}

def cost(p, total_in, cached, out):
    inp, cin, o = p
    return ((total_in - cached) * inp + cached * cin + out * o) / 1_000_000

Output from the full version of that script (which also included Terra, Luna and Astra):

chat reply (3k in, 500 out)
  Gemini 4 Argon (to Nov 30)     $0.0110 per call, $   11.00 per 1,000 calls
  Gemini 4 Argon (from Dec 1)    $0.0220 per call, $   22.00 per 1,000 calls
  Claude Sonnet 5.5              $0.0110 per call, $   11.00 per 1,000 calls
  GPT-6.1 Sol                    $0.0110 per call, $   11.00 per 1,000 calls

doc review (40k in, 3k out)
  Gemini 4 Argon (to Nov 30)     $0.1100 per call, $  110.00 per 1,000 calls
  Gemini 4 Argon (from Dec 1)    $0.2200 per call, $  220.00 per 1,000 calls
  Claude Sonnet 5.5              $0.1100 per call, $  110.00 per 1,000 calls
  GPT-6.1 Sol                    $0.1100 per call, $  110.00 per 1,000 calls

agent step (30k in of which 24k cached, 1.5k out)
  Gemini 4 Argon (to Nov 30)     $0.0294 per call, $   29.40 per 1,000 calls
  Gemini 4 Argon (from Dec 1)    $0.0588 per call, $   58.80 per 1,000 calls
  Claude Sonnet 5.5              $0.0294 per call, $   29.40 per 1,000 calls
  GPT-6.1 Sol                    $0.0294 per call, $   29.40 per 1,000 calls

The finding is dull and useful: until November 30, Argon, Sonnet 5.5 and GPT-6.1 Sol have identical list prices on all three shapes ($2 input, $0.10 cached input, $10 output). After December 1, Argon is twice as expensive as the other two at list. So the decision is not price. It is which model passes your task with fewer tokens, because a model that thinks for 8,000 tokens before answering costs more than a model that thinks for 1,000 at the same rate.

Caveats: Gemini's cache price applies to implicit hits, which you do not control; OpenAI's long-context rates apply above 272K input tokens and are not modelled; and I have not measured how many thinking tokens each model spends on a given task. For the per-1,000-request method on other models, see LLM API cost per 1,000 requests.

How should you set thinking and prompts for Argon?

Google's thinking page says the parameter is thinking_level inside generation_config, with low, medium and high, defaulting to medium. The page lists no "off" setting for gemini-4-argon, and does not say explicitly whether thinking can be disabled. So plan on always paying for some thinking.

An untested starting routine, since I could not run it:

1. Start every task at low. Run your eval set. Note pass rate and total_thought_tokens.
2. Move to medium only on the task types that fail. Move to high only if medium still fails and the task is worth the extra output tokens.
3. Keep your system prompt and few-shot examples at the start of the prompt, identical across calls, and put the changing user content last. With implicit caching, a stable prefix is the only lever you have.
4. Do not copy prompts tuned for another model without an eval. The older Gemini 3.8 prompting guide covers the Gemini 3.x habits; I have not verified which of them carry over to Argon.

Do not set a small max_output_tokens to save money. As above, that can truncate the answer while still billing the thinking.

Should you move to Argon now? A decision checklist

• You need the Live API or image/audio generation: no, use another model.
• You depend on explicit cache objects: no, or redesign around prompt-prefix stability.
• You have an eval set and a cheaper model fails it: test Argon at low and medium, and cost it at the December prices.
• You do not have an eval set: build one first. Which AI model should I use and the model pricing overview help you choose the candidates.
• You send personal data: check the paid-tier terms and storage settings first; the Gemini storage default checklist covers a change from October 26.

What I could not verify

• Benchmark results. I am not quoting any, because I did not copy them from a primary page with a date.
• Knowledge cutoff and exact release date: not stated on the model page.
• Real-world latency, thinking-token usage and behaviour of long decode: not tested, no API key.
• Google's pages were summarised by a fetch tool, then re-read for the specific numbers above. Prices and limits can change, so open the pages before you commit.

Sources, all read 2026-10-30: Gemini 4 Argon model page, Gemini API pricing, Gemini API changelog, thinking guide, OpenAI pricing, Claude Sonnet 5.5.