Gemini 4 Argon Is Live: Pricing, Benchmarks and API Tests

Gemini 4 Argon is the first model in Google's Gemini 4 line. It is live on Roar AI from today: send gemini-4-argon as the model name with your existing key. We spent the day running it through our API before announcing it. Below are the numbers people ask about first, then what we measured ourselves.

What is Gemini 4 Argon?

Google announced Argon on 30 September and opened it to developers through its API on 13 October. It is built for long, multi-step work in finance, law, coding and science, and it is a reasoning model: it thinks before it answers, and you control how much.

The full specs and the current rate are on the Gemini 4 Argon model page.

Gemini 4 Argon benchmarks

Google's own comparison table shows Argon ahead on most of the tests it published, with clear exceptions in coding. These are Google's figures, as reported by VentureBeat, and the settings behind each one are not public.

Independent results tell a similar but less lopsided story. Vals AI ranked Argon first of 43 models on its Vals Index, at 68.9%. Artificial Analysis puts it at 53 on its Intelligence Index, tied for third with GPT-6 Astra and Claude Fable 5.1.

The result we found most useful is about what Argon does when it doesn't know. On Artificial Analysis's AA-Omniscience test, which penalises wrong answers but not refusals, it hallucinated 15% of the time. GPT-6 Astra scored 51%. Argon says "I don't know" more readily than its rivals, which is what you want for research and analysis and something to plan around in an app that has no fallback when the model declines.

On agentic coding, Anthropic's newest models still lead.

Gemini 4 Argon pricing

Our rates, per million tokens:

Cached input costs $0.115 per million. Unlike Gemini 3.1 Pro, Argon has no surcharge for prompts over 200,000 tokens.

At launch, Argon costs less than Gemini 3.1 Pro on output. That will change. Google's introductory rate ends on 30 November, and its list price then doubles, which would put Argon level with Claude Opus 5.5. Our published rate follows it, and the pricing page always shows the rate in force.

What we found testing it

Reasoning effort is the setting that matters most for cost. We sent the same trivial prompt, "Say hi in five words", three times at each setting. On low, Argon answered in 11 output tokens every time. On high, it used 260 to 440, nearly all of them thinking, and took 11 to 18 seconds instead of about 4. At the default, it decides for itself: two of three runs thought for about 200 tokens before saying hello, and the third didn't think at all. Thinking is billed as output, so for classification, extraction and short replies, set reasoning_effort to low.

A grounded answer costs more in searches than in tokens. We asked three current-events questions with web search turned on. Argon ran between one and three searches per question and cited 6 to 13 sources. The searches made up 79% to 88% of each answer's cost, and a full answer came to between 2 and 6 cents. If a question doesn't need fresh information, leave search off.

Reading a page is billed as prompt. Asking Argon to summarise one web page added about 2,250 input tokens to the request, roughly half a cent at launch rates. A long page costs what a long prompt would.

Code execution checks its own arithmetic. Asked for the sum of the first 1,000 primes, Argon wrote and ran a short Python script and returned 3,682,913, which is correct. The code and its output come back in the answer, so a reader can check the working. The whole request took about 35 seconds.

Priority was faster on short requests. In a small sample, a short priority request returned in about 3 seconds against 5 for standard and 6 for flex. That is three requests, not a benchmark, so measure it against your own traffic.

Using the built-in tools

You turn on each tool the way you would declare any other:

{
  "model": "gemini-4-argon",
  "messages": [
    { "role": "user", "content": "What changed in the latest Node.js LTS release?" }
  ],
  "tools": [{ "type": "web_search" }]
}

Use url_context to have it read a URL from the prompt, and code_execution to let it compute an answer rather than estimate one.

Search results come back as citations in message.annotations, each a url_citation with the source page's own address. That is the shape the OpenAI SDKs already parse. Web search costs $16.10 per 1,000 searches, on top of tokens. URL reading and code execution carry no separate fee.

Speed tiers

Set service_tier on a request to choose how it is served:

• Standard, the default.
• flex, at half the standard rate. Requests may wait behind standard traffic, so use it for work nobody is watching in real time.
• priority, at a higher rate, for latency-sensitive calls.

If a priority request is served at standard speed, the response says so with "service_tier": "default" and you pay the standard rate. The models reference lists each tier's rate.

Building agents with Argon

Gemini 3 and later models attach a signature to each tool call they make and expect it back on the next turn. Argon does this, as do Gemini 3.1 Pro and Gemini 3.8 Flash. The signature arrives under extra_content on the tool call. If your code sends the assistant message back as it received it, there is nothing to change. If your framework rebuilds messages and drops the field, the call still succeeds, but keeping it lets the model carry its reasoning from one step to the next.

Context window and limits

Argon's context window is 1,048,576 tokens, shared between input and output. Through our API, an answer can run to 65,536 tokens. Google also offers a long-output mode that lets an answer run to a million tokens, and video input. Neither is available through our API yet.