Gemini 4 Argon reaches the API: the prices, the 1 December increase and how 1M-token outputs work

Google’s Gemini 4 Argon is now open to developers. Google first announced it on 30 September but gave access only to a vetted group of security teams. On 13 October it reached the paid tier of the Gemini API, with the model ID gemini-4-argon. If you plan to try it, three details matter more than the launch claims: the price doubles on 1 December, there is no free tier, and very long outputs only work through a separate background mode.

Where you can use it

On launch day Google listed these surfaces:

• Gemini API, paid tier. The developer post says nothing about a free tier, and the pricing page confirms that Argon is not available on the Free tier.
• Google AI Studio and Google Antigravity. In both, the developer post limits Argon to Google AI Ultra subscribers for now.
• Google Cloud. The Cloud announcement lists Gemini for business plus the Gemini platform API under the same model ID.
• The Gemini app. Google AI Pro and Ultra subscribers get Argon, along with a new tasks mode that keeps working across Google apps while you are away. Google says the app now connects to more than 40 apps.

None of the Google pages we read say whether Argon is a preview or a generally available model, or which Cloud regions serve it. If you need data residency in India, confirm the region before you build anything around it.

What it costs, and the 1 December jump

The Gemini API pricing page lists two sets of prices for Argon: introductory rates until 30 November 2026, and standard rates from 1 December. All prices are per million tokens, in US dollars. Output prices include thinking tokens.

Tier	Input (to 30 Nov)	Output (to 30 Nov)	Input (from 1 Dec)	Output (from 1 Dec)
Standard	$2.00	$10.00	$4.00	$20.00
Batch	$1.00	$5.00	$2.00	$10.00
Flex	$1.00	$5.00	$2.00	$10.00
Priority	$3.60	$18.00	$7.20	$36.00

Some details are easy to miss:

• Caching is implicit only. The pricing page says explicit context caching is not supported for Argon. Repeated input is discounted automatically: standard cached input costs $0.10 per million tokens now and $0.20 from 1 December. You cannot pin a large shared prefix yourself the way you can with explicit caches.
• No long-context tier. The pricing page lists no higher rate above a context-size threshold for Argon.
• Grounding is billed separately. Google Search grounding includes 5,000 free requests a month, shared across Gemini 3 and newer models, then costs $14 per 1,000 requests.

What this means for planning: if you benchmark Argon in October and size a budget on what you see, your bill doubles in December with no change to your code. Plan with the December prices. For overnight or bulk jobs that can wait (document processing, evals, back-fills), Batch and Flex cost half of Standard at both price levels.

For reference, from 1 December Argon’s standard price will match the list price of Meta’s new Muse Wonder on the Meta Model API: $4 input and $20 output per million tokens.

Long outputs: what “1M output tokens” means in practice

Google’s launch post says Argon’s output limit grew to 1M tokens, up from 64K. The long-decode documentation shows how that works:

• Normal streaming and non-streaming requests still stop at 65,536 output tokens. Support for more without background mode is described as coming soon.
• To go higher, set background: true (which also needs store: true) and a max_output_tokens above 65,536. The ceiling is 1,048,576 minus your input tokens.
• The server generates output in 64K “hops”. You poll or stream the interaction and join the text from every model_output step yourself.
• You pay for every token in every hop, including hops that contain only thinking. Cancelled or failed runs are billed for the work already done.
• Google recommends HTTP read timeouts of 15 to 20 minutes per stream. Background interactions can run for up to 24 hours.
• The docs also say the model does not follow token-count or word-count targets, and that long decode is not meant for chat or latency-sensitive calls.

So a 1M-token output is a batch-style job with its own plumbing: resumable streams, event IDs and stitching. You cannot get it by raising one parameter. It suits work such as large code migrations or full-document rewrites. For ordinary app traffic, treat 64K as the real limit and set a sensible max_output_tokens so a runaway reasoning chain cannot burn through your budget.

What Google claims it is good at

Google positions Argon for long, multi-step work: software engineering, legal and finance knowledge work, and security. The figures below are all reported by Google or its customers. We have not reproduced them.

• DeepSWE v1.1: 77.9%, which Google calls a new state of the art for long-horizon software engineering (launch post).
• AutomationBench (Zapier): 51.3%, ranked first; LVBench (long video understanding): 91.7% (same post).
• CWE-Bench v1: tied for first at 68% for fixing vulnerabilities, and an attack success rate of 0.7% across 15 adaptive attempts on Gray Swan’s indirect prompt-injection benchmark (Cloud post).
• Customer results: in the Cloud post, ServiceNow reports roughly 20% better task success than Gemini 3.8 Flash on its own NowAI-Bench, and Google says its engineers submitted over 35% more agentic code changes with Argon in Antigravity.

The prompt-injection result matters most if you plan to let the model read untrusted email, web pages or tickets and then act on them. It is still a vendor number, so keep your own guardrails: limit which tools the agent can call and require confirmation for irreversible actions. Google says Argon itself pauses for confirmation before irreversible steps.

Security teams: note the default block

Google is giving a version of Argon without cyber guardrails only to vetted defenders through its Fairwind Program. The general model, per the Cloud post, blocks offensive tasks such as penetration testing and exploitation by default. If your product does authorised pentesting or red-team automation, test your real workflows before you migrate. Expect refusals unless you qualify for the programme.

Should you switch now?

• Try it now if you run long agentic coding or document workflows and can use Batch or Flex. The introductory price gives you a cheap window to run your own evals until 30 November.
• Wait if you depend on explicit context caching, need a free tier for prototyping, or need a confirmed India region.
• Whatever you decide, build your cost model on the 1 December prices, and keep your current model as a fallback until your own tests confirm Google’s claims on your workload.