Gemini 4 Argon reaches the Gemini API, AI Studio and the Gemini app

Google made Gemini 4 Argon available to the public on October 13, 2026, in a post by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. On September 30, Google had first rolled out Gemini 4 Argon to trusted cyber defenders through its Fairwind Program. Developers can now call the model on the Gemini API paid tier under the ID gemini-4-argon.

What is new in the Gemini 4 Argon release

The main change since September is access. According to the Gemini 4 Argon launch post (opens in a new tab), Argon is available globally to developers and enterprises through the API in AI Studio. Google AI Ultra and Pro subscribers get it in the Gemini app and in AI Mode in Search. Ultra subscribers can also use it in Google Antigravity.

The Gemini 4 Argon model page (opens in a new tab) lists the model as Stable. Argon accepts text, image, video, audio and PDF input and returns text only. The context window is 1,048,576 tokens, shared between input and output. Synchronous requests stop at 65,536 output tokens by default. Long decode raises that ceiling to the full 1,048,576 tokens, minus the input.

Supported features on that page include function calling, structured outputs, code execution, search grounding, Google Maps grounding, URL context, file search and thinking at low, medium or high levels. Computer use is listed as Preview. The Live API, image generation and audio generation are not supported, and caching is implicit only.

Gemini 4 Argon pricing on the API

The Gemini API pricing page (opens in a new tab) lists standard rates of $2.00 per 1M input tokens and $10.00 per 1M output tokens through November 30, 2026. From December 1, 2026, those rise to $4.00 and $20.00. Output prices include thinking tokens. Cached input costs $0.10 per 1M tokens, then $0.20.

Batch and Flex requests cost half the standard rate: $1.00 input and $5.00 output until the end of November. Priority inference costs $3.60 input and $18.00 output in the same period. Argon has no free tier. Grounding with Google Search includes 5,000 free requests per month, shared across Gemini 3 and newer models, then costs $14 per 1,000 requests.

How long decode works in the Gemini API

Outputs longer than 65,536 tokens go through the Interactions API with background execution, per the long decode guide (opens in a new tab). A request sets background: true, store: true and a max_output_tokens value above 65,536. The guide suggests 262,144 for hard reasoning tasks. Leaving out store: true returns a 400 error.

The server then works in hops of up to 64K tokens, each added as a model_output step under one interaction ID. Client code polls interaction.status until the run leaves queued or in_progress. A run can last up to 24 hours. Every token in every hop is billed, including hops that only contain thinking. Cancelled or failed runs are billed for the work done before they stopped.

Who Gemini 4 Argon is for

Google's Gemini 4 Argon model card (opens in a new tab) names coding, finance, legal, cybersecurity and science as the intended work. The long decode guide recommends long decode for hard reasoning in math, science, research and formal proofs. Coding and agent work should start at the 65,536-token default, without background execution. The guide advises against long decode for latency-sensitive requests, for agents that call tools in many short turns and for prompts near the input limit.

Scientists can apply for free access to the highest Antigravity usage tier for 6 months. Google is offering this to 10,000 scientists, subject to eligibility, capacity and selection. For security teams, Google says it is expanding the Fairwind Program to hundreds of additional partners.

Teams moving traffic from another model can rerun their existing prompts against Argon first, for example with promptfoo prompt and output tests, and compare cost per run at both the November and December prices.

Benchmarks Google reports for Argon

All figures here come from Google. The launch post says Argon ranks first on the Vals Index and scores 87% on ERQA, an embodied reasoning benchmark. On Gray Swan's Indirect Prompt Injection benchmark, Google says attack success across 15 adaptive attempts fell to 0.7%.

Google also reports that teams using Argon in Antigravity raised agentic code submissions by more than 35% and cut code rollbacks by nearly a third. The post cites no independent replication of these numbers.

Limits and open questions about Gemini 4 Argon

• The model card gives a March 2026 knowledge cutoff and lists hallucinations, occasional slowness and timeouts as limitations.
• Long decode without background execution is listed as "coming soon", with no date.
• The long decode guide says Argon does not follow length targets such as word counts.
• The pricing page says Argon is not available on the Google AI Pro plan, while the launch post lists Pro subscribers for the Gemini app and AI Mode. How those two statements fit together is not explained.
• The launch post does not name Vertex AI. Enterprise access is described as Gemini for business, the API and Antigravity.
• In the model card's Frontier Safety Framework section, Google says it could not rule out the cyber and biological critical capability levels, and built safety cases assessing the risk as below them.