Gemini 4 Argon, and how it stacks up.

Google announced Gemini 4 Argon on 2026.09.30 and opened it to developers on 2026.10.13. It is the first Gemini 4 model, and it replaces Gemini 3.1 Pro (still a preview) as Google's flagship. Two things set it apart as of 2026.10: one answer can run to about a million tokens, where every rival stops at 128K, and the intro price is half of Claude Opus 5.5's. Independent rankings still put it a step behind Opus 5.5.

Watch the 60-second version on Instagram.

The spec sheet

The model id is gemini-4-argon. It reads text, images, video, audio and PDFs, and writes text only. There is no Live API. It is a sparse mixture-of-experts model with a knowledge cutoff of March 2026. It runs on the paid Gemini API, AI Studio, Android Studio and Antigravity, in the Gemini app for AI Pro and Ultra subscribers, and on Google Cloud's Gemini Enterprise Agent Platform (formerly Vertex AI). There is no free tier.

Google's announcement and model card compare Argon only with rivals; they have no Gemini 3.x column. So the Artificial Analysis Index is the only like-for-like number against older Gemini models, and it climbs from 30 to 41 to 53.

1M tokens out, and what that really means

Every model in this comparison can read about a million tokens. The difference is how much one answer can write.

Gemini 3.x stops at 64K. Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra stop at 128K. Argon can go to 1,048,576, but not by default. Out of the box it still stops at 65,536 tokens. To go further you turn on long decode: set background: true and store: true, and ask for more than 65,536 max_output_tokens. The server then writes in 64K hops, for up to 24 hours, and you poll for the result.

let job = await client.interactions.create({
  model: "gemini-4-argon",
  input: "Migrate this codebase to TypeScript…",
  background: true,
  store: true, // required by the docs, though their example leaves it out
  generation_config: { max_output_tokens: 262144, thinking_level: "high" },
});
// poll client.interactions.get(job.id) until status is no longer "in_progress"

Two things to keep in mind. The 1M window is shared by input, thinking and output, so a long prompt leaves less room for the answer. And you pay for every token in every hop, including hops that are only thinking.

The price

Until Nov 30, Argon costs $2 per million input tokens and $10 per million output tokens, and thinking counts as output. From Dec 1 it doubles to $4 / $20. Cached input is $0.10, then $0.20, and batch is half price. At the promo price it is half of Opus 5.5. After Dec 1 it matches Opus 5.5 and is still 2.5x cheaper than Fable 5.1 or Astra, both at $10 / $50 (Astra charges more above 272K input tokens).

The benchmarks, and who ran them

The first two columns come from Google's model card, and Google ran 10 of its 19 rows itself. In that table Argon wins most rows: DeepSWE 77.9%, AutomationBench 51.3% (Opus 5.5 42.5%, Astra 41.4%) and GraphWalks 256K to 1M at 84.2%, where the next best is 71.8%. It loses FrontierSWE (55.0% vs Astra's 65.5%), Terminal-bench, PostTrainBench, Terminal-Bench Science and OSWorld-2.0.

The last column is independent. On the Artificial Analysis Intelligence Index, Argon scores 53 at high effort, tied with Fable 5.1 and Astra. Opus 5.5 scores 58, and it also wins terminal coding, 66.4% to 57.4%.

What people get wrong

• "1M output means one huge reply every time." By default you get 65,536 tokens at most. A 1M answer is an async job you poll, billed for every hop.
• "Tops the benchmarks, so it's the best model." It tops Google's table, not the independent one.
• "A small max_output_tokens saves money." Set it to something like 8,192 and the answer can be cut off while the model is still thinking. To spend less, lower thinking_level instead.
• "My old request still works." It won't fail, but temperature, top_p and top_k are now silently ignored. thinking_level (low, medium by default, high) replaces the deprecated thinking_budget, and explicit caching returns an error.
• "Default timeouts are fine." Set HTTP timeouts to 15 to 20 minutes.

On a client project I'd try Argon first for the long jobs, a whole-codebase migration or a long report, where the 1M output is the point, and keep Opus 5.5 on terminal-heavy coding agents, where the independent numbers still favour it.

The short version

1. Gemini 4 Argon is Google's new flagship, on the paid Gemini API since 2026.10.13. No free tier.
2. It can write up to 1M tokens in one answer through long decode; by default it still stops at 64K, and rivals cap at 128K.
3. $2 / $10 per million tokens until Nov 30, then $4 / $20: half of Opus 5.5 now, the same after Dec 1, 2.5x cheaper than Fable 5.1 and Astra.
4. It wins most rows of Google's own table, but on the independent Artificial Analysis Index it ties at 53 and Opus 5.5 leads with 58.
5. Sampling parameters are ignored, explicit caching errors, and a small max_output_tokens can cut off the thinking. Lower thinking_level instead.

Sources: Google, Gemini 4 Argon: our next era of frontier intelligence · Google, Gemini 4 Argon is now available for developers · Gemini API pricing · Gemini API, Long decode · Gemini 4 Argon model page · Google, Gemini 4 Argon in the Gemini app · What's new in Gemini 4 Argon · Google Cloud, Developer's guide to Gemini 4 Argon · VentureBeat · dsebastien · Gemini thinking docs · Google DeepMind, Gemini 4 Argon model card · Artificial Analysis, Gemini 4 Argon · Artificial Analysis leaderboard · Gemini models list · Gemini 3.1 Pro Preview model page · Gemini 3.8 Flash model page · Claude models overview · Claude pricing · OpenAI, GPT-6 Astra model page · Artificial Analysis, Gemini 3.1 Pro Preview · Artificial Analysis, Gemini 3.8 Flash · Apidog, Argon benchmarks methodology