# Gemini 4 Argon explained: the benchmarks, the two prices, and why only cyber defenders get it first > Source: > Published: 2026-10-01 14:43:02+00:00 # Gemini 4 Argon explained: the benchmarks, the two prices, and why only cyber defenders get it first Google says Gemini 4 Argon leads on coding, agent and long-video tests. This page puts its two prices next to the models it competes with, sorts where it wins and loses, and explains why the first users are cyber defenders and not developers. **This explains reporting by** [Google, Gemini 4 Argon (Koray Kavukcuoglu, September 30, 2026)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/). Read the original first: [https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/) ## In one minute - Gemini 4 Argon is Google DeepMind's new top model. Today only trusted cyber defenders in Google's Fairwind Program can use it. - Paid API customers and Google AI Ultra subscribers are next. Google gives no date. - Intro price is $2 per million input tokens and $10 output, the same list price as GPT-6.1 Sol. The standard price after that is $4 and $20, the same as Claude Opus 5.5. - Google's numbers put Argon at 77.9 percent on DeepSWE v1.1, against 74.1 for GPT-6 Astra and 74.2 for Opus 5.5. It trails on four other tests. - Independent testers found Argon used more than twice as many output tokens per task as Astra, so the cheap token price may not mean a cheap task. ## What shipped, and who can use it Google announced Gemini 4 Argon on September 30. The post is signed by Koray Kavukcuoglu of Google DeepMind. It is rolling out first to a set of trusted cyber defenders through Google's Fairwind Program. Google says it is also taking part in the US government's voluntary process for pre-release model access. Google says paid API customers and Google AI Ultra subscribers come next, then wider developer, business and consumer access "as soon as possible". There is no date. So for most readers, today's news is a preview. You can plan for Argon, but you cannot test it yet. ## The two prices, next to the competition Token prices are per million tokens. A token is a chunk of text, roughly three quarters of a word. - Gemini 4 Argon, intro: $2 input, $10 output, cached input 95 percent off. - Gemini 4 Argon, standard: $4 input, $20 output. - GPT-6.1 Sol: $2 input, $10 output, $0.10 cached (OpenAI, September 29). - Claude Opus 5.5: $4 input, $20 output (as listed by VentureBeat). - GPT-6 Astra: $10 input, $50 output (as listed by VentureBeat). The intro price matches OpenAI's new mid-tier model to the cent. The standard price matches Anthropic's flagship. Google has not said when the intro price ends. ## Where Argon wins and where it loses Every score below is from Google's launch materials, as tabled by VentureBeat. None was run by an independent lab. - DeepSWE v1.1 (long coding tasks): Argon 77.9, Astra 74.1, Opus 5.5 74.2. - AutomationBench (business workflows): Argon 51.3, Astra 41.4, Opus 5.5 42.5. - LVBench (long video): Argon 91.7, Astra 87.5, Opus 5.5 83.7. - CWE-bench v1 (finding software weaknesses): Argon 68, Astra 68, Opus 5.5 67. A tie. - FrontierSWE v2: Argon 55.0, Astra 65.5. Argon loses. - Terminal-Bench Science 0.1: Argon 57.6, Astra 68.1. Argon loses. - Terminal-bench 4.0: Argon 57.4, Opus 5.5 66.4. Argon loses. - PostTrainBench: Argon 45.3, Opus 5.5 49.3. Argon loses. VentureBeat counts Argon leading or tied on 13 of 18 tests. The losses cluster on long terminal and science work, the same area where Astra still leads GPT-6.1 Sol. ## Cost per task beats cost per token A model that thinks longer writes more tokens. You pay for every one. Latent Space's AINews reports Artificial Analysis testing in which Argon used about 62,000 output tokens per task, against about 27,000 for Astra. Its Intelligence Index score was 53, the same as Astra. At the intro price, 62,000 output tokens cost about 62 cents. At Astra's $50 output price, 27,000 tokens cost about $1.35. At Argon's standard price, the same 62,000 tokens cost about $1.24. So Argon is cheaper per task than Astra on those numbers, but the gap is far smaller than the 5x token price gap suggests. Against GPT-6.1 Sol, which has the same intro price, the answer depends on how many tokens each uses on your work. ## Why the first users are cyber defenders A model that is very good at reading code is also very good at finding holes in it. Argon ties for first on CWE-bench, a test of finding software weaknesses. Google says Argon can find, validate and patch critical vulnerabilities on its own, and that defenders in Fairwind get it without the cyber guardrails other users will have. The logic is a head start. Defenders get time to find and fix holes before the same skill is in everyone's hands, including attackers'. Argon is not alone. OpenAI rated GPT-6.1 Sol Critical for cybersecurity on September 29, and on the same day Anthropic reported that the open GLM-5.3 model can build working exploits. The whole frontier is crossing this line at once. ## The 1 million output tokens, read closely Google says Argon can write up to 1 million output tokens, up from 64,000. Latent Space notes this uses an experimental feature called Long Decode Continuation, which pauses a long answer and resumes it across API calls. It reports that Vals lists 262,000 tokens as the standard maximum. Treat 1 million as a special mode until Google's developer docs say otherwise. ## Who is affected | Case | Status | |---|---| | Cyber defenders in Google's Fairwind Program | Access now, without the cyber guardrails other users will get. | | Paid Gemini API customers | Next in line. No date. Plan evals now so you can test on day one. | | Google AI Ultra subscribers | Next in line with API customers. No date. | | Teams on GPT-6.1 Sol or Opus 5.5 | Nothing to change today. Argon's intro price matches Sol and its standard price matches Opus. | | Everyone running internet-facing software | Frontier models now find bugs well. Shorten your patch window. | ## What to do - Pick 20 of your own real tasks now, so you can run them through Argon the day you get access. - Track cost per finished task, not price per token, when you compare models. - Do not plan on the $2 and $10 intro price lasting. Budget at $4 and $20. - Shorten patch times on anything exposed to the internet. Models that find bugs are spreading fast. ## What is still unknown - Every benchmark score comes from Google's own launch materials. No independent lab had run the full set when this page was written. - Google has not given a date for API or AI Ultra access, or an end date for the intro price. - The tokens-per-task figures come from Artificial Analysis as reported by Latent Space, not from Google. - Competitor prices for Opus 5.5 and Astra are as listed by VentureBeat, not read off each vendor's price page. - What Fairwind membership requires is not described in Google's post. ## Sources - [Google, Gemini 4 Argon (Koray Kavukcuoglu, September 30, 2026)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/) — the original report - [VentureBeat: Google unveils Gemini 4 Argon, but in limited release](https://venturebeat.com/technology/google-unveils-gemini-4-argon-retaking-benchmark-lead-over-openai-and-anthropic-but-in-limited-release) - [Latent Space AINews: Gemini 4 Argon, GDM's answer to Astra and Fable](https://www.latent.space/p/ainews-gemini-4-argon-gdms-answer) - [Anthropic: GLM-5.3 and the spread of advanced cyber capabilities](https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities) [AI News Report](https://theainewsreport.com/)· every headline, every morning.