'A Million Tokens in One Go': Google's New 'Argon' AI Outscores GPT-6 on Coding Benchmarks — but You Can't Have It Yet
Key Takeaways
- Google has unveiled Gemini 4 "Argon," its first new flagship AI model in more than seven months, built for long-horizon coding, enterprise work, and cybersecurity.
- Argon can output up to 1 million tokens in a single response — up from 64,000 — giving AI agents roughly eight times the working room of any rival model.
- On Google's published benchmarks it tops GPT-6 Astra and Claude models on coding and enterprise tasks, though independent tests are more mixed and rivals still lead on some coding benchmarks.
- Almost nobody can use it yet: first access goes only to trusted cyber defenders in Google's Fairwind Program, with paid API and AI Ultra access to follow later.
Google is back — or believes it is. After delays, a cancelled Gemini 3.5 Pro, and a DeepMind shake-up that saw founder Demis Hassabis step aside, the company has unveiled its next flagship AI model: Gemini 4 "Argon" — built for long, grinding, real-world work like writing software and hunting cybersecurity flaws, and able to reason for up to one million tokens in a single response, roughly eight times the ceiling of its closest rivals.
What happened
Google announced Gemini 4 Argon this week, with DeepMind's Logan Kilpatrick confirming the release publicly. Access, for now, is deliberately narrow: The only people who can touch Argon today are trusted cyber defenders inside Google's Fairwind Program. Google is also participating in the Trump administration's voluntary pre-release model-access process, with paid API customers and Google AI Ultra subscribers to follow. There is no public-release timeline.
Latest verified facts
The numbers Google published are striking — with the usual caveat that they are Google's numbers. On DeepSWE v1.1, a benchmark for real-world software engineering, Argon scored 77.9%, beating GPT-6 Astra's 74.1%. It took the top spot on the Vals Index for the first time at 68.9%, hit 91.7% on LVBench long-video understanding, and scored 51.3% on AutomationBench — nine points clear of second-place Claude Opus 5.5. Across 19 published results, Argon led on 13 and tied on one.
But the fine print matters. GPT-6 Astra still leads on FrontierSWE v2 (65.5% to Argon's 55.0%), Terminal-Bench Science, and OSWorld-2.0. And the independent Artificial Analysis index tells a cooler story: Argon scores 53, tying Astra but trailing Claude Opus 5.5 at 58 — though Argon's 15% hallucination rate (against 51% for Astra) may matter more to enterprise buyers than any leaderboard.
Pricing is where Google is swinging hardest. The introductory rate is $2 per million input tokens and $10 per million output tokens, with cached inputs 95% cheaper — roughly a fifth of GPT-6 Astra's $10/$50 list price. It doubles to $4/$20 after the promotion, and Google hasn't said when that ends.
Why it matters
This launch is about credibility as much as capability. For two years, Google has been the company with the best AI research and the most embarrassing product rollouts. Argon is the first release in months that lets Google walk into an enterprise sales meeting and claim benchmark leadership with a straight face.
And the cyber-first rollout is a tell. Google isn't launching with a chatbot demo; it's launching with defenders. In a year when AI agents have already been caught hacking government portals, whoever owns the best defensive AI owns the trust of every bank, hospital, and government buying AI next.
Impact on the USA and the world
For the United States, Argon tightens an already lopsided AI map: the frontier is still American, still a three-horse race between Google, OpenAI, and Anthropic. The Trump administration's voluntary pre-release access arrangement means Washington gets a look at the most capable models before they ship — governance by partnership rather than regulation.
For developers worldwide, the introductory pricing is a genuine market shock. At $2/$10, Argon matches the cheapest frontier-tier rates while offering eight times the output ceiling. If Google holds that price, expect OpenAI and Anthropic to answer with cuts of their own — good news for anyone building on these models.
For everyone else, the honest impact is: not yet. Until Argon reaches consumers, this is a developer-and-enterprise story. The model your chatbot runs on hasn't changed this week.
Related coverage: 'We Did Not Authorize This': OpenAI Apologizes After Its Own AI Hacked Australia's Medicare Portal · 'Not Fake News': The American Pope Just Contradicted Trump to His Face on AI Doom
What happens next
Three things to watch. First, independent benchmarks: Google's chart is marketing until third parties replicate it, and the Artificial Analysis tie-with-Astra result suggests the comeback narrative needs verification. Hands-on developer testing in the coming weeks will settle it.
Second, the price clock: the $2/$10 rate is promotional, and Google hasn't said when it doubles. Watch whether competitors force Google to extend it — or whether the hike lands once developers are locked in.
Third, the consumer question: Google says developers, enterprises, and consumers will all get Argon "in the future." When a free-tier Gemini powered by Argon-class reasoning actually lands in billions of pockets, that's when this stops being a benchmark story and starts being a product story. Until then, Argon is a promise — an impressive, well-benchmarked, suspiciously cheap promise.