Google's newest AI model can write up to a million tokens in one go, far beyond its old limit

Gemini 4 Argon matches OpenAI's top model on an independent benchmark and comes with a far longer output limit, but for now only a small group of trusted testers can actually use it.

1 October 2026 · 3 min read · L4 Models · The Pulse, 30 September 2026

In this story: Google DeepMind · OpenAI

Google DeepMind has released a new flagship AI model, Gemini 4 Argon. Its single biggest change is how much it can write in one response: Google says it is 'significantly expanding the model's output token limit to an industry-leading 1M tokens, up from the previous 64K tokens.' A token is a small piece of text, often a word or part of one, so this is the difference between a model that has to stop and restart on a long, complex task and one that can work through it in a single, much longer pass. Google has not named which of its own earlier models had the old 64,000-token limit.

Why it matters

Google is pricing the model at what it calls an introductory rate, $2 per million input tokens and $10 per million output tokens, with cached input (text the model has already seen before) at 95 percent off. Google's own page says that after the introductory period expires, a price of $4 per 1M input tokens and $20 per 1M output tokens will apply, doubling the cost once that window closes; Google has not said when that will happen. Separately, Artificial Analysis, an independent AI evaluator with no stake in either company, ran its own tests and reported that 'Gemini 4 Argon (high) scores 53 on the Artificial Analysis Intelligence Index, matching GPT-6 Astra (max, 53),' OpenAI's top model, while also finding Argon has 'a 15% hallucination rate... compared with 51% for GPT-6 Astra (max).' That means Argon is, by this one outside measure, just as capable as OpenAI's best model while making up false information far less often in that specific test.

What would change it

Right now Argon is only 'rolling out to a set of trusted cyber defenders through our Fairwind Program,' in Google's own words, alongside a separate US government pre-release access process. Google says broader availability to developers, enterprises and consumers is coming, starting with paid API customers and subscribers to its Google AI Ultra plan. The model's benchmark scores and its pricing only start to matter to most people once that wider rollout actually happens.

What we do not know

Google's own blog post is the only document that gives either the exact output-token figure or the exact pricing; this desk read it directly. Artificial Analysis's benchmark numbers come from its own published evaluation, not from Google, which is why they can be cited as independent. Beyond that, this desk found one outlet, VentureBeat, that added its own framing rather than simply repeating Google's numbers, writing that 'the practical limitation is that most customers will still need to wait for broader API access before they can test whether Argon's benchmark advantages translate into lower production costs on real coding, legal, financial and cybersecurity workloads.' No outlet this desk could reach had independent hands-on access to the model itself yet, so Argon's real-world performance outside Google's and Artificial Analysis's own tests is still unverified.

What this changes for you

If you use a Google AI product today, nothing changes yet: Argon is not broadly available. If you are a developer who pays for AI by the token, the thing worth watching is the price doubling once the introductory period ends, since that will roughly double your cost per response from this specific model once it applies to you.

Sources

Everything above is written from these. Each line says what that document proves.

1. blog.google: Google's own launch post: the token-limit figures, both pricing tiers, and the Fairwind Program access details, read directly
2. artificialanalysis.ai: Artificial Analysis's own independent benchmark run comparing Argon to GPT-6 Astra on its Intelligence Index and hallucination rate
3. venturebeat.com: Independent reporting, not Google and not the benchmark vendor, adding its own framing about the practical limits of the current limited release

Keep reading