Sign in Subscribe
news

2026-10-03 Gemini 4 Argon - Google is back at the top, but you can't use it yet

Sebastien Dubois

Sebastien Dubois

— 5 min read

Canonical version: 2026-10-03 Gemini 4 Argon - Google is back at the top, but you can't use it yet.

On 30 September, Google DeepMind announced Gemini 4 Argon, its new frontier model. Demis Hassabis called it "a huge step forward across key capabilities" on X. Koray Kavukcuoglu, who wrote the announcement, says it's already changing how Google works and builds internally.

According to Artificial Analysis, Google is back among the top three labs: it scores Argon 53 on its Intelligence Index, level with GPT-6 Astra and one point above GPT-6.1 Sol. That's 23 points above Gemini 3.1 Pro Preview, Google's previous non-Flash model. And it's the first model above the Flash class Google has shipped in more than seven months.

Right now, Argon is only rolling out to "trusted cyber defenders" through Google's Fairwind Program.

What Google announced

  • A model built for long, multi-step work: real-world software engineering, enterprise knowledge work (legal, finance), and cybersecurity defense
  • 1M output tokens, up from 64K. The model can think and write for hundreds of thousands of tokens in one go. A new Gemini API feature, Long Decode Continuation, resumes long responses across calls so they don't time out
  • An introductory price of $2 / $10 per million input/output tokens, with cached input 95% off. After the promotion, it doubles to $4 / $20. No end date yet
  • A phased rollout: cyber defenders first, while Google takes part in the U.S. government's voluntary pre-release access process. Then paid API customers and Google AI Ultra subscribers. Then everyone else, "as soon as possible"

Google also shared what Argon already does inside the company. A team of Argon agents analyzed fleet-wide telemetry and freed over 300 TiB of memory across Google's data centers (with 500 TiB to 1 PiB expected in total). Others are migrating C/C++ code to Rust, up to 800K+ lines for the Fuchsia Zircon kernel. And on a quantum computing problem, it beat the published baseline by 40% "in a matter of minutes".

Benchmarks

Google's comparison table pits Argon against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Argon wins most rows. The largest gap is on legal: 19.6% on Harvey's Legal Agent Benchmark, vs 6.7% for the next best model. It also leads on the Vals Index (economic value across finance, coding, legal and tax), DeepSWE v1.1 (77.9%), long context (84.2% on GraphWalks between 256K and 1M tokens), and long video understanding (91.7% on LVBench).

It loses on five: FrontierSWE v2, Terminal-Bench 4.0, PostTrainBench, Terminal-Bench Science and OSWorld-2.0. These are mostly agentic tasks in a terminal, where Opus 5.5 and Astra still lead.

Artificial Analysis measured a 15% hallucination rate on AA-Omniscience, vs 51% for GPT-6 Astra and 54% for GPT-6.1 Sol. Argon doesn't know more facts (its accuracy is actually lower than Astra's). It declines to answer more often when it doesn't know. For agents that run unattended, that matters a LOT.

At the intro price, an Artificial Analysis task costs $1.99, 60% of what Astra costs for the same score. But Argon uses 62k output tokens per task vs 27k for Astra, so once the price doubles, it'll cost about 1.2x more than Astra. And GPT-6.1 Sol is 2.7x cheaper per task for one point less.

Safety approach and rollout

Google's reasoning for the slow rollout covers four areas: misuse (cyber and CBRN), prompt injection, misalignment monitoring, and hardened sandboxes. On misalignment, Google monitors Argon's chain-of-thought and actions and stops execution when needed. It also monitored training runs without feeding the findings back into training, so the model wouldn't learn to hide its reasoning. Google asks the rest of the industry to do the same.

Trusted defenders get Argon WITHOUT cyber guardrails. Through Wiz's Scan for Good program, it already found a critical vulnerability in healthcare software used by hospitals worldwide, one that previous frontier models had missed.

As The Verge notes, the announcement came a day after OpenAI's DevDay (Dots agent, GPT-6.1 Sol), in a week where OpenAI also said it wouldn't release a planned GPT-6.1 Astra model over safety concerns. And Bloomberg reported that Google is dealing with internal skepticism about how well Gemini 4 performs in key areas such as coding.

What Hacker News thinks

The main thread passed 1,600 points and 1,100 comments. Recurring arguments from HN commenters:

  • The most upvoted thread argued that no lab has a moat: labs keep leapfrogging each other, so the "first one to the top wins everything" theory looks wrong. Several commenters drew the practical lesson: keep your workflows model- and provider-agnostic
  • Many were annoyed by an announcement without availability or a date. Some called the "too capable to release" framing a marketing playbook copied from OpenAI. Paying subscribers pointed out that the Gemini app still offered them Gemini 3.6, while 3.7 and 3.8 had already been released
  • Some doubted the benchmarks. Commenters brought up Gemini 3, which benchmarked great but needed constant babysitting in agentic work. Others noted that every lab's chart shows it winning most rows, and that Google compared at maximum reasoning settings
  • Moving 800K+ lines of Zircon to Rust at Google scale got a lot of attention, along with questions about what that means for C++ and Carbon
  • On pricing, some praised the per-token price (5x cheaper than Astra), others pointed out it's a temporary discount and that the cost per task is above Opus 5.5 (high) on the Artificial Analysis charts
  • Others complained about being pushed into Google's own tools (Antigravity), and about how to stop Google from training on your data. The common answer was to use the API
  • Some early testers were positive. One commenter who had access for a few weeks called it the first Gemini model they could hand complex research-heavy tasks to: not 100% reliable, but cheaper to verify than to do the work themselves

In the Artificial Analysis thread, opinions split between "it caught up but doesn't move the frontier" and "it's tied with Astra and Opus 5.5 (high) at a good price, with far fewer hallucinations". Several commenters expect a higher reasoning tier or an Ultra version later.

My take

A third lab at the top means more competition on price and capability.

I can't evaluate a model I can't access. Until then, here's what I'd do:

  • Wait for public API access and independent coding evals before changing anything. Terminal-based agentic coding is where Argon trails in Google's own table
  • Track it for knowledge work: legal, finance, long documents, long videos and charts are where its scores are highest
  • If the 15% hallucination rate holds up in real use, it's the best reason to try it
  • Budget with the post-promotion price of $4 / $20
  • Keep your tools model-agnostic: the leaderboard changed three times in a month

The full breakdown (key facts, benchmarks, positioning vs other models, safety approach, caveats) is in my Gemini 4 Argon note.

That's it for today! ✨

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe