Gemini 4 Argon widely available: Google back among the leaders

On 13 October Google made Gemini 4 Argon widely available. The model was already unveiled on 30 September, but until now it was only accessible to selected cybersecurity teams through Google's Fairwind programme. Now developers and companies can use it through the Gemini API and AI Studio, and consumers with an AI Pro or Ultra subscription in the Gemini app and AI Mode in Search. Ultra subscribers also get it in the programming environment Antigravity.

For Google it is an important moment. Gemini 3.1 Pro dates from February; the promised successor Gemini 3.5 Pro was delayed for months and was eventually replaced by a new model. According to Artificial Analysis, Argon is Google's first new large model in more than seven months. In that time, OpenAI (GPT-6 Astra) and Anthropic (Claude Opus 5.5 and Fable 5.1) pulled ahead.

What does Google itself say?

According to the model card, Argon is a large sparse mixture-of-experts model: a transformer in which only part of the network is active for each piece of text. It was trained from scratch, not refined from an earlier model, and it processes text, images, audio and video. It can handle up to 1 million tokens of input and output combined, and can write an answer of up to 1 million tokens. The previous limit was 64,000. The model's knowledge runs up to March 2026.

In Google's own comparison, Argon beats GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on 13 of the 19 benchmarks, and ties on one. The clearest leads:

• Agents and knowledge work: on AutomationBench (automating tasks) Argon scores 51.3 per cent against 42.5 for Opus 5.5 and 41.4 for Astra. On Harvey's legal agent benchmark the difference is even larger: 19.6 per cent against at most 6.7 for the rest.
• Long context: when searching texts of 256,000 to 1 million tokens (GraphWalks), Argon reaches 84.2 per cent, against 71.8 for Astra and 66.8 for Opus 5.5.
• Science: on LABBench2, which tests biological research work, Argon scores 88.8 per cent against 85.4 for Astra and 73.1 for Opus 5.5. On RiemannBench (mathematics) it is 76.0 against 72.0.
• Programming: on DeepSWE v1.1 Argon reaches 77.9 per cent, against 74.2 for Opus 5.5 and 74.1 for Astra.

But Argon also loses on parts that matter. On FrontierSWE v2, a harder programming test, it scores 55.0 per cent against 65.5 for Astra. On Terminal-Bench 4.0, in which a model has to work independently in a terminal, Opus 5.5 is clearly better with 66.4 per cent than Argon with 57.4. And on Terminal-Bench Science, with clearly defined research tasks, Argon stays more than ten points behind Astra, even though Google gave the model six times as much time as standard.

The method also deserves a caveat. On some of the benchmarks Google ran Argon itself and took competitors' scores from rankings or from their own publications. That makes the comparison less clean.

What do independent measurements say?

The independent figures are more sober. On the Intelligence Index of Artificial Analysis, Argon scores 53, equal to GPT-6 Astra and Claude Fable 5.1, but behind Claude Opus 5.5 with 58. Compared with Gemini 3.1 Pro (30), though, that is an enormous leap for Google itself. On the Agent Arena ranking, where people judge agents in practice, Argon ranks eighth. On Arena's text ranking, based on human preference, Argon was provisionally at the top when it was unveiled.

What stands out is the low hallucination rate: in a test by Artificial Analysis, Argon makes up an answer in 15 per cent of cases, against 51 per cent for Astra. This is partly because Argon more often says it does not know; it does not give the right answer more often.

Where is the real step forward?

Not in raw intelligence, but in three other things.

Price. Until 30 November, Argon costs $2 per million input tokens and $10 per million output tokens; after that $4 and $20. GPT-6 Astra costs $10 and $50. Argon does write longer answers: per task in the Artificial Analysis index it used about 62,000 output tokens against 27,000 for Astra. Per task, Argon now costs $1.99 against $3.26 for Astra; at the normal price that becomes $3.98. So the price advantage is temporarily large and limited after that.

Long context. A million tokens in one answer, and good scores when working with very long texts, make it suitable for tasks such as rewriting entire codebases. As an example, Google mentions a refactor in the Chrome codebase, of around 40 million lines, in which months of work would have been done in weeks.

Science and the real world. Google claims that Argon helped design five new enzymes that proved active in the lab, made a quantum calculation more than a hundred times cheaper, and found counterexamples to a conjecture from nuclear fusion physics. It also learned a skill from video (throwing small bags, as in the game cornhole) that was transferred to Apptronik's humanoid robot Apollo. These are all claims by Google itself, not yet independently confirmed.

Safety

Google says that in its own safety testing it cannot rule out that Argon reaches critical thresholds in the areas of chemical and biological weapons, cyberattacks and situational awareness. That is why extra layers of protection have been built in. The model has been trained to put irreversible steps to a human first, and a monitor keeps an eye on its reasoning and actions. Notably, Google says it deliberately did not penalise the model's train of thought during training, so that it stays honest and checkable.

What does this mean for the road to the singularity?

Argon is not a leap to a new level, but it confirms a pattern: three labs are now close together at the top, and every few months the frontier moves. That makes progress less dependent on one company, and the competition pushes down the price per task. More important is where Argon does stand out: independent work over long periods and large amounts of information, and contributions to real laboratory research. Those are exactly the skills needed for AI that speeds up science, and eventually AI research itself. That Argon lags behind on the hardest programming and terminal tests shows that this last step has not yet been taken.

What is the technological singularity? Read the explanation →