Gemini 4 has arrived: Argon is live, and most people still cannot use it

Gemini 4 landed on September 30, 2026

Google announced Gemini 4 on September 30, 2026, and the first model of the generation is Gemini 4 Argon. It is the company's first new flagship since the Gemini 3 series arrived in November 2025. The ten months in between belonged to the Flash line and to multimodal variants, not to a new top-end model.

What makes the launch unusual is the order of access. Argon was not switched on inside the Gemini app for everyone. Google opened it first to vetted cyber defenders through a controlled early-access program called Fairwind.

Paid API customers and Google AI Ultra subscribers are named as the next groups to get in, and general consumers are given no date at all. Google says the wider rollout follows once the safety guardrails are further refined.

The reported facts in this article come from Google's public announcement of September 30, 2026 and the coverage published on October 1 and 2, 2026.

The flagship slot sat empty for ten months

Gemini 3 Pro arrived in November 2025. Through the first nine months of 2026 Google shipped Flash refreshes, a native Gemini app for Windows, live speech models and agentic video understanding, while the top of the line stayed where it was. OpenAI and Anthropic both kept shipping frontier models during that window.

On July 22, 2026 Sundar Pichai told analysts that the next step up would require much larger base models, and he named coding and agentic coding as the areas where Google had to close a gap. A day earlier, Google's own blog had already confirmed that an ambitious pre-training run for Gemini 4 was under way.

The organization changed as well. In August 2026 Demis Hassabis stepped back from the Google DeepMind chief executive role to become Alphabet's chairman and chief scientist, and Koray Kavukcuoglu took over DeepMind. On September 24 he said Argon had entered post-training and that he hoped for a launch much earlier than the end of the year. Six days later the model was out.

• The Gemini 3 series shipped in November 2025, ten months before Gemini 4.
• Coding and agentic coding were named as the gap Google had to close.
• Argon moved from a post-training statement to a public launch in six days.

Twelve outright wins across eighteen evaluations

Google published a cross-model table with the announcement, comparing Gemini 4 Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Argon leads twelve evaluations outright and ties for first on one more. GPT-6 Astra takes three outright wins and Claude Opus 5.5 takes two.

The lead is broad rather than absolute. Argon trails GPT-6 Astra by 10.5 points on FrontierSWE v2 and on Terminal-Bench Science 0.1, and trails Claude Opus 5.5 by 9 points on Terminal-bench 4.0. The table below puts the biggest wins and the three misses side by side, with the gaps calculated rather than quoted.

Every figure here is Google's own reported result. No independent party has published a replication of the full comparison yet, so read the table as the vendor's case rather than a settled ranking.

Google is already running Argon on its own infrastructure

The most concrete evidence in the announcement is not a benchmark row. It is what Google says the model already does inside the company, where thousands of employees use it for specialised coding, deep research and writing.

A quantum computing team had a subroutine whose cost is measured by the product of qubits and gates. Argon found a new optimisation within minutes that beat the previously published baseline by 40 percent.

A second deployment analysed performance telemetry from Google data centers and carried out memory optimisations automatically. With the full set of changes deployed, Google reports more than 300 TiB of memory released and expects eventual savings between 500 TiB and 1 PiB.

• The C and C++ to Rust migration covers re2, libgav1 and the Fuchsia Zircon kernel, more than 800,000 lines of code.
• For libgav1, Argon rewrote roughly 32,000 lines of SIMD code in safe Rust.
• The rewritten decoder keeps video output identical and runs 2.7 times faster than the earlier Rust port.

These are internal results reported by Google rather than independent measurements, and the announcement gives no external audit of them.

Why cyber defense went first

Security is the capability Google chose to lead with. Argon scores 68 percent on CWE-bench v1, tied for first, and Google says it can discover, verify and patch vulnerabilities on its own. Vetted defenders and internal teams receive a version with some cyber guardrails relaxed so they can reach the full defensive capability.

Wiz, Google's cloud and AI security platform, already uses Argon in a program called Scan for Good that scans critical public infrastructure for free. In an early demonstration the model found a severe flaw in medical software used by hospitals worldwide that earlier frontier models had missed.

Holding the release back is the point rather than a delay for its own sake. Google describes four directions of guardrail work and a runtime mitigation layer that watches the model's reasoning and its actions and stops a task when it drifts away from the user's intent. The same kind of monitoring runs during training, and abnormal behaviour raises an alert to a dedicated incident response team. Google also avoids feeding that monitoring back into training, to reduce the chance of a model slowly learning to evade it.

The security figures are vendor-reported as well, and no third party has reproduced the penetration testing result.

Introductory pricing undercuts the nearest rivals

Google listed introductory API rates of 2 US dollars per million input tokens and 10 US dollars per million output tokens, with cached input discounted by 95 percent. Standard rates after the introductory period step up to 4 and 20 dollars. Consumer plan pricing has not been announced.

Against the published rates of the models in its own comparison table, the introductory tier is the cheapest of the group by a wide margin, and even the standard tier matches Claude Opus 5.5 rather than undercutting it.

Costs on this site always come from its own server-side quote, and the Star Task ledger is the final billing record. No price on this page is a quote for work performed here.

Who gets in, and when

Access moves in phases. Fairwind partners come first, then paid API customers and Google AI Ultra subscribers, and only then developers, enterprises and consumers as Google finishes the safety work. No public date has been given for the free or Pro tiers of the Gemini app.

While the rollout is gated, the useful habit does not change. Read the announcement for direction, then check what your own workspace can actually call. Gemini 3.8 Flash is the model this site's chat workspace runs today, and the status page collects every announced Gemini 4 fact in one place.