Everything we know about Google's Gemini 4 Argon

Why it matters

The competitive claims currently help Google more than buyers: without API access or independent replication, developers cannot test whether Argon's coding and long-output gains hold up in production. Routing early access through cyber defenders lets Google gather safety evidence, but delays startups' ability to build against the model. Teams planning around the $2/$10 introductory rates should account for the stated $4/$20 post-intro pricing, which is 2.5 times higher.

Status, October 8th, 2026: Announced. Not generally available. Confirmed access is Google internal teams plus vetted cyber defenders in the Fairwind Program. Paid API and Google AI Ultra are the named next tier. Google has not dated that tier.

Argon is the first Gemini 4 model, and the first new frontier Gemini since 3.1 Pro. Google skipped 3.5 Pro over the summer and shipped 3.8 Flash instead. The public name is Gemini 4 Argon. Koray Kavukcuoglu, SVP of Google DeepMind and Google’s chief AI architect, wrote the announcement. Sundar Pichai posted the early look the same day.

The access reality

What X is saying today

Treat all of this as rumor until a model card or API changelog lands.

• Oct 9th–10th window. The most repeated claim this week is paid API and Ultra access around October 9th or 10. No Google account has confirmed it.
• “Launching tonight.” Low-engagement posts are calling a drop between 11 p.m. and midnight IST on October 8th. Nothing in official channels matches that.
• Thinking tiers. Screenshots circulating today show Argon as the first Gemini with xhigh and max effort options, plus selectable context windows, with 256K described as the normal quota. Unconfirmed.
• Plan reshuffle. A separate rumor says free tiers lose Flash and Pro, Plus loses Pro, and Argon hits Ultra first, then Pro. Also unconfirmed.
• Cancellation rumor. One post claims Argon was killed internally in favor of a Gemini 4.5 press release in two months. Single low-engagement claim. It conflicts with the Sept 30th blog, Fairwind rollout, and published pricing. Weight it near zero unless Google says otherwise.

A few Pro users have posted app screenshots this week. Those are consistent with a small early-access test, not a general rollout.

What the model is for

Google’s framing is long-horizon work, not a chat upgrade. Three named jobs:

• Real-world software engineering
• Enterprise knowledge work, especially legal and finance
• Defensive cybersecurity: find, validate, and patch vulnerabilities

Secondary claims cover creative writing, chart reading, long video, and actions over a series of documents.

The spec that changes workflows is output length. Argon can generate up to 1 million tokens in one response, up from 64K on earlier Gemini models. Google’s argument is that a migration or a long analysis needs room to think and write in one trajectory. Input context was not clearly specified in the launch post. Some trackers list 1M input. Today’s X screenshots imply selectable windows rather than one fixed number.

Architecture, parameter count, active parameters, and knowledge cutoff are unpublished.

Benchmarks Google published

All figures below are Google’s, from the Sept 30th post and the chart Sundar posted. Several rivals were scored by Google. Independent replication is not out, because almost no one outside Fairwind can run the model.

Google also says Argon beats 3.8 Flash Cyber on an internal vulnerability-discovery set across 20 languages, and on Wiz’s black-box web pentest benchmark. Wiz is an early Fairwind deployer and has said it found significant flaws in widely used healthcare software.

What Google says it already did inside Google

These are internal, unaudited claims. They indicate Google's intended use cases but are not yet evidence of results outside the company.

• Quantum. Researchers gave Argon a subroutine. In minutes it beat the published baseline by about 40% on qubits and operations.
• Data centers. A team of Argon agents read server telemetry and freed over 300 TiB of memory, with 500 TiB to 1 PiB projected once rolled out.
• Migrations. C and C++ to Rust, including up to 800,000+ lines of the Fuchsia Zircon kernel, still under automated and manual audit before production.
• libgav1. Replaced 32,000 lines of SIMD. Google says the Rust port is 2.7× faster than the existing Rust port, with identical video output. Comparison to the original C++ is not stated.

Pricing

Introductory, for the public launch whenever it happens:

• $2 / million input tokens
• $10 / million output tokens
• Cached input at 95% off the input rate, so about $0.10 / million during the intro window

After the introductory period: $4 input / $20 output per million. That matches Anthropic’s Opus 5.5 list price. Google has not said how long the intro lasts. Startups that price a product on the $2/$10 rate are baking in a future margin cut.

Why it is gated

Google’s stated reason is capability, not a soft launch.

• Frontier Safety Framework. Refuses harmful cyber and CBRN requests, allows dual-use research.
• Separate monitors on chain-of-thought and actions that can stop a run if it leaves the user’s intent.
• Claims the lead on Gray Swan’s indirect prompt-injection benchmark, via automated red-teaming and adversarial training.
• Sandbox isolation for high-risk training and evals, per Google’s agent-control roadmap.
• External and internal red teams, manual and automated.

The U.S. voluntary pre-release process is running in parallel. That is why Fairwind comes before the API.

What we still do not know

• Public date for API or Ultra
• Whether the Gemini app gets it in the same wave
• Input context window, and whether thinking effort is a user control
• Intro-pricing duration
• Independent benchmark replication
• Whether the 300 TiB and Zircon migrations survived audit
• Rate limits, latency, and tool-use behavior outside Google

Bottom line

Argon has a real Sept 30th announcement, a 1M output ceiling, and a benchmark table that puts it ahead of Opus 5.5 and GPT-6 Astra on the tests Google chose to show. It is still a cyber-defender preview: Argon is in Fairwind, priced on paper at $2/$10, and the public tier is what X is guessing at.

If the Oct 9th–10th rumor is right, the next thing to check is the model card and the API name. Until then, nothing outside Fairwind is testable.