Google Releases Gemini 4 Argon: Flagship Returns After Seven Months, Security Experts First

Seven Months Later, Google Finally Ships a New Flagship

On September 30, Google DeepMind SVP and chief AI architect Koray Kavukcuoglu announced on the official blog: the next-gen flagship model Gemini 4 Argon is here.

It’s Google’s first flagship model update in over seven months. The last was Gemini 3.1 Pro. Per media timelines, the planned 3.5 never shipped at all.

Argon’s positioning is specific: long-horizon software engineering, enterprise knowledge work (legal, finance), and cybersecurity defense. Google says releasing safely at this capability tier requires staging.

First Users Are “Gatekeepers,” Not You

On launch day, ordinary users couldn’t touch Argon. First access went only to trusted cybersecurity defenders in the Fairwind program.

Google has the rollout order planned: paid API users and Google AI Ultra subscribers next. No timeline given; the free tier wasn’t mentioned.

One more detail: Google says it’s participating in the US government’s voluntary pre-release model access process. Government and security experts get a look before the market does.

The Biggest Change: 1M Tokens in a Single Output

Argon outputs up to 1 million tokens in a single response. The previous generation capped at 64K. Long reasoning chains and long code rewrites no longer get cut off midstream.

Benchmarks are Google’s own: DeepSWE v1.1 (software engineering) 77.9%, CWE-bench v1 (vulnerability repair) 68% tied for first, AutomationBench (enterprise automation) 51.3%, LVBench (long-video understanding) 91.7%.

Third-party tester Artificial Analysis gave Argon 53 on its Intelligence Index, matching GPT-6 Astra’s (full-context) top score. In Google’s own comparison table, Argon leads GPT-6 Astra, Claude Fable 5.1, and Opus 5.5 on 13 of 19 items.

Don’t take the numbers at face value. Bloomberg reported Google employees with direct model access saying real-world coding performance doesn’t match what the benchmarks imply. Google responded that the characterization is inaccurate. Both claims are on the record — judge for yourself.

On guardrails, Google lists four: refusing CBRN and cyberattack misuse (including monitoring internal model activations), resisting indirect prompt injection (0.7% success rate after 15 attack attempts), monitoring chain-of-thought and actions, and sealed sandboxes for high-risk training and evals. All required coursework before general release.

Pricing: Half-Price at Launch, but Only Temporarily

Argon’s launch pricing: $2 per million input tokens, $10 output. Cached input is 95% cheaper at $0.10. After the launch window, it rises to $4 and $20.

Compare: Claude Opus 5.5 lists at $4 and $20. Argon’s launch price is exactly half. The price war continues.

Real use cases already exist. Security company Wiz used Argon through the free Scan for Good program to find a critical vulnerability leaking sensitive personal data in hospital software. Google says no earlier flagship model caught it.

Google uses Argon internally too: agents cleaned out over 300 TiB of data-center memory, migrated 800,000+ lines of Fuchsia Zircon kernel code from C/C++ to Rust, and rewrote the libgav1 decoder 2.7x faster. All Google’s own numbers.

What to Watch Next

Two dates to track: when paid API users and Ultra subscribers get access, and how long the launch pricing window lasts. A million-token response costs $10 in output fees now — $20 after the window. Budget on long-term pricing.

And some context: in the same week, OpenAI halted GPT-6.1 Astra over failed safety testing. Two leading labs — one cutting a flagship for safety, one staging a flagship release for safety. Oddly aligned directions.