On Wednesday, 30 September 2026, Google DeepMind announced Gemini 4 Argon, its new frontier model. The announcement post was signed by Koray Kavukcuoglu, Google DeepMind SVP and chief AI architect. Two things set the launch apart from previous Gemini releases. First, the name: Argon is the first Gemini to carry a codename instead of the familiar Pro/Flash tier labels. Second, almost nobody can use it yet. The model is rolling out first to “trusted cyber defenders” through Google’s Fairwind Program, with paid API customers and Google AI Ultra subscribers promised access later and no date given.
So this post is about a model most developers cannot call today. It is still the most consequential AI release of the week, partly because of what it can do and partly because of how Google decided to release it.
What Google shipped
The headline technical change is the output limit: 1 million tokens per response, up from 64,000 on earlier Gemini models. Google, gHacks and heise all report the figure. (SiliconANGLE described it as a context window, which looks like a mix-up; Google’s own post talks about output.) That is a 15x jump, and the point is not longer essays. A model that can emit a million tokens in one turn can, in principle, produce an entire multi-file migration, a long chain of tool calls, or an extended reasoning trace without the orchestration layer having to stitch many shorter turns together.
Google says Argon was built for “long, multi-step tasks” in three areas: software engineering, enterprise knowledge work, and defensive cybersecurity. That last area includes autonomously finding, validating and patching vulnerabilities.
Pricing (per Google’s post, confirmed by SiliconANGLE and heise):
| Input / 1M tokens | Output / 1M tokens | |
|---|---|---|
| Introductory rate | $2 | $10 |
| Standard rate (after the intro period) | $4 | $20 |
| Cached input | 95% off the input rate | — |
The standard rate is exactly what Anthropic charges for Claude Opus 5.5. The introductory rate undercuts it by half. Google has not said how long the introductory period lasts.
The internal case studies
Google’s most developer-relevant claims come from internal deployments, not benchmarks:
- C/C++ to Rust migration. Argon-driven agents translated code ranging from libraries such as re2 and libgav1 up to the 800,000-line Fuchsia Zircon kernel. According to SiliconANGLE and heise, the libgav1 AV1 decoder work replaced about 32,000 lines of hand-written SIMD code with safe Rust that auto-vectorises. The result decodes 2.7x faster than the previous hand-written Rust port. Google says the rewrites go through automated and manual checks plus emulation tests before production.
- Fleet memory. Analysis across Google’s server fleet freed more than 300 TiB of memory, with projected savings of 500 TiB to 1 PiB.
- Quantum algorithm optimisation, where Google says Argon beat a baseline by 40%.
- A real vulnerability. Wiz, which Google acquired this year, used Argon in its “Scan for Good” work on critical infrastructure. It reports finding a critical flaw that exposed personal data in healthcare software used by hospitals worldwide, a bug earlier frontier models missed. The vendor has not been named, which makes sense if a patch is still in progress.
None of these claims has been independently verified. The Rust migration is still the one to watch. Memory-safe rewrites of legacy C/C++ are among the most expensive problems in the industry, and an 800k-line kernel is not a toy.
The benchmarks: strong, but not a sweep
Google published a table of 19 benchmarks against GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. Outlets count its wins slightly differently (SiliconANGLE says 12 of 18), but the pattern is consistent:
| Benchmark | Argon | GPT-6 Astra | Opus 5.5 |
|---|---|---|---|
| DeepSWE v1.1 (agentic coding) | 77.9% | 74.1% | 74.2% |
| Vals Index (knowledge work) | 68.9% | 63.1% | 67.0% |
| AutomationBench | 51.3% | 41.4% | 42.5% |
| CWE-bench v1 (vuln remediation) | 68.0% | 68.0% | 67.0% |
| FrontierSWE v2 | 55.0% | 65.5% | 62.3% |
| Terminal-Bench 4.0 | 57.4% | 58.2% | 66.4% |
Argon’s wins cluster in enterprise knowledge work, long-context retrieval and video understanding. On GraphWalks at 256K–1M tokens it scores 84.2% against 71.8% for Astra. Its losses cluster in terminal-driven agent work, which is where coding agents actually live: it trails Opus 5.5 by nine points on Terminal-Bench 4.0 and Astra by ten on FrontierSWE v2. The headline cyber number, CWE-bench, is a tie with Astra, not a lead.
Independent write-ups have flagged methodology issues. About half the rows come from third-party leaderboards. The rest mix Google’s own runs with competitor numbers from other sources. On LVBench, Argon reportedly saw video at one frame per second, while competitor scores were produced under different frame budgets. On Terminal-Bench Science, Argon reportedly got a verifier timeout six times the default and still came in behind. None of this invalidates the results. It does mean “state of the art” holds benchmark by benchmark, not across the board.
Why the rollout matters more than the scores
The more important story is the release model. Fairwind opened on 3 September with Gemini 3.8 Flash Cyber and, according to SiliconANGLE, now counts more than 650 organisations, including CrowdStrike and Palo Alto Networks. Those vetted defenders get Argon without its usual cyber guardrails. Everyone else waits while Google “hardens” four layers of safeguards: refusing harmful requests, resisting prompt injection, monitoring chain-of-thought for misalignment, and hardening sandboxes. Google also says it is taking part in the U.S. government’s voluntary pre-release access process, the framework behind the White House accord with frontier labs.
This is a pattern now, not a one-off. A frontier lab ships its strongest model to defenders before attackers can buy it through a public API. Microsoft’s Digital Defense Report this week argued that attackers already had the AI advantage. Staged access is the industry’s answer, and Argon is the clearest example so far. For developers it changes planning. “Released” no longer means “available”, and the gap may now be measured in weeks or months.
Open questions
- When can developers use it, and at what real cost? There is no public model ID and no end date for the introductory pricing. Reasoning-token billing is not documented either. With a 1M-token output ceiling, how reasoning tokens are billed will decide the actual bill.
- Will the safeguards hold? Releasing a model without guardrails to 650+ organisations is a large trust perimeter. Access controls are only as strong as the weakest member’s security.
- No model card yet. At the time of writing, Google’s model-card index had no Argon entry.
- Do the Rust results generalise? Google’s own code, tests and emulators are an unusually good environment for automated migration. Teams with thinner test suites should expect less.
Argon is a strong model with a real edge on long-horizon knowledge work and a credible claim on large-scale code migration. It is not a clean sweep for coding agents. The most important thing about it may be the precedent its release sets.
Sources
- Google — Gemini 4 Argon announcement (Koray Kavukcuoglu)
- SiliconANGLE — Google’s new frontier AI model Gemini 4 Argon goes to cybersecurity defenders first
- heise online — Google’s Gemini 4 Argon focuses on cybersecurity
- SecurityWeek — Google Launches Gemini 4 Argon With Guardrail-Free Access for Vetted Defenders
- gHacks — Google Launches Gemini 4 Argon With a 1 Million Token Output Limit
- DataCamp — Gemini 4 Argon: Benchmarks, Pricing, and Access
- Emergent — Gemini 4 Argon Benchmarks: Every Score, Its Source