Gemini 4 Argon Opens as a Fairwind-Gated Cyber Frontier — Guardrails Come Later
Google DeepMind’s Koray Kavukcuoglu named Gemini 4 Argon for long-horizon software engineering, legal/finance knowledge work, and cyber defense. Trusted Fairwind defenders and U.S. voluntary pre-release get it without cyber guardrails first; paid API and AI Ultra wait on safeguard iteration. Lab-claimed benches lead DeepSWE and AutomationBench; CWE-bench v1 ties peers at 68%. No independent Arena Elo yet.
Google just named the frontier SKU that the September 28 post-training tease only hinted at — and the access order is the story: unguarded for vetted defenders first, guarded public API later.
Google DeepMind SVP and Google chief AI architect Koray Kavukcuoglu announced Gemini 4 Argon on September 30 as the lab’s next frontier model for long-horizon software engineering, enterprise knowledge work (legal and finance), and cybersecurity defense. Initial access runs through the Fairwind program for trusted cyber defenders and the U.S. government’s voluntary pre-release process; Google says internal teams already use it. Broader availability for paid API customers and Google AI Ultra subscribers is planned only after further guardrail iteration. This is the named Argon launch — not a restatement of the earlier “Gemini 4 in post-training” interview tease.
Access and pricing
| Cohort | Cyber guardrails | Status |
|---|---|---|
| Fairwind trusted defenders + Google internal | Off (full defensive cyber capability) | Rolling out now |
| U.S. voluntary pre-release testers | Phased / feedback loop | Active |
| Paid API + AI Ultra (developers, enterprises, consumers) | On (misuse/CBRN refusals + monitors) | After iteration |
Introductory API pricing: $2 / $10 per million input/output tokens, with cached input at 95% off input. After the intro period: $4 / $20. Output context expands to 1M tokens (from 64K) so the model can sustain hundreds of thousands of tokens in a single trajectory.
What Google claims it can do
Company-reported highlights (not independent Arena Elo):
| Bench | Claimed result | Caveat |
|---|---|---|
| DeepSWE v1.1 | 77.9% (SOTA) | Lab figure; long-horizon SWE |
| AutomationBench (Zapier) | 51.3%, #1 | Lab figure |
| LVBench (long video) | 91.7% SOTA | Lab figure |
| Vals Index / Vals Finance / Harvey Legal | “Leading” | Lab framing; sector weights claimed |
| CWE-bench v1 | 68%, tie for first | Google: tie at 68%; SecurityWeek names peers GPT-6 Astra and Grok 4.7 |
Google also cites internal wins: quantum subroutine spacetime cut 40% vs a published baseline in minutes; Argon agents claimed >300 TiB memory freed so far (est. 500 TiB–1 PiB); Rust migrations including libgav1 SIMD rewrite claimed 2.7× faster than a prior Rust port with identical video output — all company anecdotes, not third-party audits.
Defender path and the hospital-software claim
Building on September’s 3.8 Flash Cyber / Fairwind split, Google trained Argon for autonomous find/validate/patch of critical vulnerabilities and will release it without cyber guardrails to trusted defenders and internal teams. Wiz (via Scan for Good) used Argon to uncover a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide — a risk Google says prior frontier models missed. Neither Google nor SecurityWeek names the software or states remediation status.
On discovery vs 3.8 Flash Cyber, Google claims gains on an internal multi-language vuln bench (20 languages) and on Wiz’s internal black-box pen-test bench (attack surface, vulns, PoC evidence). Those remain vendor/partner internal scores.
Safeguards before broad GA
Before public API / Ultra, Google lists four safeguard tracks: misuse and CBRN refusals with activation monitoring (Frontier Safety Framework); leading claim on Gray Swan IPI for indirect prompt injection; chain-of-thought / action misalignment monitors that can stop execution; and hardened, sealed sandboxes for high-risk evals. Dual-use framing is explicit: unguarded for vetted defenders, refused for offensive cyber/CBRN on the public path.
Claims vs checks
All headline benches above are Google-reported. SecurityWeek (October 1) restates Fairwind access, the unnamed hospital vuln, and the CWE-bench peer names. Independent LMArena / Chatbot Arena Elo and Artificial Analysis listings for Argon were not verified for this piece — no independent ranks are asserted here. Open questions include a named CVE/vendor for the hospital claim, whether cyber guardrails differ by SKU at GA, and third-party CWE-bench / Vals / Arena replication.
Limits
- Benchmarks and internal productivity anecdotes are lab-claimed only; no independent Arena Elo in sources used.
- Hospital-software product is unnamed; remediation unknown.
- CWE peer names (Astra / Grok 4.7) come from SecurityWeek, not the Google blog’s own wording.
- Dual-use: unguarded defender access is intentional policy, not a leak — treat carefully; this is not an exploit how-to.
- Distinct from any September 28 post-training tease coverage; this slug is the Argon launch only.