Introducing Gemini 4 Argon (limited rollout to cyber defenders)
B · Useful, limited Google DeepMind's launch post for a frontier model released first to vetted cyber defenders; notable for the staged-access pattern, but capability claims and internal results are self-reported.
Summary based on the full text·Original: about 8 minutes to read
In short
Google's announcement (30 September 2026) of Gemini 4 Argon, a frontier model for long-horizon coding, finance and legal work, and cyber defence. Access starts with trusted cyber defenders through the Fairwind Program; developers, enterprises and consumers come later, starting with paid API customers and Google AI Ultra subscribers. Vetted defenders and Google's own teams get the model without cyber guardrails. Google says it is in the US government's voluntary pre-release access process and is adding monitoring of chain-of-thought and actions to stop out-of-bounds behaviour.
Key numbers
- 1M tokensoutput token limit, up from 64K
- 77.9%score on DeepSWE v1.1, long-horizon software engineering
- $2 / $10introductory price per million input / output tokens, rising to $4 / $20 afterwards
- over 300 TiBmemory freed across Google data centres by Argon agents applying optimisations from profiling telemetry
- 2.7xspeed-up of a memory-safe Rust video decoder after Argon agents rewrote 32K lines of SIMD code
As reported by the source.
English·Web page·opens blog.google
What it reports
- Output limit raised to 1M tokens from 64K.
- Benchmarks: DeepSWE v1.1 77.9%; AutomationBench (Zapier) 51.3%, ranked first; LVBench 91.7%; CWE-bench v1 68%, tied for first. Claimed lead on the Vals Index, Vals Finance Agent v2 and Harvey's Legal Agent Benchmark.
- Internal use: a quantum subroutine optimised 40% beyond the published baseline; over 300 TiB of memory freed fleet-wide, with 500 TiB to 1 PiB estimated in total; C/C++ to Rust migrations up to 800K+ lines for the Fuchsia Zircon kernel, under automated and manual audit.
- Wiz used Argon to find a critical vulnerability exposing personal data in hospital software that earlier models had missed.
- Pricing: introductory $2 input and $10 output per million tokens, cached input at 95% off, then $4 and $20.
- Safeguards: refusal of cyber and CBRN misuse, activation monitoring, prompt-injection hardening (leading on Gray Swan's IPI benchmark), misalignment monitors, sealed sandboxes.
How they got it
Google's internal evaluations and deployments, public and partner benchmarks, and a partner case from Wiz. No evaluation details or system card content are included in the post.
Limits
- Self-reported by the vendor; internal productivity figures cannot be checked.
- No comparison tables against named competitor models are given.
- Release is limited, so most readers cannot test the claims yet.
Why it matters for companies
Two frontier labs now release their strongest cyber capabilities first to vetted defenders, with weaker or guarded versions for everyone else. Security teams should expect tiered access, verification programmes and different capability levels by customer status, and plan procurement accordingly. The 1M-token output limit and the internal migration examples point at long, autonomous engineering jobs; the stated audit and review steps around those rewrites are the part worth copying.
Summary based on the full text. Read the original for the full argument and data.Read the original on blog.google