Google DeepMind unveils Gemini 4 Argon, its new frontier model - Dealroom
High confidence: full text extraction produced 2688 characters.
What's the deal? On 30 September 2026, Google DeepMind announced Gemini 4 Argon, the new frontier model from Alphabet | Google, built for long-horizon work across software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defence. It is not generally available yet: access is limited to trusted cyber defenders through the Fairwind Program, launched on 2 September 2026 with more than 650 partners, including government agencies, critical infrastructure operators and security firms. Paid API customers and Google AI Ultra subscribers come next. Google is also taking part in the US government's voluntary pre-release model access process. The model launches at an introductory price of $2 per million input tokens and $10 per million output tokens (cached input 95% off), rising to $4 and $20 after the introductory period. The output token limit rises to 1M tokens, up from 64K.
Benchmarks: Google's own table puts Argon ahead of OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Claude Opus 5.5 on most tests: 77.9% on DeepSWE v1.1 (Astra 74.1%, Opus 5.5 74.2%), 68.9% on the Vals Index, 51.3% on Zapier's AutomationBench, 19.6% on Harvey's Legal Agent Benchmark and 91.7% on LVBench. It ties for first on CWE-bench v1 (68%). It trails on FrontierSWE v2, Terminal-bench 4.0, PostTrainBench, Terminal-Bench Science and OSWorld-2.0. The official numbers are well below the figures in a leaked table that circulated in late September 2026 (for example 88.7% on DeepSWE v1.1).
Cyber and internal use: Argon can find, validate and patch software vulnerabilities on its own, and trusted defenders get it without cyber guardrails, standalone or inside CodeMender. Wiz used it through its Scan for Good initiative to find a critical flaw exposing personal data in healthcare software used by hospitals worldwide, which earlier frontier models had missed. Inside Google, Argon agents freed more than 300 TiB of data-centre memory (500 TiB to 1 PiB expected in total), beat a published quantum-subroutine baseline by 40%, and are porting C/C++ codebases to Rust, up to the 800K+ line Fuchsia Zircon kernel. A Rust rewrite of the libgav1 video decoder now runs 2.7x faster than the earlier port.
Why it matters: Argon is Google's first new flagship since it cancelled Gemini 3.5 Pro, and it brings Google back into the race at the frontier. The gated, defenders-first rollout follows similar moves by Anthropic (Project Glasswing) and OpenAI (Daybreak). Google says it is strengthening safeguards against misuse, prompt injection and misalignment before a wide release.
Read more: Google DeepMind on X · Google · Google (Fairwind) · 9to5Google · Techzine