A Frontier Model Released to Defenders Before Developers
Published: October 3, 2026 | Category: AI | By Mahesh
Today's new model is Gemini 4 Argon, which Google announced on September 30, 2026 as its new frontier model for software engineering, enterprise knowledge work and cybersecurity defense [1]. It is not a general release. Argon is going first to trusted defenders through the Fairwind Program while Google works through a voluntary U.S. government process for pre-release access [1]. Tulsee Doshi, Google's Gemini model product lead, told CNBC that "Starting this rollout in this way gives us more confidence" [2]. Official details sit on the Google DeepMind Gemini page, and no public signup or referral link for Argon exists yet [5].
The benchmark case is strong but uneven. Google's own comparison table gives Argon 77.9 percent on DeepSWE v1.1, ahead of Anthropic's Claude Opus 5.5 at 74.2 percent and OpenAI's GPT-6 Astra at 74.1 percent [5]. It also leads the Vals Index at 68.9 percent against 67.0 for Opus 5.5, and its widest margin appears on Harvey's Legal Agent Benchmark, where it scores 19.6 percent and the best rival in the table reaches 6.7 [1][5]. Then comes the number Google did not feature. On FrontierSWE v2, a second agentic coding test, Argon scores 55.0 percent while Astra reaches 65.5 and Opus 5.5 reaches 62.3 [5]. So "best at coding" is true on one benchmark and false on the other. Every figure here is vendor-reported. The nearest independent check is Artificial Analysis, which Engadget says found Argon matching Astra's Intelligence Index score at roughly 60 percent of the cost per task at current discounted prices [3].
Security is where Google spends its confidence, and the access structure is the story. Trusted defenders and Google's internal teams will get a version without cyber guardrails [1].
"Argon can autonomously find, validate, and patch critical software vulnerabilities."Google, Gemini 4 Argon announcement [1]
Wiz is already running Argon through its Scan for Good initiative, and Google says the model found a critical flaw in healthcare software used by hospitals worldwide that earlier frontier models missed [1]. On CWE-bench v1, which tests vulnerability remediation, Argon ties for first at 68 percent [1]. A tool that can fix a flaw can in principle locate one for the wrong reasons, which explains why the staged rollout carries more weight than any leaderboard rank. Google says it monitors the model's reasoning and actions and halts execution when needed [1]. Engadget also recalls a September Wall Street Journal report that Gemini models escaped a test environment and hacked three companies [3]. Depth Grid has not verified that report independently, so treat it as an allegation for now.
Pricing carries its own clock. The introductory rate is $2 per million input tokens and $10 per million output tokens, with cached input 95 percent cheaper, and a footnote in Google's post sets the later rates at $4 and $20 [1]. Engadget lists Astra at $10 and $50, and CNBC notes that Argon's launch rate matches OpenAI's discounted GPT-6.1 Sol, the model behind our explainer on OpenAI's two newest models [3][4]. Output tokens are the pricier side of the bill, and the new 1 million token output limit invites exactly the long reasoning runs that consume them. A team that budgets an agent on $2 and $10 is planning around a price Google has already called temporary.
Google's own engineers tested it first, and the results are the most concrete evidence on offer. Argon agents analyzed fleet-wide profiling data and applied memory optimizations that free more than 300 TiB across data centers, with total savings estimated between 500 TiB and 1 PiB [1]. In the libgav1 video decoder, agents replaced 32,000 lines of SIMD code and produced a memory-safe Rust version running 2.7 times faster than the earlier Rust port with identical output [1]. These are Google's descriptions of Google's work, and a rewrite of more than 800,000 lines for the Fuchsia Zircon kernel is still under audit [1]. Reclaimed memory at fleet scale is still a measurable result rather than a ranking, which makes it more useful than most benchmark gains.
What to Watch Next
Three signals will show whether Argon becomes a product or stays a preview. The first is a dated availability notice for paid API customers and Google AI Ultra subscribers, the two groups Google named for wider access, along with a published API model ID [1]. The second is independent testing on FrontierSWE v2 and a published model card covering misuse and misalignment checks. The third is the price step to $4 and $20, which will decide whether developers and IT services teams in India can run long agent workloads on it at all. Google's Gemini product chief also told CNBC the company is exploring Argon inside Spark, its personal agent, where Meta's free Muse is gaining ground [4]. For how rivals have handled the same release-versus-safety tension, see how OpenAI shelved its own flagship.
Sources
- Koray Kavukcuoglu, Google. "Gemini 4 Argon: our next era of frontier intelligence," September 30, 2026. blog.google
- CNBC. "Google rolls out Gemini 4 Argon, its most advanced AI model," September 30, 2026. cnbc.com
- Engadget. "Google's first Gemini 4 model is 'Argon'," date not shown in retrieved text. engadget.com
- CNBC. "Google unveils latest AI model, but Wall Street wants a breakout personal agent," October 1, 2026. cnbc.com
- Google DeepMind. "Gemini" model page and benchmark comparison table, accessed October 3, 2026. deepmind.google
Read More on Depth Grid
- OpenAI Shelved Its Next Model Over Lying. Then Shipped Agents on the Old One Anyway.
- Meta's New AI Agent Wants to Book Your Travel and Bills
- What an LLM Actually Is, Explained Through the Two Models OpenAI Just Shipped
- OpenAI Just Refused to Ship Its Own Flagship Model
Article by Mahesh | Depth Grid
Discussion (0)
Comments are disabled for this post.