Frontier claims are not a checkout button.
On September 30, 2026, Koray Kavukcuoglu announced Gemini 4 Argon as Google’s new frontier model: long-horizon workflows across software engineering, enterprise knowledge work (legal, finance), and cybersecurity defense. The same post rolls Argon out to “a set of trusted cyber defenders through our Fairwind Program,” states active engagement in the U.S. government’s voluntary process for pre-release model access, and quotes introductory pricing of $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. After the introductory period, $4 / $20 applies. Output token limit expands to 1 million tokens, up from 64K (Google Blog, September 30, 2026).
That is a lot of product shape for a surface most operators cannot touch this week. Fairwind is the gate. Paid API customers and Google AI Ultra subscribers are named as the later expansion path, “as soon as possible,” without a dated GA (Google Blog, September 30, 2026; Fairwind Program page, accessed October 2, 2026).
If your eval ticket says “stand up Argon on Ultra / API and bench it against Opus and Astra,” you are staffing a queue that does not exist yet for you. This review is primary-source only. No hands-on. Evidence limits are part of the product truth.
What Google shipped on paper
Argon is framed for deep reasoning across complex, long-horizon workflows. Internal Google examples in the launch post include quantum algorithmic optimization (one example beating a published baseline by 40% in minutes), fleet memory efficiency work (over 300 TiB freed once rolled out, with estimated 500 TiB to 1 PiB total savings), and large C/C++ to Rust migrations including profile-guided SIMD rewrites on libgav1 that Google says run 2.7x faster than a prior Rust port with identical video output (Google Blog, September 30, 2026).
Benchmark claims Google published in that post (treat as vendor-reported): DeepSWE v1.1 at 77.9% (state of the art for long-horizon software engineering in Google’s framing); leading on the Vals Index; leading on Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark in Google’s description; AutomationBench (Zapier) #1 at 51.3%; LVBench long video understanding at 91.7% SOTA; CWE-bench v1 at 68%, tied for first. Fairwind / internal cyber configurations can run without cyber guardrails for trusted defenders. Wiz is named as already using Argon through Scan for Good, with a claimed critical healthcare software vulnerability find prior models missed (Google Blog, September 30, 2026).
InfoWorld’s October 1 independent piece adds comparative numbers and buyer caution. It reports Argon at 68.9% on the Vals Index versus Claude Opus 5.5 at 67.0%, DeepSWE v1.1 at 77.9% versus 74.2% for Opus 5.5, and PostTrainBench where Opus 5.5 leads 49.3% to Argon’s 45.3%. On CWE-bench v1 it restates the 68% tie with Grok 4.7, GPT-6 Astra, and Claude Opus 5.5. Pareekh Jain (EIIRTrend / Pareekh Consulting) told InfoWorld to treat scores as hints not proof, and to build business cases on the eventual $4/$20 price, not the intro rate. He compared intro pricing favorably to Opus 5.5 at $4/$20 and GPT-6 Astra at $10/$50, and advised evaluating on own workloads rather than switching outright (InfoWorld, Nidhi Singal, October 1, 2026).
Tech Insider (Marcus Chen, October 1) restates the access table cleanly: internal Google and Fairwind live; paid API and Ultra planned with no date; public/free not announced. It flags the absence of a full public head-to-head table versus GPT-6 Astra or Claude Sonnet 5.5 for independent verification, and notes Fairwind members can pair Argon with CodeMender. It also connects the gated sequencing to Google’s broader cyber-first posture after earlier 2026 testing incidents reported elsewhere. Use Tech Insider for access framing and the missing-scorecard caution. Use Google primary for exact feature and price claims. Use InfoWorld for independent pricing and eval advice (Tech Insider, October 1, 2026).
Fairwind is the product you can actually buy into (if eligible)
Google DeepMind’s Fairwind Program page (accessed October 2, 2026) is the access doc. It gives high-priority defenders early access so they get a head start against cyber threats before general availability. Priority partners: governments and national cyber authorities; critical infrastructure operators (healthcare, telecom, energy, financial networks); core technology platforms. The page says Google currently works with over 650 partners globally. A set of Fairwind partners get exclusive access to Gemini 4 Argon and can use it with CodeMender. Participating orgs agree to user-level authentication, phishing-resistant MFA, access controls, and may only grant Argon to internal cybersecurity, incident response, or penetration testing teams, with tracking of employee access. Permitted dual-use tasks include authorized threat simulation, reverse engineering, and malware analysis for defensive and academic research. Malicious tasks such as creating malware are not permitted. Access is managed; redistribution or resale is not allowed. Zero data retention is supported when accessed as a managed model on Gemini Enterprise (Fairwind Program page).
That is not Ultra. That is not a developer API key. That is a vetted defender program with an apply form and due diligence. Non-eligible orgs are pointed to CodeMender with publicly available models and other Google AI Threat Defense products (Fairwind Program page).
Google’s launch post is explicit about why the gate exists: safely releasing frontier capabilities requires a phased approach; U.S. government voluntary pre-release engagement is active; guardrails iterate with early tester feedback before developers, enterprises, and consumers get access (Google Blog, September 30, 2026).
Pricing without a cart
Introductory price: $2 / $10 per million input / output tokens. Cached input: 95% off input token price. After intro: $4 / $20. Google has not specified when the introductory period ends (Google Blog, September 30, 2026; InfoWorld, October 1, 2026).
InfoWorld’s Jain line is the operator rule: build the business case at $4/$20. Intro pricing is competitive marketing, not a permanent COGS assumption. Cost per successful business outcome beats cost per token (InfoWorld, October 1, 2026).
Until paid API and Ultra open, those price lines are planning inputs for a future SKU, not this week’s invoice.
Evidence limits (read before you brief)
This review did not run Argon. Fairwind and Google internal are the live surfaces. Independent third-party benches on a public endpoint are not available yet; Tech Insider flags the missing published comparison table against named rivals (Tech Insider, October 1, 2026). Vendor benchmarks and internal Google anecdotes are real claims with dates. They are not your workload. InfoWorld’s mixed PostTrainBench result is a reminder that “frontier” is not uniform across every task class (InfoWorld, October 1, 2026).
Do not invent Ultra availability. Do not claim hands-on. Do not paste Google’s quantum or memory-TiB stories into a client deck as proof of your roadmap.
What works / what breaks / who it is for
What works on paper: long output headroom (1M tokens), vendor-reported strength on long-horizon SWE and several knowledge-work benches, cyber defense tooling for Fairwind with CodeMender pairing, intro pricing that looks aggressive versus named premium rivals if and when you can buy it (Google Blog; InfoWorld).
What breaks for most teams this week: no public API, no Ultra seat, no independent confirmation path, unclear intro-period end date, and a Fairwind eligibility fence that excludes ordinary developers and most marketers (Google Blog; Fairwind; Tech Insider).
Who it is for now: organizations that clear Fairwind criteria (government, critical infrastructure, core platforms, defensive academic labs as described) and can staff authenticated cyber IR / pen-test use under Google’s terms. Who should wait: everyone else planning a general coding or knowledge-work swap this week.
What operators will get wrong
Treating the blog post as GA. It is not. Fairwind first (Google Blog; Fairwind).
Staffing an Ultra or paid-API eval this week. Those surfaces are named as next, not live (Google Blog; Tech Insider).
Building forecasts on $2/$10 forever. Jain’s $4/$20 planning rule is the diligence item (InfoWorld).
Copying vendor benches into a bake-off without your own tasks. PostTrainBench already shows a loss to Opus 5.5 on that slice (InfoWorld).
Ignoring the no-cyber-guardrails Fairwind configuration when briefing security leadership. That is a deliberate defender design, and it is exactly why broad release is gated (Google Blog; Tech Insider).
Confusing CodeMender-with-public-models (available guidance for non-Fairwind orgs) with Argon access (Fairwind page).
What to do this week
If you might be Fairwind-eligible, apply with a real defender brief: which IR / pen-test teams, MFA posture, permitted dual-use tasks, CodeMender pairing plan, and zero-retention needs on Gemini Enterprise (Fairwind Program page).
If you are not eligible, do not open an “Argon pilot” ticket. Open a “watch Fairwind expansion + paid API / Ultra” ticket with a dated revisit. Put Jain’s eval grid on your existing stack: task success, accuracy, agent reliability, cost per successful outcome, latency, security/governance, privacy, integration, lock-in (InfoWorld).
Price the future SKU at $4/$20 in any TCO model. Treat $2/$10 as introductory marketing until Google dates the end of intro (Google Blog; InfoWorld).
Separate cyber-defender interest from marketing/GTM interest. Marketers should track when Ultra and API open, not pretend Fairwind is a growth channel. Security teams should track Fairwind + CodeMender, not confuse them with consumer Gemini features.
Keep a one-page evidence limits note in the eval folder: primary sources only, no hands-on, vendor benches, no public head-to-head table yet (Tech Insider).
The announcement is not the access. Argon’s launch post is a capability and pricing preview wrapped around a defender gate. Staff the gate you can enter, or staff the watchlist for the SKU you cannot buy yet. Do not staff a phantom Ultra bench and call it diligence.
