Loading page…
Paid API customers and Google AI Ultra subscribers are next in line, but Google has not set a date. Selected defenders will get the model without cyber guardrails.
Listen to this story
Google is using internal engineering work and a defender-only external rollout to test Gemini 4 Argon before broader access. The model has supported tasks including a memory optimization Google says freed more than 300 TiB and a Rust migration spanning over 800,000 lines, but the company says critical code still requires audits, testing, and review. A new 1-million-token output limit is intended to support longer runs; it is not an input-context claim. Paid API customers and Google AI Ultra subscribers are next, with no availability date announced.
Wiz is testing Argon through Scan for Good; Google says an early demonstration found a critical healthcare-software flaw exposing personal information.
Google reports scores of 77.9% on DeepSWE v1.1, 51.3% on AutomationBench, and 68% on CWE-bench v1; these results are not independently assessed.
Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached inputs priced 95% below standard rates.
Google’s next frontier model is reaching cybersecurity defenders before ordinary customers. On September 30, 2026, Google introduced Gemini 4 Argon through a restricted rollout aimed at trusted defenders, while it strengthens safeguards for wider access. The model targets software engineering, legal and financial work, and finding and fixing security flaws.
The release follows substantial use inside Google. The company says thousands of employees have used Argon for specialized coding, deeper research and writing. Its announcement describes agents working on engineering tasks that range from memory optimization across data centers to large code migrations.
One team of agents analyzed fleet-wide performance data and applied memory optimizations that Google says freed more than 300 TiB of memory after rollout. Other agents are migrating C/C++ software to Rust, including work spanning more than 800,000 lines of the Fuchsia operating system’s Zircon kernel.
Those rewrites are not being treated as ready-made production code. Google says critical systems undergo automated and manual audits, testing in emulated environments, and review before deployment. Argon’s role is to carry more of the engineering work, not eliminate those checks.
To support longer tasks, Google raised the output limit to 1 million tokens from the previous 64,000-token limit it cites. Tokens are the units used to measure model input and output. This is an output allowance—not a claim about how much source material the model can accept. Google says the extra room lets Argon sustain deeper reasoning in one run.
The initial rollout runs through Google’s Fairwind Program. Google says Argon can autonomously find, validate and patch critical software vulnerabilities. For trusted defenders and its own internal teams, the company will release the model without cyber guardrails so they can use its full defensive capabilities.
Wiz is already using Argon through Scan for Good, its initiative to find and fix high-risk exposures in critical public infrastructure for free. Google says an early demonstration uncovered a critical flaw exposing personal information in healthcare software used by hospitals worldwide—a vulnerability previous frontier models had missed.
Google’s published scores include 77.9% on DeepSWE v1.1, which tests extended software-engineering tasks, and 51.3% on Zapier’s AutomationBench, which measures end-to-end business execution. It also reports a joint first-place score of 68% on CWE-bench v1, a vulnerability-remediation test. These are Google-reported results, not an independent assessment of Argon’s performance.
Google says it is participating in the U.S. government’s voluntary pre-release model-access process and gathering early tester feedback. Its safety work includes several distinct controls:
For developers planning ahead, Google lists introductory API prices of $2 per million input tokens and $10 per million output tokens. Cached input tokens are priced 95% below the standard input rate. The announcement gives no end date for the introductory pricing.
Broader availability will start with paid API customers and Google AI Ultra subscribers, Google says. It has not announced a date for that step. For now, the external rollout remains focused on trusted defenders and testers whose feedback will shape the systems surrounding the model.
Story updates
Google is promising an AI model for difficult, long-running professional work. Its September 30, 2026 announcement of Gemini 4 Argon comes with a concrete result: Vals AI ranks it first among 41 models on the Vals Index, at 68.9%. But the lead is not universal across tasks, and access is starting with trusted cybersecurity defenders—not the developers and consumers Google eventually intends to serve.
The Vals Index covers finance, coding, legal and tax work, with sectors weighted by their contribution to U.S. GDP, according to Google. Vals’ evaluation places Argon ahead of Claude Sonnet 5.5 at 67.04%, Claude Opus 5.5 at 66.97% and Claude Fable 5.1 at 65.83%. That establishes an aggregate lead in this evaluation, rather than a clean sweep of the underlying professional tasks.
Argon’s Vals Index result cost $15.68 per test, below all three Claude models immediately behind it. Sonnet cost $21.34, Opus $32.14 and Fable $28.71. The advantage is therefore not just the highest score: in this test, Argon also costs less than its nearest ranked competitors.
Those test costs should be kept separate from token rates. Vals lists $4 per million input tokens and $20 per million output tokens. Google says Argon will launch at introductory rates of $2 and $10, respectively. The two sources describe different pricing terms; Google’s introductory offer is not the rate listed on Vals’ model page.
Argon has clear strengths within the suite. It takes first place on Finance Agent v2 at 65.40%, ahead of Gemini 3.8 Flash at 61.44%. It also ties GPT-6 Astra for first on IOI with a perfect 100%. Those wins sit alongside second-place finishes in application coding, code migration, tax and cybersecurity.
Coding: Argon scores 91.91% on Vibe Code Bench, behind Sonnet 5.5 at 92.39%. On Code Migration, its 68.17% also trails Sonnet’s 69.83%.
Cybersecurity: Argon scores 77.86% on CyberBench, placing second among 43 models and just 0.12 percentage points behind GPT-6 Sol.
Weaker finishes: Vals ranks Argon 15th of 106 on MedScribe at 87.43%, and seventh of eight on CUA-bench at 4.83%.
The legal results also require care. Google describes leading performance on Harvey’s Legal Agent Benchmark, which it identifies as legal research and drafting. Vals’ published result instead puts Argon fifth of 73, at 19.58%. That difference makes the named benchmark and published score more useful than the launch post’s broad description of legal strength.
Google says it raised Argon’s output limit to 1 million tokens from the previous 64,000, giving the model more room for extended reasoning and generation in a single run. Vals lists a different evaluation ceiling: 262,000 output tokens, alongside a 1-million-token context window. The context window and output cap are separate figures, and neither should be substituted for the other.
Vals ran its evaluations with reasoning effort set to high and temperature set to 1.0. Its benchmark results therefore describe that configuration and its listed output cap—not a test of Google’s full advertised million-token output allowance. The larger allowance is a launch capability claim; the leaderboard measures the settings Vals actually used.
Google is rolling Argon out through its Fairwind Program to trusted cyber defenders. For selected defenders and internal Google teams, the company says it will release the model without cyber guardrails so they can use its full defensive capabilities. That is a specific access arrangement, not the policy Google describes for broader availability.
Before that wider release, Google says it is strengthening safeguards against misuse and malicious instructions, monitoring the model’s reasoning and actions, and hardening isolated testing environments. It is also participating in the U.S. government’s voluntary pre-release model access process. These are ongoing preparations, rather than a claim that safety work is finished.
Paid API customers and Google AI Ultra subscribers are first in line for the broader release. Google says early testers’ feedback will help it refine safeguards before expanding to developers, enterprises and consumers. It has given no specific date for that next stage.
blog.googleEditorial analysis
The important distinction is between announcing a model and making its capabilities broadly usable. Our reading is that Google is treating trusted defenders as both early customers and a proving ground for the next release stage. That creates a demanding test: feedback from specialists using Argon without cyber guardrails must inform safeguards for a much wider audience. Watch the first paid API release for how Google translates that experience into access rules and operating controls. Benchmark scores describe task performance; the broader rollout will reveal which parts of that performance Google is prepared to offer under general-use restrictions.
Loading discussion...
Join the conversation
Explain whether restricted access is enough to justify removing those limits.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.