Google Launches Gemini 4 Argon Frontier AI Model
The first release in the Gemini 4 series introduces a massive 1-million-token output limit, specialized cybersecurity features, and competitive benchmark performance.
Key takeaways
- Google announced Gemini 4 Argon, the first model in its Gemini 4 generation, featuring a 1-million-token output limit.
- Argon introduces introductory pricing of $2 per million input tokens and $10 per million output tokens, alongside a 95% discount for cached input tokens.
- The model tied for first on the CWE-bench vulnerability remediation benchmark at 68% and scored 77.9% on DeepSWE v1.1.
- Early deployment is restricted to trusted partners in Google's Fairwind Program, with broader release planned for API users and Google AI Ultra subscribers.
On September 30, 2026, Google announced the release of Gemini 4 Argon, the first model in its next-generation Gemini 4 series. According to Google's official announcement, the frontier model is built to sustain deep reasoning across long-horizon workflows spanning software engineering, enterprise knowledge work in finance and law, and cybersecurity defense. The company is now rolling out early access through a phased safety initiative.
Rather than launching immediately to the broad public, Google is distributing Argon first to vetted cyber authorities and defense organizations through its Fairwind Program. The company stated that it is actively participating in the U.S. government's voluntary pre-release model evaluation framework to assess frontier risks prior to wider availability.
Technical Specifications and 1M Output Window
Google expanded the model's output token limit to 1 million tokens, a substantial increase from the prior 64,000-token threshold. As reported by Engadget, this output capacity is significantly larger than OpenAI's GPT-6 Astra, which offers 128,000 output tokens.
In terms of pricing, Google announced an introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%. Independent testing by benchmarking firm Artificial Analysis cited by Engadget found that Gemini 4 Argon matches GPT-6 Astra's Intelligence Index score at 60% of the cost per task under current promotional rates. Artificial Analysis also recorded a 15% hallucination rate for Argon, compared to 54% for GPT-6 Astra and 54% for GPT-6.1 Sol.
Benchmark Performance Across Engineering and Enterprise Tasks
Google claims Gemini 4 Argon establishes leading marks across several real-world task evaluations. On the DeepSWE v1.1 benchmark, which tests long-horizon software engineering capabilities, Argon achieved a state-of-the-art score of 77.9%. On Zapier's AutomationBench evaluating end-to-end business function execution, the model ranked first with a score of 51.3%.
For visual knowledge tasks, Google reported a score of 91.7% on LVBench, a benchmark measuring long video comprehension. The company also reported top scores on the Vals Index—which weights economic performance across finance, legal, tax, and coding based on contributions to U.S. GDP—as well as domain-specific tests including the Vals Finance Agent v2 and Harvey's Legal Agent Benchmark.
Internal Deployments and Data Center Optimization
Within Google, engineering teams have already integrated Gemini 4 Argon into infrastructure and research workflows. According to Google, in one quantum computing optimization task, researchers used the model to optimize spacetime resources (qubits × gates) for bottlenecked subroutines, beating the published baseline by 40% in minutes.
Google reported that autonomous Argon agents analyzed fleet telemetry across data centers to identify memory efficiencies, freeing over 300 TiB of memory with projected total savings between 500 TiB and 1 PiB, while Business Insider reported that Argon helped engineers free up more than 300 tebibytes of memory without new hardware.
Google teams are also using the model to migrate legacy C and C++ codebases into memory-safe Rust. These efforts include libraries like re2, video decoder libgav1, and the Fuchsia Zircon kernel spanning over 800,000 lines of code. Google reported that for libgav1, Argon agents replaced 32,000 lines of SIMD instructions with safe Rust through iterative compiler analysis, yielding a decoder that runs 2.7 times faster than the initial Rust port.
Defensive Cybersecurity and Frontier Safeguards
Cybersecurity defense is a focal point of Gemini 4 Argon's training. The model tied for first place on the CWE-bench v1 vulnerability remediation benchmark with a 68% score, matching GPT-6 Astra and Grok 4.7. In early trials with cloud security firm Wiz under its Scan for Good initiative, the model identified an unpatched vulnerability in widely deployed healthcare software that exposed sensitive personal information.
To allow trusted defenders to find, validate, and patch critical flaws, Google is providing Fairwind Program participants with versions of Argon that omit cyber guardrails. For general safety, Google stated it has updated mitigations against chemical, biological, radiological, and nuclear (CBRN) misuse, strengthened resistance against indirect prompt injections on the Gray Swan benchmark, and deployed chain-of-thought monitoring to halt execution if misalignment is detected.
Google indicated that broader access for developers and enterprises—starting with paid API customers and Google AI Ultra subscribers—will begin following initial evaluations by early testers.
Frequently asked questions
What is Gemini 4 Argon's output token limit?
Gemini 4 Argon features an output token limit of 1 million tokens, expanded from the previous 64,000-token limit.
How much does Gemini 4 Argon cost to use?
Google announced introductory API pricing of $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached input tokens.
Who currently has access to Gemini 4 Argon?
Access is currently limited to trusted cybersecurity defenders and government partners via Google's Fairwind Program, with future access planned for paid API users and Google AI Ultra subscribers.
How did Gemini 4 Argon perform on cybersecurity benchmarks?
Gemini 4 Argon tied for first place on the CWE-bench v1 vulnerability remediation benchmark with a top score of 68%.
Sources
- Gemini 4 Argon: our next era of frontier intelligenceGoogle · Sep 30, 2026 · Official
- Google launches Gemini 4 Argon to reclaim the AI frontierBusiness Insider · Sep 30, 2026
- Google's First Gemini 4 Model Is 'Argon'Engadget · Oct 1, 2026
How this story was made: the newsroom picked it up from Google News, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (31 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 3, 2026 at 00:13 UTC
