Google withheld general release of Gemini 4 Argon on 30 September 2026, granting access only to vetted cybersecurity specialists and US government testers under a White House accord.
Argon expands output capacity to 1 million tokens per response, up from 64,000 tokens in prior models, allowing systems to generate roughly 750,000 words without interruption.
The system scored 77.9% on DeepSWE v1.1 and 68% on CWE-bench, while internal teams used it to rewrite 32,000 lines of libgav1 code and free over 300 TiB of data center memory.
Bloomberg reported that some internal employees observed performance struggles on coding tasks, though Google and chief AI architect Koray Kavukcuoglu defended the model.
Google introduced Gemini 4 Argon on Wednesday, 30 September 2026, restricting initial access to vetted cybersecurity experts and internal teams. The tech company submitted the model to the voluntary pre-release vetting process established by the US government, following an agreement signed at the White House by tech leaders including Alphabet chief executive Sundar Pichai and Anthropic chief executive Dario Amodei. The cautious rollout mirrors measures taken after Washington temporarily forced Anthropic to suspend public access to its Claude Mythos and Claude Fable models in June 2026. Google designed Argon to refuse instructions related to cyber-attacks or chemical, biological, and nuclear weapons, while monitoring internal reasoning to prevent misalignment.
“Safely releasing frontier capabilities at this level requires a phased approach.”
Gemini 4 Argon raises the single-generation output limit to 1 million tokens, an increase from the 64,000-token ceiling in previous Gemini models. One million tokens corresponds to approximately 750,000 English words, enabling automated agents to complete prolonged tasks without dividing requests into dozens of smaller prompts. To support continuous generation, the API introduces a mechanism that pauses extended answers and resumes them across successive calls. In technical testing, Google stated that Argon scored 77.9% on the DeepSWE v1.1 benchmark for complex software engineering and 68% on CWE-bench for repairing software vulnerabilities. The model also ranked first on the Vals Index, which measures economic impact across legal, tax, financial, and coding domains.
Data from the article text
Google reported that thousands of its engineers already employ Argon for internal operations, including optimizing data center memory management to release more than 300 tebibytes (TiB). Automated agents used the system to translate 32,000 lines of libgav1 video decoder code from C and C++ into Rust, yielding a 2.7-fold increase in processing speed, and analyzed over 800,000 lines of the Zircon kernel within the Fuchsia operating system. In early partner testing, researchers used Argon to detect a vulnerability in global hospital software that leaked personal records. However, Bloomberg reported that some internal employees found the model encountered difficulties in real-world coding assignments despite strong benchmark marks. Google disputed the report, expressing confidence in the system led by Koray Kavukcuoglu, who succeeded Demis Hassabis as head of Google DeepMind in August 2026.
“Argon is a complete model, with cutting-edge capabilities across multiple areas.”
The debut marks Google's first Gemini 4 release since the launch of Gemini 3 nearly a year prior, after an anticipated June release failed to materialize. Argon enters direct competition with rival models including Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra and GPT-6.1 Sol. Data compiled by Artificial Analysis indicates that running tasks on Argon costs $1.99 per benchmark task during its introductory promotion, representing 60% of the cost of GPT-6 Astra. Argon consumes an average of 62,000 output tokens per task compared to 27,000 tokens for Astra, with standard task pricing set to double to $3.98 after the promotional period ends. Google plans to broaden access to Google AI Ultra subscribers and paid API customers after concluding early safety evaluations.
Pollar is free and carries no ads in the text. Its members keep it going.
Support PollarAI & Tech · Mountain View and Washington
Google has unveiled Gemini 4 Argon with a 1-million-token output capacity, limiting initial deployment to select cybersecurity specialists and federal reviewers under a voluntary White House safety framework.
AI text based on reports by EL MUNDO, Gizmodo, The Guardian and other publishers. Changed .