Google is moving its flagship AI push to Gemini 4 Argon after scrapping plans to release Gemini 3.5 Pro, Reuters reported, closing out a delayed upgrade cycle during which OpenAI and Anthropic released newer frontier models.
Announced September 30, Argon is built for long-running software engineering, enterprise knowledge work, and cybersecurity defense.
It is not broadly available yet: Google is starting with selected cyber defenders through its Fairwind Program before expanding access to developers, enterprises, and consumers. Paid API customers and Google AI Ultra subscribers are expected to be first in line, but Google has not announced a date.
A reset after Gemini 3.5 Pro
Google’s previous Pro-class flagship, Gemini 3.1 Pro, arrived on February 19. At Google I/O on May 19, the company introduced Gemini 3.5 Flash and said Gemini 3.5 Pro was already being used internally and would roll out the following month.
That June release never happened. By July 21, Google said 3.5 Pro was still being tested with partners and would become broadly available when ready. In the same announcement, it disclosed that work had already begun on what it called its most ambitious pre-training run yet for Gemini 4.
Google ultimately scrapped plans to release 3.5 Pro, according to Reuters. In the meantime, OpenAI launched GPT-6 Astra on September 3, followed by Anthropic’s Claude Opus 5.5 on September 22.
Argon therefore represents more than a routine model upgrade. Google has effectively moved past the delayed 3.5 Pro and into a new generation aimed at long-running agent work, coding, professional workflows, and cyber defense.
Argon competes with Astra and Opus, but does not sweep them
Google’s own benchmark table shows Argon posting the highest score on 13 of the 19 listed benchmark rows, with particularly strong results across knowledge work, long-context tasks, multimodal understanding, and some coding evaluations.
It leads the Vals Index, AutomationBench, Vals Finance Agent v2, Harvey’s Legal Agent Benchmark, DeepSWE v1.1, Vibe Code Bench, LABBench 2, RiemannBench, both GraphWalks tests, Agent’s Last Exam, Chartography, and LVBench. On CWE-bench v1, Argon ties GPT-6 Astra at 68%.
The results are not a clean sweep. Astra leads FrontierSWE v2, Terminal-Bench Science 0.1, and OSWorld-2.0, while Claude Opus 5.5 leads Terminal-bench 4.0 and PostTrainBench.
On DeepSWE v1.1, Argon scores 77.9%, compared with 74.2% for Opus 5.5 and 74.1% for Astra, but on FrontierSWE v2, Astra reaches 65.5% against Argon’s 55.0%.
Google is also emphasizing what Argon can do over longer periods rather than only its benchmark scores. The model supports up to 1 million output tokens in a single run, up from 64,000 previously.
Google says teams using Argon internally have applied agents to jobs including data-center memory optimization and large C and C++ migrations to Rust.
Cybersecurity is central to the initial rollout. Google says Argon can autonomously find, validate, and patch software vulnerabilities, and selected Fairwind defenders will receive access to its full cyber capabilities while the company continues strengthening safeguards before a wider release.
For users in the Philippines, there is no Argon-specific rollout date yet. The Philippines is already a supported market for both the Gemini API and Google AI Ultra, the two channels Google plans to prioritize when wider access begins, but that does not confirm when Filipino developers or subscribers will actually receive Argon.