Google has released Gemini 4 Argon, a frontier model it positions for long, complex workflows in software engineering, legal and financial knowledge work, and cyber defence. Unlike its consumer launches, Argon will roll out in stages, reaching trusted cyber-defence partners first through the Fairwind Program.
The headline change is the output ceiling. Argon can generate up to one million tokens in a single response, up from 64,000 on its previous generation. Most frontier models cap output near 128,000 tokens, so a one-million-token single generation is rare. Google says the room lets the model reason across a long task trail and produce results running to hundreds of thousands of tokens, aimed at large code migrations, deep research and multi-step enterprise processes.
Benchmarks and pricing
On the DeepSWE v1.1 software-engineering benchmark, Argon scored 77.9 per cent, the current high. It ranks first on the Vals Index, which weights finance, coding, law and tax by each sector’s contribution to US GDP. It also leads Vals Finance Agent v2, the Harvey Legal Agent Benchmark and Zapier’s AutomationBench, where it scored 51.3 per cent. On long-video understanding, measured by LVBench, it scored 91.7 per cent.
API pricing starts at US$2 per million input tokens and US$10 per million output tokens, with cached input priced 95 per cent below standard. Google says thousands of staff already use Argon for debugging, algorithm design and large code-base migration. In one quantum-computing test it lifted a published baseline by 40 per cent in minutes. On a data-centre memory project, Argon’s agent freed more than 300 tebibytes of memory, heading to a total saving of 500 TiB to 1 pebibyte.
Cyber defence first
In cyber defence, Argon can find, verify and patch critical software flaws on its own. On the CWE-bench v1 vulnerability-fix test it tied for first at 68 per cent. Google notes the gap between benchmark scores and real use remains, and Argon will join the US government’s pre-release voluntary review before a broader rollout to developers, enterprises and consumers.
Editor’s note: This is an adapted translation of the original 36Kr (Machine Heart) report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://36kr.com/p/4006503753830529.