Google introduces Gemini 4 Argon for long-running AI tasks

Google announced Gemini 4 Argon on September 30. The model is designed for tasks that require many steps, with a stated focus on software engineering, financial research, legal work and defensive cybersecurity.

Google's introductory pricing is $2 per million input tokens and $10 per million output tokens. After the introductory period, the listed prices rise to $4 and $20 respectively. Reuters reports that Google has not given a date for a public release.

That is one of the most concrete changes from previous models. A much larger output budget gives the model room to keep a single work trajectory going across many steps instead of constantly breaking the task into separate requests.

One million tokens does not automatically mean one million useful tokens. For agents, the important part is what happens along the sequence: whether the model keeps state, uses tools correctly and avoids compounding errors as the task grows.

According to Google, a group of Argon agents analysed data-center telemetry and found changes that freed more than 300 TiB of memory after rollout. The company estimates total potential savings of 500 TiB to 1 PiB.

Google also cites C/C++ to Rust migrations on codebases exceeding 800,000 lines, including Fuchsia's Zircon kernel. In another case, Argon worked on 32,000 lines of SIMD code in libgav1; Google says the resulting Rust implementation runs 2.7 times faster than the previous port while producing identical video output.

These are company-reported results rather than independent tests. The useful detail is that Google also describes the surrounding process: automated tests, manual audits, emulation and review before changes are rolled out.

The model is also being evaluated on finance, legal work and business automation. On Zapier's AutomationBench, which measures end-to-end task execution across business tools, the Argon High configuration is currently listed at 51.29%.

Cybersecurity is the other major focus. Google says Argon can find, validate and patch software vulnerabilities. Initial access therefore runs through the Fairwind Program, which is limited to governments, selected Google Cloud customers and cybersecurity partners.

Google reports 77.9% on DeepSWE v1.1, a benchmark for long-horizon software engineering tasks. But an independent Epoch AI review rated the benchmark “Flawed” after finding problems in at least 23 of the 113 tasks it examined.

Reuters also notes that Argon remains behind competitors on two of the four coding benchmarks included in Google's launch material. The picture is therefore less straightforward than any single score suggests.

Restricted access makes it difficult to know how the model behaves on long-running tasks designed by external teams, with different tools, data and failure modes.

Once access widens, the most useful measurements will be less dramatic than leaderboard scores: completed-task rate, errors accumulated across a sequence, recovery after tool failures, cost per task and the number of human interventions required.