Google's Gemini 4 Argon: Cheaper Than GPT-6 Astra on Paper, but Most Benchmarks Are Google's Own

Google has introduced Gemini 4 Argon, a frontier model pitched at long-running enterprise work, with a headline price well below OpenAI's and a rollout that starts with cybersecurity defenders rather than the public. The numbers are impressive, but nearly all of them come from Google.

What Google Announced

Google says Argon is "designed for complex, long-running tasks across software engineering, legal and financial work, and cybersecurity." The biggest spec change is output length: Argon can produce up to 1 million output tokens in a single request, up from a previous limit of 64,000. Google showed that capacity off with internal projects. According to coverage from eWeek and TUN, they include migrating more than 800,000 lines of code and finding datacenter memory optimizations (eWeek reports more than 300 terabytes freed; Yahoo Tech's account says improvements that could save 1 PiB across Google's data centers). Yahoo Tech adds that Argon helped quantum researchers beat a published optimization baseline by 40% and helped convert C/C++ code to Rust that ran 2.7 times faster than the previous version.

The date is murky. TUN says the launch was Sept. 30, eWeek describes an Oct. 5 announcement, and CNBC was already asking on Oct. 2 whether Google's new model could catch up to OpenAI and Anthropic, so the model appears to have been announced in early October.

Pricing

Reported introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted 95%. eWeek reports those rates later rise to $4 and $20, and contrasts them with $10 and $50 for OpenAI's GPT-6 Astra. We did not independently verify the GPT-6 Astra figures. TUN also cites a cost of about $1.99 per task on the Vals benchmark, a measure that is only meaningful for that specific test.

The Benchmarks, and Who Reported Them

Google's self-reported results, as listed by TUN, include 77.9% on DeepSWE v1.1, 91.7% on LVBench, and 68% on CWE-bench v1. eWeek's analysis of Google's own comparison table says Argon beat GPT-6 Astra in 14 of 19 comparable rows, lost 4, and tied 1. The one result that was independently verified, per TUN, is a 65.4% score on Vals Finance Agent v2, ranked first. The coverage we read, including Yahoo Tech's, offered no independent analyst commentary beyond Google's own claims.

A Staged, Security-First Rollout

Argon is not generally available. Google is releasing it first to "trusted cyber defenders" through its Fairwind Program, then to paid API customers and Google AI Ultra subscribers, with no date for wider access. Google says it is "strengthening safeguards against misuse, prompt injection and model misalignment before broader availability." That echoes the security themes in our agent security deep-dive and the White House pledge covered in our report on the new AI czar.

How to Evaluate a New Model Like This

This section is practical advice, not reporting. When a vendor-reported benchmark table lands, the useful move is to ignore the headline score and test the model on your own tasks. Pick 20 to 50 real examples, run them against the model you use now, and compare quality, cost per completed task, and failure modes. Our AI power-user playbook explains how to build that evaluation set and when to route work between models, and our guide to paid AI tools covers the subscription side.

Why It Matters

• Price pressure at the frontier. Reported pricing is a fraction of what eWeek lists for GPT-6 Astra, though it may rise after the introductory period.
• Long outputs change what is possible. A 1 million-token output limit targets large code migrations and long research documents.
• Treat the scores with care. Nearly all benchmark data is self-reported; only one result is reported as independently verified.
• Security gates the release. Starting with vetted defenders signals how seriously the cyber capabilities are being handled.