Gemini 4 Argon costs 1/5 per token but per-task bills may exceed Astra once discounts end

新智元 · wechat · 2026-10-01

After seven months, Google shipped Gemini 4 Argon, its first non-Flash closed model, scoring 53 on Artificial Analysis' intelligence index—tying GPT-6 Astra and Fable 5.1—with the lowest hallucination rate (15%) among 45+ models and a tie for first on CWE-bench.

The interesting part is the math:

Pricing looks cheap: $2/$10 per million input/output tokens—half of Opus 5.5, one-fifth of Astra
But Argon averages 62K output tokens per task vs Astra's 27K; at intro discounts a task costs $1.99 (60% of Astra) yet 2.7x Sol's $0.72
At standard prices the same task would cost $3.98, about 1.2x Astra—cheap per token, expensive per task
Low hallucination has a cost: 50% accuracy on AA-Omniscience, 13 points below Astra's 63%
Agent ability is uneven: 77.5% tops Automation Bench-AA, but 57% on Terminal Bench 4 trails Sonnet 5.5, Opus 5.5 and Astra
Internal wins include freeing 300+ TiB of data center memory and a 2.7x-faster Rust libgav1 after rewriting 32K lines of SIMD code, though the 800K-line kernel migration is still ongoing.

Verdict: Google is back among the top three labs, but long-term adopters should benchmark token usage and discount terms against their own workloads.

Original post →

Gemini 4 Argon costs 1/5 per token but per-task bills may exceed Astra once discounts end

More from Models