同基准三模型对比：Opus 5.5 每任务 5.98 美元、Gemini 4 Argon 1.99、V4.1 Flash

Prasenjit Sarkar @stretchcloud原推发布于 2026.10.18

在 X 上打开原推

Same benchmark, three models, one model costs 22 times more per task. That's the frame tekario built around the Artificial Analysis Intelligence Index this week, turning it into a literal depth race: Claude Opus 5.5 scores 58 points at $5.98 a task, Gemini 4 Argon scores 53 at $1.99, DeepSeek V4.1 Flash scores 39 at $0.27. Gemini lands at roughly 27 points of intelligence per dollar, the best ratio of the three. DeepSeek gets two thirds of Opus's score for under five cents on the dollar. Except that $1.99 number for Argon won't last. Google's own footnote on the September 30 announcement says the introductory $2 input and $10 output pricing per million tokens doubles to $4 and $20 once the promo period ends, with no end date published. Artificial Analysis is already telling people to budget at the standard rate: run an agent task at 200K input and 60K output tokens and you're paying about a dollar per run today, roughly two dollars once the discount lapses. At a thousand runs a month that's a $620 bill becoming $1,240, and Argon isn't even generally available yet outside trusted testers and Google AI Ultra subscribers. This is becoming the standard move across the category. Google priced Argon's promo to land exactly on GPT-6.1 Sol's $2/$10 rate and half of GPT-6 Astra and Claude Fable 5.1's $10/$50, then set the real price to match Claude Opus 5.5's $4/$20 line for line once the teaser period ends. Everyone's headline number is the one in the launch post. Nobody's headline number is the one on the invoice in month two. What matters for a team picking a default model isn't the per token rate on launch day. It's points of intelligence per dollar measured against the bill you'll actually pay in steady state, not the promo one. Argon's real test starts the day the discount expires. https://x.com/tekari0/status/2111532349171372311

这条是引用着另一条写的：「Gemini 4 Argon 与 Opus 5.5、DeepSeek V4.1 性价比对比」