邮件分流实测：GPT-6 Astra 月费 2250 美元 vs Jev 6.3 美元

Donatello @DonatelloXPro原推发布于 2026.10.14

在 X 上打开原推

I'M STILL IN SHOCK. GPT-6 ASTRA: $2,250 A MONTH. JEV: $6.30. SAME F*CKING ANSWER. 357X CHEAPER. 03:12. A new email lands in an inbox nobody's watching. The AI agent has one question to answer: does this matter, yes or no. An agent running all night asks that thousands of times. Is this email relevant. Who owns this ticket. Is this action risky. Every answer has two or three options. Not one of them needs a model that writes. Most people route all of it through their smartest model anyway. Then they open the bill. At 10,000 decisions a day, 500 tokens each, the month looks like this. GPT-6 Astra at $10 in and $50 out: $2,250. GPT-6.1 Sol at $2 in and $10 out: $450. Jev at $0.042 in, with output free: $6.30. The difference is in what each model sends back. Astra writes you a sentence explaining why the email doesn't matter. Jev sends back one number, the probability that it does. Your own code holds the line at 0.5 and makes the call. That split matters more than the price. When the threshold lives in your code, you can move it. Too many false alarms at 4 AM, raise it to 0.6. Missing important emails, drop it to 0.4. With a writing model, the line lives inside a paragraph you have to parse every time. The 03:12 email scored under the line. Skipped, logged, nobody woke up. Over a year that's $27,000 against $75.60, for the word "no." I used to send every yes/no to the best model I had. Now it only gets the work that needs words. The full routing setup is below and free. Almost nobody has it saved. Save it before your next API bill lands.