GPT-6 Astra 每任务成本与省 token 五条规则
Artificial Analysis 指数显示 Astra 每任务约 2.7 万 token、$1.35，比 Opus 5.5 和 Fable 5.1 便宜，但分数仍落后；帖中给出控制输入长度、缓存、限制输出等省钱规则。
Sam Altman's new model uses 4x fewer tokens than Claude Opus. and costs a third of Fable GPT-6 ASTRA: $1.35 A TASK. CLAUDE FABLE 5.1: $3.90. SAME BENCHMARK RUN you pay for every token a model writes to sound smart. Astra barely writes any Artificial Analysis ran the full index: GPT-6 Astra → ~27K tokens, $1.35 Opus 5 → ~73K, $1.83 Opus 5.5 → ~119K, $2.38 Fable 5.1 → ~78K, $3.90 honest part: Fable 5.1 still scores higher on that index. Astra wins on what you pay per answer and the cheap price only holds if you follow 5 rules. steal this: 1. stay under 272K input. cross it and the WHOLE request bills at 2x input and 1.5x output. 500K in + 32K out = $12.40 for one call 2. reuse your context. cached input is $1 per 1M instead of $10 3. cap output. Astra is short by default. type "explain in detail" and you're paying Claude prices on an OpenAI model 4. hand repeat work to a Dot. it runs Astra on its own cloud computer with 4,000+ apps and keeps going after you close ChatGPT 5. write Custom Rules first: allow, approve, prohibit. password changes always wait for you OpenAI's own numbers vs Fable 5.1: BenchCAD Vision2Code → 95.9 vs 84.3 Terminal-Bench Science → 64.6 vs 52.6 I moved my research runs to Astra and kept every prompt under 272K. full price grid is in the image, ready to copy save this before your first $30 call ↓
这条是引用着另一条写的：x.com/donatelloxpro/status/2109320196649619552