Total Cost of Agentic Tasks is Rising
ArtificialAnlys · x · 2026-07-14
At the AI Engineer World’s Fair, Artificial Anlys shared a key insight: 固定水平的“智能”正在变便宜，但完成复杂智能体任务的总成本反而在上升.
This is due to factors beyond just token price:
- 多轮交互: Agentic tasks often require multiple turns, and previous outputs feed back into the context, causing costs to accumulate.
- token 效率差异: Different models consume varying amounts of tokens for the same task; even with a lower unit price, the overall cost might be higher.
- prompt caching 命中率: Long tasks repeatedly pass the same instructions, research materials, and tool outputs, making cache hits increasingly impactful on the total price.
To profile such workloads, they introduced the AA-Briefcase benchmark to measure model performance in 长周期知识工作. Tasks span multi-week projects and thousands of input files, breaking down the costs of tool calls, turns, and tokens.
Related event: True AI Cost Depends on Task Completion Rates(2 posts)→
More from coding & agent
- Cheaper OpenAI Agents API alternative: sandbox service undercutting E2B by 46% — airesearch12 · 2026-09-11
- His agent kill switch ran for months before he found it was wired to nothing — AnvilandCode · 2026-09-11
- Kernel's Browser Agents Can Now Pay Online Using Aliases, Never Touching Card Data — jeff_weinstein · 2026-09-11
- OpenAI opens up agent sandboxes: BYO or pick from Cloudflare, E2B, Modal, Vercel and more — threepointone · 2026-09-11
- SocialCrawl MCP lets agents search Reddit, YouTube, TikTok, X with one API key — dooddyman · 2026-09-11
- Astra builds a surprisingly polished Catan game in three.js, reusing past UI and 3D assets — FinanceYF5 · 2026-09-11