Asana cut a browser agent's cost 76x to $0.47 per run via prompt caching
Codeblix_Ltd · reddit · 2026-10-10
Asana reports cutting one browser-agent workflow's estimated model cost from at least $36.21 to $0.47 per run (76x) and runtime from 22.5 minutes to about 4 minutes — without changing models.
The trick:
- Keep the agent's growing page history stable long enough for prompt caching to kick in
- Prune screenshots in batches instead of rewriting the history every step
Evidence and caveats: a 144-run study across GPT-6.1 Sol and three other frontier models, plus a 12-run follow-up. Every best-condition run reached the same 192 facts, but only 3-4 runs per condition were used and the baseline included capped runs.
Takeaway: measure cache reads, history churn, cost, runtime, and answer quality together — a cheaper agent that fails more tasks isn't cheaper.
Related event: Asana Cuts Codex Agent Costs via Stable Prompt Caching(3 posts)→
More from coding & agent
- Developer Urges Watching This Agentic Coding Video at 2x This Weekend — jarrodwatts · 2026-10-11
- Vercel CEO Builds a Working App in 20 Minutes From Idea With AI Coding — zeeg · 2026-10-11
- Codex ships real-time Steering: turn it on now in the ChatGPT desktop app — daniel_mac8 · 2026-10-11
- Who cleans up coding agents' branches and worktrees? Assay shares a workflow — fuckme · 2026-10-11
- User spends all day with Grokbot managing his network and running Claude Code sessions — Rasmic · 2026-10-11
- Vibes-coding with Codex now replaces the mental overhead of WordPress plugin setups — tinyfool · 2026-10-11