# ECONOMICS.md Every cost and metric quoted on stream, in one place, with the rule each one supports. Figures are what speakers said out loud in September 2026; treat them as order-of-magnitude, not a price list. --- ## What work cost | Work | Cost | Manual equivalent | Source | |---|---|---|---| | Full sales case-study slide deck | $20–30 | 4–5 hours | Amrita, day 2 | | Mid/complex support ticket resolved | $1–2 | more than that for a human | David, day 2 | | Low-complexity billing ticket, after bucketing and batch scripts | ~$0.20 | — | David, day 2 | | Comparable support-agent products | $1–10 per ticket | — | David, day 2 | | Finding a cheaper PG&E plan | ~60 seconds of attention | ~$1,000/year saved | Matthew Berman, day 2 | The support numbers are the useful pair: the same job went from $1–2 to $0.20 once low-complexity tickets were sorted into a bucket and handled by a script instead of a reasoning pass. Cost is a property of the setup, not of the tool. --- ## What made the bills go up | Driver | Quoted | Rule | |---|---|---| | A routine every 15 minutes | "comes out to 100 times a day"; three of those is "hundreds of messages a day" | Once or twice a day is the default. Prefer webhooks and inbound signals to blind polling. | | Bots in a group chat | "eager", "love to talk", "speak over each other", get expensive fast | One bot tags two others once. Group chats only when you want a debate. | | UI work through the bot's browser | filling a web form by clicking costs more than an API-native path | Use a connector when one exists. The browser is the fallback. | | Too many bots | keep the team lean, "not 45 bots" | One chief of staff over 10–20 narrow specialists worked; nobody needed middle managers. | | Throwaway verification scripts | rewritten every run, token-costly, non-reproducible | Build one CLI the agents call instead. | | Agents that `sleep 300` | four minutes lost per wait on tests that finish in one | Watch running agents; a 5-minute check routine is for this, not for reporting. | Frequency and group chats are where the tokens go. Everything else is second order. --- ## Throughput | Metric | Number | Source | |---|---|---| | PStack PRs merged to production in one month | 2,500 (~83/day) | Lauren, day 1 | | Share of merged PRs at the company coming through Grok Bot | double-digit percent | PM session, day 1 | | First mobile app version, one engineer | 3 weeks | Ling, day 1 | | Cloud agents one engineer managed by hand before orchestration | 15 | Ling, day 1 | | CI auto-fix: page a human only after | 10 minutes unresolved | day 1 | | PRs in the 72-hour build | 433 | day 3 | | Commits from one host in three days | 157 | Roshan, day 3 | | Bot employees across 7 businesses | 22 | Jenny Co, day 1 | | SDR prospects per day | 50, top 5 actioned each morning | Simon, day 2 | --- ## Launch day, for scale Thursday Arena went live at ~10:00 AM Pacific from one X post. | When | Number | |---|---| | Launch hour | 1,908 games — the best hour of the day | | Sustained afternoon | 400–500 games/hour | | End of day | ~4,000 games, 6,000+ public matches, ~30,000 page views, ~2,000 users | | Practice → X sign-in conversion | ~8% | | Feedback mix | 71% bugs, 16% praise | | Revenue | $0 (one theoretical $1 test bid on a broken ad auction) | Traffic: roughly 50/50 mobile and desktop, 44% iOS, referrals almost all from X. Eight people came from Bing.