Route for usable capacity

  • Default even distribution can leave allowance unused when an account resets. Theo changes routing to favor accounts whose next reset is closest, while retaining enough accounts to handle five-hour limits. 18:04 18:34
  • For the Claude 20× plan shown, he says consuming a full five-hour window uses 40% of the Fable seven-day allowance, or 20% of the overall seven-day allowance. Those are different buckets. 18:55 19:21
  • He recommends enabling Codex WebSocket support in the proxy to reduce message round-trip delays. The video supplies no measured latency improvement or exact patch. 20:25 20:51
  • Session/account affinity means keeping successive prompts in one thread on the same account. His explanation is that caches are account-specific, so switching accounts forces a cache rebuild. 20:51 21:19
  • Frequent switches per prompt or tool call waste cache writes. He says the default affinity generally works and caches last five minutes; inspect cache read/write usage when modifying routing. 21:19 21:47

Two different reset behaviors

  • In his description, a Codex reset refills capacity and restarts the weekly clock at seven days. Simultaneous resets can align all accounts and leave a long wait to the next renewal. 26:18
  • Codex has immediate resets and banked resets that the user can trigger later. His explanation that off-hours capacity makes immediate resets cheaper is an economic hypothesis, not a verified provider policy. 26:51 27:14
  • A Claude reset refills remaining capacity without changing the existing renewal time. A reset just before regular renewal therefore creates a short window to use the extra allowance. 27:35 28:02
  • Theo treats unused subscription allowance as expiring capacity: prioritize useful work before renewal. This reverses the incentive of metered API use, where reducing token consumption reduces the bill. 28:19 28:43