the honest mechanics

How Standard Compute rate limits work: pacing, not windows

“Unlimited” has become a nearly meaningless word in this market — every plan has some boundary. Here is ours, stated plainly: a fixed monthly compute budget, no windows inside the month, and smooth pacing instead of hard walls.

Normal use
Full speed. No meter ticking per request, no windows to think about — requests run as fast as the models answer.
Extreme sustained load
Pacing engages: requests keep succeeding but are spaced out. Your agent cruises instead of sprints — no 429s, no dead loops.
Budget reached
The honest ceiling. Generously sized per tier; the dashboard shows your position at all times. Upgrade raises it.
Limit modelWho uses itMid-task failure mode
Per-minute 429sOpenAI/Anthropic APIs, OpenRouterRequest rejected — agent loop breaks, retries cascade
Rolling windows (5h + weekly)Claude Pro/Max, Codex, OpenCode GoHard stop mid-sprint until the window resets
Credit quotasMiniMax, GLM, Synthetic, Chutes, Router9Quota exhausts — work stops or bills overflow
Monthly budget + smooth pacingStandard ComputeNone mid-task: requests keep succeeding, pacing spaces them under extreme load

FAQ

Does Standard Compute have rate limits?
Not in the classic sense. There are no 5-hour session windows, no weekly caps, and no per-minute 429 walls to hit. Each plan has a fixed monthly compute budget; requests run at full speed under normal use. Only extreme sustained load ahead of the monthly budget engages pacing — requests are spaced out smoothly rather than rejected, so agents keep working instead of crashing on errors.
What does pacing feel like compared to a 429 error?
A 429 breaks your agent's loop: the request fails, retries pile up, tasks die mid-run. Pacing keeps every request succeeding but adds spacing between them when sustained load runs far ahead of the plan's budget. Your agent finishes its task — it just cruises instead of sprints. No error handling required.
How is this different from Claude Max or Codex windows?
Window-based plans meter usage inside rolling windows (5-hour sessions, weekly caps) — hit the wall mid-task and work stops until reset, no matter what you pay. Standard Compute has one budget for the whole month with no intra-month windows: a heavy Tuesday doesn't lock your Wednesday.
Is there an optional pacing mode I control?
Yes — Smart pacing is an optional dashboard switch that spreads your budget evenly across the month, useful for always-on agents you don't want draining the budget in the first week. It's off by default: without it, requests run at full speed until the monthly budget is reached.
So what happens if I use the whole monthly budget?
The budget is sized generously per tier and most users never touch the ceiling. If you do reach it, that's the honest boundary of 'flat-rate' — no plan anywhere has infinite compute. Upgrading a tier raises the budget; the dashboard shows exactly where you stand at all times.

More: the coding subscription · what “unlimited” honestly means · why agents get rate limited