Module 1 · Throughput under limits · Lesson 1

Rate limits and quotas

How model providers ration capacity (RPM, input and output tokens per minute, usage tiers, spend caps and acceleration limits), how the Claude API enforces and reports them, and how to compute which limit binds your fleet before it binds in production.

70 min 1 video 3 exercises 7-question quiz Answers

Setup for the exercises:Install Claude Code Python with uv Get an OpenRouter key OpenRouter or a local model?Local models with Ollama

By the end of this lesson you can

Explain how RPM, ITPM and OTPM are enforced with token buckets, and why a per-minute limit can reject a short burst
Compute which limit binds a workload, and how prompt caching changes the answer on the Claude API
Read the rate-limit headers and usage fields to measure headroom per model and workspace
Tell apart a rate-limit 429, an acceleration-limit 429, a spend-cap 429 and a 529, and respond to each correctly
Plan workspaces, tiers and model mix so that one workload cannot starve another

On the last Friday of the month the platform completes a few hundred purchases across a dozen firms. Every completion wakes several agents: one reconciles the completion statement against the lender's redemption figure, one drafts the post-completion letters, one prepares the stamp duty return, one queues the Land Registry application. At 10:40 the error rate jumps. Nothing is down. The provider is doing exactly what it said it would: your organisation has used its allowance of output tokens for this minute, and every extra request comes back 429. This lesson is about knowing that allowance in advance, measuring how close you are to it, and telling the different "no"s apart.

Rate limits and quotas

Rate limits and quotas

Rate limits and quotas

Sign in to keep reading.

Sign in to keep reading.