The short version: when Anthropic says your tier gets 2 million input tokens per minute, they don't mean 2 million tokens from every request you send. For all current Claude models — Haiku 4.5, Sonnet 5, Opus 4.8, and Fable 5 — cached input tokens don't count toward that limit. Only uncached tokens do.
If you're building anything with a substantial, reusable system prompt, this changes the math considerably. Let's walk through it.
What the tier consolidation actually did
Before June 26, Sonnet and Haiku users hit lower rate limits than Opus users at the same usage tier. After June 26, all three model classes share the same numbers:
Start tier: 1,000 requests per minute, 2 million input tokens per minute, $500/month spend cap.
Build tier: 5,000 RPM, 5 million ITPM, $1,000/month cap.
Scale tier: 10,000 RPM, 10 million ITPM, $200,000/month cap.
That's a genuine improvement if you're building on Haiku 4.5 ($1 per million input tokens) or Sonnet 5 ($2 per million through August 31). You now get the same rate limit headroom as someone running the more expensive Opus 4.8 ($5 per million).
But the tier numbers themselves are almost beside the point. Here's why.
Cache-aware ITPM: what it means in practice
Anthropic's rate limit page describes a feature they call "cache-aware ITPM." The mechanics: when you cache part of your prompt — your system instructions, a knowledge base, a long document — and subsequent requests read from that cache, those cache-read tokens don't count toward your input token rate limit.
Their own example from the docs: "With a 2,000,000 ITPM limit and an 80% cache hit rate, you could effectively process 10,000,000 total input tokens per minute (2M uncached + 8M cached), because cached tokens don't count towards your rate limit."
That's a 5x multiplier on your headline limit. And you get it by designing the system correctly.
A concrete example — the FAQ bot
Say we build a customer FAQ bot for your business. It has a 40,000-token system prompt — your knowledge base, product details, policies, tone guidelines. Every conversation starts with that context.
Without caching: each conversation sends roughly 40,200 tokens as uncached input (40,000 context + ~200 user message). On the Start tier (2M ITPM), you'd cap out around 47 conversations per minute before hitting the limit.
With caching: the 40,000-token knowledge base gets written to cache once. Every subsequent request reads from cache. Only the 200-token user message counts toward ITPM. Suddenly your effective capacity is around 10,000 conversations per minute — same tier, same price, different architecture.
The cost math shifts too. Cache reads cost 10% of the standard input price. Those 40,000 cached tokens, which would run $0.04 per conversation at standard Haiku 4.5 rates, cost $0.004 instead. Over 10,000 conversations per month, that's $360 you don't spend.
The rate limit number in your console is a floor. Caching is the multiplier that makes it real.
This is one of those things that separates a quickly-built prototype from a workflow designed to actually scale. The prototype sends the full context every time. The production system caches it. They look identical to the end user and have very different unit economics.
What this means right now
Two things are true simultaneously. First, the June rate limit changes give small business AI workflows meaningfully more room to grow before anyone needs to talk to Anthropic's sales team about a custom limit. Second, Sonnet 5's introductory pricing ($2/$10 per million input/output tokens) runs through August 31. After that, the standard rate kicks in at $3/$15 — a 50% increase.
Combine those two facts and the next seven weeks are a reasonable time to scope and start building an AI-powered workflow: the model costs less, the rate limits are higher than they've ever been for Sonnet-class work, and if you build it with prompt caching from day one, your effective throughput is much larger than the headline numbers suggest.
We've had a few clients ask recently whether they'd "hit a wall" if business picks up. With the Build tier ($1,000/month spend cap, 5,000 RPM, 5M ITPM), and an architecture that uses caching correctly, the answer for most small business volumes is no. The wall is considerably further than it looks.
If you're thinking about building something and want a sanity check on whether the rate limits or costs would actually constrain your use case, that's the kind of question we dig into in a discovery call. We can also usually tell you whether you need a Sonnet-class model or whether Haiku 4.5 would do the same job for half the price — which is often the more useful question anyway.
For the per-token cost math on specific workflows, our earlier post on what AI agents actually cost to run still holds up; just substitute in the current model prices.
— Cole
Sources
- Anthropic API Rate Limits documentation — verified June 26, 2026 tier structure and cache-aware ITPM details
- Anthropic Pricing documentation — model pricing, Sonnet 5 introductory rate through August 31, 2026