When a team plans for a throughput target — say, "we need to classify 3,000 documents a minute" — the instinct is to pick the model first and figure out capacity later. It pays to run the logic the other way. On every platform, the most capable models carry the tightest limits, and the fast, cheap models carry the loosest ones. The familiar advice "use the cheapest model that meets your quality bar" is usually justified on cost grounds; the less-quoted second half is that it simultaneously maximizes how many requests per minute your existing quota allows.
The spread on the Claude API
Anthropic's published rate limits illustrate the gap. The figures below are the documented defaults as of July 2026 and are subject to change — treat the ratios, not the absolute numbers, as the lesson, and check the rate-limits page for current values.
| Tier | Haiku 4.5 / Sonnet 5 / Opus 4.x class | Claude Fable 5 |
|---|---|---|
| Start | 1,000 RPM · 2M ITPM · 400K OTPM | 1,000 RPM · 500K ITPM · 100K OTPM |
| Build | 5,000 RPM · 5M ITPM · 1M OTPM | 2,000 RPM · 1.5M ITPM · 300K OTPM |
| Scale | 10,000 RPM · 10M ITPM · 2M OTPM | 4,000 RPM · 4M ITPM · 800K OTPM |
Within the same usage tier, Haiku 4.5 supports 2.5x the RPM of Fable 5 at Build and Scale, with over three times the input-token throughput and more than double the output throughput. Combine that with list pricing — Haiku 4.5 at $1/$5 per million input/output tokens versus Fable 5 at $10/$50 — and the requests-per-minute-per-dollar gap between the two ends of the lineup is enormous. Fable 5 exists for workloads that need the highest available capability; pointing it at a high-volume classification queue burns both budget and quota.
The same pattern is sharper on Microsoft Foundry: on pay-as-you-go subscriptions, Haiku 4.5 gets double the default RPM and ITPM of the Opus-family and Sonnet 5 models, while Fable 5 defaults to zero — it is simply not available at that subscription level (see the Foundry pay-as-you-go vs Enterprise gap). On Bedrock the shape differs again — Opus 4.7 and 4.8 on the runtime endpoint have no RPM quota at all and are governed purely by token quotas — but the principle of checking per-model limits before committing holds everywhere.
Separate pools are free parallelism
There is a second, less obvious throughput bonus in mixing models. Rate limits are enforced per model class: the Opus 4.x versions share one combined bucket, but Sonnet 5 has its own separate limit, and Haiku 4.5 has its own pool (separate even from the older Haiku 3.5). A pipeline that routes easy items to Haiku 4.5 and hard items to Opus 4.8 is not just cheaper per item — it is drawing on two independent quota pools at once, so the two workloads never throttle each other. Each model family has its own quota pool explores this, and routing to Haiku covers the dispatch implementation.
What "meets quality requirements" has to include
Right-sizing only works if the quality bar is evaluated honestly, and that evaluation should include hard constraints, not just output quality:
Context window. Haiku 4.5 has a 200K-token context window where Fable 5, Opus 4.8, and Sonnet 5 have 1M. If your requests routinely carry hundreds of thousands of tokens of context, Haiku is out regardless of quality — no throughput math rescues a request that doesn't fit.
Max output. Haiku 4.5 tops out at 64K output tokens versus 128K for the 1M-context models.
Feature needs. Capabilities like adaptive thinking vary by model — Haiku 4.5 supports extended thinking but not adaptive thinking, for example. Check the models overview for the workload's specific requirements.
A practical sizing loop: prototype the task on a mid-size model (Sonnet 5 is the balanced workhorse, currently at introductory pricing through August 31, 2026), build a small evaluation set, then test whether Haiku 4.5 passes it. If it does, you have just multiplied your effective capacity and divided your cost. If it almost does, consider whether prompt improvements close the gap before reaching for a bigger model — and remember that prompt caching raises effective throughput on any model you pick, since cache reads don't count toward ITPM on current models.
Where to go next
The model lineup explainer covers what each tier is for, and capacity planning turns a model choice plus a traffic forecast into a quota request.