On every platform that serves Claude, rate limits are allocated per model or per model family, not as one org-wide pot. A team hammering Opus 4.8 into a 429 wall may simultaneously be sitting on an untouched Haiku 4.5 allowance of the same nominal size. Understanding exactly how the buckets are drawn on your platform turns model choice from a pure quality/cost decision into a capacity decision too.
How the buckets are drawn, platform by platform
Claude API: limits apply per model, with documented sharing rules. The Opus limit is a combined bucket across Opus 4.8, 4.7, 4.6, and 4.5 — spreading traffic across Opus versions buys you nothing. Sonnet 4.6 and 4.5 share a bucket, but Sonnet 5 has its own separate limit, and Haiku 4.5 has its own pool distinct from the older Haiku 3/3.5 models. Fable 5 is its own bucket with markedly lower defaults: at the Start tier, 500K ITPM versus 2M for the Opus/Sonnet/Haiku classes (published figures, subject to change). Every model class also scales differently by tier.
Google Vertex AI: the bucket boundary is the model lineage. Claude models launched after May 26, 2026 share one quota bucket per lineage per location — the base_model dimensions are anthropic-claude-opus, anthropic-claude-sonnet, anthropic-claude-haiku, and anthropic-claude-fable. All Opus versions drain one bucket; Haiku traffic never touches it. Global-endpoint quota and each multi-region-endpoint quota are additionally independent of each other.
Amazon Bedrock: quotas are defined per model on each endpoint ("Bedrock Mantle input tokens per minute for model"), and the two inference endpoints — bedrock-runtime and bedrock-mantle — track quota separately even for the same model.
Microsoft Foundry: quota is subscription-level and shared across all Global Standard deployments of the same model and version — but different models have different default allocations (for example, 40 RPM / 40K ITPM for Opus-family and Sonnet 5 on pay-as-you-go versus 80 / 80K for Haiku 4.5 and Sonnet 4.6/4.5, per Microsoft's published defaults, subject to change).
The routing play
The strategic consequence: every request you route to Haiku 4.5 is quota you did not spend from the Opus, Sonnet 5, or Fable bucket. High-volume, low-complexity traffic — classification, extraction, routing decisions, first-pass triage, formatting — rarely needs a frontier model. Sending it to Haiku does three things at once: it conserves the scarcer premium-model quota for genuinely hard work, it adds an entire independent TPM pool to your effective org-wide throughput, and it cuts unit cost (Haiku 4.5 lists at $1/$5 per million tokens versus $5/$25 for Opus 4.8 and $10/$50 for Fable 5 — list prices, subject to change).
The scarcity ordering matters most at the top. Fable 5's rate limits are a fraction of the other classes' on the Claude API, and on Foundry pay-as-you-go it defaults to zero. If any part of your workload truly requires Fable 5, treat its bucket as the binding constraint of the whole system and design everything else to stay out of it.
Caveats before you celebrate
Separate buckets are not a loophole for identical work. Version-shuffling inside a family usually doesn't help — the Opus versions share one bucket on the Claude API, and all Opus versions share one lineage bucket on Vertex. Sonnet 5 having its own Claude API bucket separate from Sonnet 4.6/4.5 is a genuine exception, but the two generations use different tokenizers, so migrated traffic consumes tokens differently (see the tokenizer article). Haiku 4.5 also has a 200K context window versus 1M on the larger current models, so it cannot absorb your long-context traffic. And quality is the ceiling on all of it: routing to a smaller model only conserves quota if the smaller model's answers are actually acceptable — evaluate before you migrate, not after.
Also remember that on the Claude API, rate limits are shared across inference_geo values — routing traffic between "us" and "global" draws from the same per-model pool, so geography is not another bucket.
Where to go next
Model tiering covers the quality/cost side of the same routing decision, and prompt caching is the other big multiplier on whatever bucket you're in. For platform-specific quota anatomy, see Bedrock quota types, Vertex quota types, and Foundry quota types.