The Bedrock throughput wall
Budget review is not what stops most Bedrock projects. Throttling is. Every on-demand model carries a per-account, per-region ceiling on tokens per minute and requests per minute, and crossing either returns an error rather than a queue. These are the published defaults — the number you start from, and the number that decides whether you need a support case before launch or never.
Read the gap before the table
AWS Service Quotas exposes plenty of Bedrock limits that are account-wide or cannot be attributed to one model. Those are deliberately not attached to anything here — a guessed quota is worse than a missing one. So 86 available models show no row, and that means AWS publishes no per-model default for them, not that they are unlimited. For those, the only reliable number is the one in your own Service Quotas console for your own account.
The figure worth knowing before you design around it: image models are metered on requests alone, and the default is small. Stable Image Conservative Upscale defaults to 2 requests per minute in us-east-1. A batch job that walks a product catalogue will hit that in the first few seconds.
Token-metered models
Each model's most generous published region. "Runs / min" converts the token ceiling into requests of 12,000 tokens — a realistic agentic call once a system prompt, tool definitions and some history are in the window — and "binds" names whichever limit runs out first at that size.
| Model | Tokens / min | Requests / min |
|---|---|---|
| DeepSeek V3.2 DeepSeek | 100M | 10,000 |
| GPT OSS Safeguard 120B OpenAI | 100M | 10,000 |
| GPT OSS Safeguard 20B OpenAI | 100M | 10,000 |
| Kimi K2 Thinking Moonshot AI | 100M | 10,000 |
| MiniMax M2 MiniMax | 100M | 10,000 |
| MiniMax M2.1 MiniMax | 100M | 10,000 |
| MiniMax M2.5 MiniMax | 100M | 10,000 |
| Ministral 14B 3.0 Mistral AI | 100M | 10,000 |
| Mistral Large 3 Mistral AI | 100M | 10,000 |
| Nemotron Nano 3 30B NVIDIA | 100M | 10,000 |
| NVIDIA Nemotron 3 Super 120B A12B NVIDIA | 100M | 10,000 |
| Qwen3 Coder Next Qwen | 100M | 10,000 |
| Qwen3 Next 80B A3B Qwen | 100M | 10,000 |
| Qwen3 VL 235B A22B Qwen | 100M | 10,000 |
| Writer Palmyra Vision 7B Writer | 100M | 10,000 |
| Mistral Large (24.07) Mistral AI | 300K | 400 |
Request-metered models
Image and video models are not billed or limited per token, so only a request rate is published. These are the smallest numbers on the page by a wide margin.
Planning against these
Size the workload in tokens per minute, not requests per day
The unit that meets the limit is tokens per minute at peak, and peak is rarely the daily average divided by 1,440. An agent loop that fires eight model calls per user action multiplies the request count without anyone noticing in the design review. The calculator works in the same units.
A cross-region inference profile widens the ceiling
Routing through a profile spreads traffic across several regions' capacity instead of one, which is a real answer to a throughput problem. It is also the thing that moves where your data is processed, so it is not free — check what it commits you to before reaching for it.
Caching buys throughput, not just money
Cached input tokens are cheaper, but on a TPM-bound workload the more valuable effect is that they still count against a ceiling you can raise more easily than you can raise latency. Batch is the other lever: it runs against separate capacity and is the right home for anything that does not need an answer this second.
Questions
What are the default Bedrock quotas for a model?
Bedrock meters on-demand access per account, per region and per model with two limits: tokens per minute (TPM) and requests per minute (RPM). Hitting either returns a ThrottlingException rather than a slower response. AWS publishes a model-specific default for 30 of the 116 models in the catalogue; the rest fall under limits that Service Quotas does not attribute to a single model.
Which limit runs out first, tokens or requests?
For anything long-context, tokens. At a 12,000-token request — a realistic agentic call with a system prompt, tool definitions and some history — a model needs an RPM above its TPM divided by 12,000 before requests become the binding constraint. The table on this page shows which one binds for each model at that size.
Can Bedrock quotas be increased?
The token and request rates are adjustable: you request an increase through AWS Service Quotas or your account team, and it is judged against capacity in that region. That is exactly why the default matters — it tells you whether the conversation needs to happen before you ship, or not at all. Some quotas are not adjustable, and Service Quotas marks those.
Do quotas differ by region?
Yes, and they are published per region — this page reports each model's most generous region so the headline is the best case, with the region named. A model at its default in 15 regions is not necessarily at the same default in all of them, so check the region you will actually deploy in on the model's own page.
Quotas come from the AWS Service Quotas API, matched to models by name, and are account defaults rather than your account's current values. Methodology.