Token pricing determines the cost of processed work. Requests per minute determines whether that work can enter the system on time. For interactive products, agents, and bursty batch pipelines, an RPM limit can become the binding constraint long before the monthly token budget is exhausted.
This is why capacity belongs next to price in every provider comparison.
1. RPM Limits Determine Whether Demand Can Be Served
Consider an application that needs 600 requests per minute during its busy period. Provider A has a lower token price but allows only 100 RPM. Provider B costs more per token and allows 1,000 RPM.
Provider A is not automatically the lower-cost option. The application must queue requests, obtain several accounts or capacity allocations, add other providers, or accept dropped traffic. Each response may be inexpensive, but the system cannot deliver the required product at the required time.
RPM is especially important for short requests. A classification or routing call may use very few tokens, so TPM consumption remains low while request count grows quickly. Agent workflows amplify the issue because one user action can trigger several model calls. One hundred user operations per minute may become 800 model requests per minute if the average workflow has eight steps.
Bursts matter too. An hourly average of 100 RPM says little about a product that receives 1,000 requests in the first minute after an event. Buyers should understand whether the provider uses a fixed window, rolling window, or token bucket and how much short-term bursting it permits.
An account limit is still not the same as backend capacity. A provider may allow 5,000 RPM contractually while latency rises before the workload reaches that level. The limit describes admission; a load test describes observed service behavior.
2. Evaluate RPM Together With TPM and Concurrency
RPM counts request arrivals. TPM describes token volume. Concurrency describes how many requests remain active. Any one of them can be the first constraint.
The relationships are approximately:
TPM = RPM × average tokens per request
concurrency = requests per second × average request duration
At 1,000 RPM, a 200-token request represents about 200,000 TPM. A 5,000-token request at the same RPM represents about 5 million TPM. If requests take longer because outputs are large, concurrency rises even when RPM stays fixed.
This changes provider selection by workload. Short DeepSeek classifications may need high RPM but modest TPM. Long document analysis may hit TPM or concurrency first. A multi-step agent may encounter all three: high request count, repeated context, and several active streams.
Ask providers for every relevant limit and verify how the API reports it. Record 429 responses separately from upstream errors. Test whether excess requests are rejected quickly or wait in a queue, because the second behavior can appear as a latency problem rather than an explicit rate limit.
Capacity planning should also leave headroom. Operating continuously at the stated maximum provides no room for output-length variation, traffic spikes, or slower requests that increase concurrency.
3. Price the Capacity You Can Actually Use
Token Forge Cloud tested increasing DeepSeek traffic with fixed pacing rather than inferring capacity from an account limit. Under a short-request workload, the highest stable observed levels were:
| Model | Stable tested load | Token throughput | Success rate | TTFT p95 |
|---|---|---|---|---|
deepseek-v4.1-flash |
5,000 RPM | >1.21M TPM | 100% | 1.97 s |
deepseek-v4-pro |
2,500 RPM | >617K TPM | 100% | 2.82 s |
kimi-k3 |
5,000 RPM | ~5M TPM | 100% | 1.26 s |
glm-5.3 |
4,200 RPM | ~5.4M TPM | 99.95% | 5.62 s |
glm-5.3-flash |
2,400 RPM | ~3.0M TPM | 100% | 5.29 s |
The Kimi K3 and GLM rows come from separate tests, the Kimi K3 load test report and the GLM load test report: about 1,000 input and up to 300 output tokens per request, Kimi K3 with reasoning off, and GLM at low reasoning (it cannot be turned off). GLM is held to a TTFT p95 of 8 seconds rather than 5.
The stability rule required at least 99% success and TTFT p95 no higher than five seconds. This makes the RPM figure more informative: it names the service quality maintained at that load for the tested workload.
A provider comparison should therefore calculate two things. First, estimate token spend from real prompt and output distributions. Second, verify that the endpoint sustains the required RPM, TPM, and concurrency without violating latency or completion targets.
If a cheaper provider covers only part of the traffic, include the cost of overflow routing and duplicate infrastructure. If queueing delays make a result arrive too late to be useful, count it as a failed business outcome even if the token price was low.
The official price table tells you the rate charged for accepted work. The RPM envelope tells you whether enough work can be accepted to run the application. When demand is higher than capacity, the second number is more important.