Team planner / API comparison

API vs self-hosting: calculate your break-even

Published 7 September 2026 · Updated 10 September 2026

A GPU can look inexpensive by the hour and still be expensive for your team. The difference is how much useful work it does while you are paying for it. A coding assistant used on weekdays and a customer-facing service running around the clock need different budgets, even if they use the same model.

To compare self-hosting with an API, work out the monthly bill, the API bill for the same workload, and whether the GPU can deliver that workload at an acceptable speed. A break-even calculation is only useful when all three agree.

Build a monthly budget you can check against an invoice

Start with the full instance rate, including the CPU and memory needed to serve the model. Add persistent storage, any separately billed network or gateway resources, platform fees and the time your team spends operating the service. Include startup and idle hours in the compute line.

Monthly self-hosting cost =
  instance hourly rate × billed hours
  + storage + network + platform fee + operating time

Stopped compute and deleted infrastructure are different budget states. AWS bills EBS storage while an instance is stopped. RunPod also bills retained storage, with different rates for volume disks and network volumes. Put the resources you keep overnight into the estimate.

A worked example: weekdays versus always on

The calculator starts with a hypothetical $2/hour instance, $20/month for storage and a $39/month platform fee based on LLM Hangar Lab. Both site calculators use a 730-hour planning month; eight-hour weekdays use 22 working days. Operating time, network charges and tax are excluded from this first example so you can add your own.

The eight-hour weekday schedule reduces compute hours by about 76% before startup. Fixed costs remain. The ten-hour business schedule uses 217 serving hours, an average-month estimate. Use Custom for your actual calendar, serving hours and starts. For always-on service, a first boot can reduce serving time within the month's billed hours.

Add two operating hours at $100/hour and the $422 subtotal becomes $622/month. At the same $1.50 blended API rate used below, break-even rises from 281 million to about 415 million tokens. The calculator shows infrastructure, platform and operating time separately.

A recorded deployment: GLM-5.3-Flash on four H200s

Our 27 August 2026 deployment record gives a concrete starting point. We launched the native FP8 build on RunPod in US-GA-2 with four H200 GPUs, tensor parallelism 4 and a 32K served context. The observed rate was $18.36/hour and the cold boot took about 34 minutes. These are historical observations.

The chat checks establish that the endpoint answered. Their input lengths and concurrency were not recorded as a controlled workload, so they do not establish sustained aggregate throughput or a capacity limit. Load the recorded deployment above to explore costs with this rate and boot time; storage and other costs remain explicitly editable assumptions. A four-GPU configuration also needs an appropriate platform plan: replace the example fee with the applicable quote.

Calculate the API bill from your token mix

Input and output tokens can have very different prices. Cached input, batch processing, longer contexts and additional tool charges can change the bill again. Use the tariff you would actually buy and count each category separately.

API cost = (uncached input tokens × input price
          + cached input tokens × cached-input price
          + output tokens × output price) / 1,000,000
          + other usage charges

For a simple example, assume hypothetical prices of $1 per million input tokens and $3 per million output tokens, with three input tokens for every output token and no caching. The blended price is $1.50 per million total tokens. A $422 monthly self-hosting bill breaks even at about 281 million tokens per month. At a blended price of $0.30, it takes about 1.41 billion.

This calculation assumes the models deliver comparable results. If one model needs extra attempts, longer reasoning or human corrections, compare the cost per completed task as well. A cheaper token can produce a more expensive answer.

Check that your GPU can reach break-even

In the $422 example, serving 281 million tokens during 176 available hours requires roughly 444 total tokens per second while the service is open. That is a capacity target derived from the budget, not a performance claim. It combines input and output tokens, so it must be checked using the same three-to-one mix.

Replay representative requests at normal and peak concurrency. Record successful requests, input and output throughput, time to first token and the slowest responses your users are likely to encounter. vLLM's serving benchmark exposes these metrics. A throughput result achieved only with long queues may be acceptable for an overnight batch and frustrating for an editor.

For the three-to-one input/output mix above, 444 combined tokens/second means roughly 333 input and 111 output tokens/second across the whole service. An individual request's decode speed is a different measurement. Enter measured aggregate throughput in Advanced to draw a monthly capacity limit on the chart.

The flat hosting line assumes one fixed deployment. Once you need another instance, its costs must be added and break-even recalculated. The monthly capacity estimate also assumes sustained work throughout serving hours; bursts, outages and uneven demand require headroom.

Include the work that remains after deployment

Provisioning automation reduces setup work, but someone still owns model evaluation, access decisions, updates, incident response and application support. Estimate those hours explicitly. Avoid treating a platform subscription as a substitute for the whole operating budget.

LLM Hangar runs deployments in your cloud account and provides budget controls and stop and resume. Your provider bills the infrastructure separately. The cost calculator is a useful starting estimate; a short representative run and the provider invoice make the decision more reliable.

Questions people ask