Self-host vs API break-even calculator
Short answer: Self-hosting wins when utilization stays above 11.1% for these inputs. With the example inputs — $2.00 per GPU-hour, 500 output tokens/sec and 50% utilization — one GPU serves 657,000,000 output tokens a month for $1,460, which is $2.22 per 1M output tokens versus $10.00 per 1M output on Claude Sonnet 5.5. The break-even GPU rate at 50% utilization is $9.00/hour.
Every number in the receipt is an estimate from the formula below. The only sourced price is the API list rate; the GPU cost and throughput are your inputs.
The formula
With a GPU or multi-GPU node billed at G dollars per hour, generating T output tokens per second while active, active U percent of the time, and an API price P per 1M tokens:
self-host $/1M output tokens = (G × 1,000,000) ÷ (T × 3,600 × U)
break-even U = (G × 1,000,000) ÷ (T × 3,600 × P)
break-even G = (P × T × 3,600 × U) ÷ 1,000,000
U is a fraction in the formulas (50% = 0.5) and prices are USD per 1M tokens. The 730-hour month cancels out: monthly GPU cost is G × 730 and monthly output is T × 3,600 × 730 × U, so dividing cost by tokens removes the month length. That is why the result is a rate, not a monthly bill.
| Symbol | Meaning | Where it comes from |
|---|---|---|
| G | GPU or node cost per hour | Your quote or invoice |
| T | Output tokens per second while serving | Your measured throughput on your model and batch size |
| U | Share of wall-clock time the GPU is generating | Your serving logs or queue metrics |
| P | API price per 1M tokens | Official provider pricing page, checked 2026-11-15 |
Reading the result
- If your real utilization is above the first break-even, the GPU is cheaper than the API output price at these inputs. Below it, the API is cheaper.
- If your GPU quote is above the break-even $/hour, the API is cheaper at the utilization you entered.
- The blended row compares a workload with 3 input tokens per 1 output token. That rate is lower than the output price, so it is the harder test for self-hosting.
- A break-even above 100% means one GPU cannot generate enough tokens in a month to match the API rate at these inputs.
Choose inputs you can defend
- Use the all-in hourly cost of the box: rental or amortised hardware plus power, networking and the operations time you would not otherwise spend.
- Measure T with the context length and batch size you expect, on the model you would actually serve. Vendor benchmarks for a different model or batch size are not a substitute.
- Measure U from logs. Peak load is not utilization; a GPU idle half the day is at 50%, not 100%.
- For a multi-GPU node, enter the node's hourly cost and its aggregate output tokens per second as one unit.
What the estimate leaves out
- Hardware purchase, replacement, spare capacity, redundancy and disaster recovery.
- Storage, egress, networking, rack space and electricity beyond the hourly cost you enter.
- Operations: deployment, monitoring, on-call, security, upgrades and model evaluation.
- Input/prefill processing. The self-host side counts generated output tokens only, which biases the comparison in favour of self-hosting on prompt-heavy workloads.
- Failures, retries, cold starts and queueing that reduce effective throughput; quantisation and batching effects on quality.
- API-side caching, batch discounts, long-context surcharges, negotiated rates and rate limits. The API rate used here is the standard list price.
Where the API price comes from
The model selector uses the same verified list prices as the price table; every model page links its official provider page and check date. The default, Claude Sonnet 5.5, comes from Anthropic pricing, checked 2026-11-15. No GPU price is published on this page; GPU costs are inputs because they depend on your hardware, region and contract. How prices are checked and how unverified ones are excluded is documented in the verification method.
Questions
Is self-hosting cheaper than an API?
Not in general; it depends on the GPU rate, utilization and the API price. With the example inputs on this page ($2.00 per GPU-hour, 500 output tokens/sec, 50% utilization), one GPU costs $2.22 per 1M output tokens against $10.00 output on Claude Sonnet 5.5, so self-hosting is cheaper above 11.1% utilization. Change the inputs to match your quote and logs.
What utilization should I use?
The share of wall-clock time the GPU is actually generating tokens at load. Read it from serving logs or queue metrics over a full week, not from a peak-hour sample. A box that is busy four hours a day and idle the rest is at about 17%.
Which GPU or model should I use?
This page does not recommend hardware and does not publish GPU prices or throughput benchmarks. Enter the hourly cost of the box you would rent or buy and the tokens/sec you measured on the model and batch size you would serve.
Why compare with both the output price and a blended price?
Tokens/sec measures generation, so the output price is the like-for-like API rate. A blended price includes cheaper input tokens, which is the right comparison for prompt-heavy work but a harder test for self-hosting. Input/prefill processing is not counted on the self-host side, so the estimate already leans in self-hosting's favour.
Does the 730-hour month matter?
No. Monthly cost is G × 730 and monthly output is T × 3,600 × 730 × U, so 730 cancels when you divide cost by tokens. The result is a rate per 1M tokens; any month length gives the same rate.
Next step → See how the API price changes with the token mix
Price data last checked 2026-11-18 against official sources · how we verify
Price-change alerts
One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.