Every AI team eventually asks the same question: should we keep paying a frontier lab per token, or run an open-weight model ourselves? The honest answer is that it depends on three things: how many tokens you actually use, how busy you can keep a GPU, and which baseline you compare against.
Published break-even estimates vary by roughly fifty times, because they assume different models, hardware, utilization and engineering costs. We summarize what the 2026 analyses agree on, show the formula, and give a decision framework. All prices and break-even figures below are vendor or blog estimates, not benchmarks we ran.
How Token Pricing Actually Works
Frontier APIs charge per million input and output tokens, with output priced several times higher than input. The sources we reviewed list frontier-tier prices in the range of roughly $2-5 per million input tokens and $12-30 per million output tokens, while cheaper open-weight models served by third-party APIs cost a fraction of that. Prices change often, so treat these as orders of magnitude.
Self-hosting flips the cost structure. You pay for hardware or rented GPU time whether you use it or not, so cost per token is mostly a fixed cost divided by how many tokens you actually serve.
The Break-Even Formula
A simple model: cost per million tokens = monthly infrastructure cost ÷ (tokens served per month in millions). Add engineering and operations on top. Utilization is hidden inside the denominator: a GPU used 20% of the time serves five times fewer tokens than one used 100%.
Illustration, not a benchmark: suppose a rented GPU server costs $1,500 a month and, at healthy utilization, serves 1.5 billion tokens a month. That is about $1 per million tokens before engineering time. At 20% utilization the same server serves 300 million tokens and the cost rises to about $5 per million, which can be worse than a mid-priced API.
What the 2026 Analyses Say
Across the guides we reviewed, the common finding is that self-hosting pays off only at high, steady volume (SitePoint, Aliteq and others). The numbers differ:
| Baseline | Reported break-even |
|---|---|
| Frontier APIs (premium tier) | Roughly 2-6 million tokens per workday in one analysis; 10-20 million per day in another, for a 27B-70B open model on a rented H100-class GPU |
| Cheap hosted open-weight APIs | Often 50 million+ tokens per day, because the API is already cheap |
| Utilization threshold | About 60% or higher to beat frontier pricing, per one guide |
The lesson is to compare against the right baseline. If you are paying frontier prices for tasks a smaller open model handles well, the saving is large. If you can already call a cheap hosted open model, the case for running hardware yourself is weak on cost alone.
The Costs Calculators Leave Out
Engineering. One guide estimates two to four weeks of initial work for serving, batching, monitoring and fallbacks.
Operations. The same source puts ongoing overhead around 10-20% of a senior engineer's time, plus on-call risk when the model server is part of a product.
Idle capacity and peaks. You must provision for peaks and pay for the troughs. APIs scale to zero for you.
Quality gap and drift. You own evaluations, upgrades and safety filtering when a better open model ships.
Reasons to Self-Host That Are Not About Cost
Privacy and sovereignty. Regulated data may not be allowed to leave your environment. See our guide to the EU AI Act and the 2026 regulation map for why this is rising.
Latency and offline use. On-device or on-premises models avoid network round trips and work without connectivity. See edge computing.
Customization. Fine-tuned small models can beat general ones on narrow tasks and become a product moat; we cover this in fine-tuned SLM monetization.
Predictability. A fixed monthly bill is easier to budget than usage that spikes with traffic.
A Practical Decision Framework
Start with the API. Prototype on a frontier model to learn what quality you need.
Measure tokens and spikes. Log tokens per task and per day for a month before buying anything.
Test an open model on your own data. If it passes your evaluation, route that task to a cheap hosted open model first.
Self-host when volume is steady and large, or when data rules require it. Keep 60%+ utilization as the target.
Use a hybrid router. Send routine work to a local or open model and escalate hard cases to a frontier model. Costs fall without capping quality. If you want help sizing this, talk to us or look at our AI services.
Frequently Asked Questions
Is self-hosting an LLM cheaper than using an API?
Only at sufficient, steady volume. Several 2026 analyses put break-even against frontier APIs between about 2 and 20 million tokens per day, depending on model size and hardware, and much higher against cheap hosted open-weight APIs. At low volume the API is usually cheaper once engineering time is included.
Are open-weight models as good as frontier models?
For many well-defined tasks such as classification, extraction, summarization and routine coding help, yes, especially after fine-tuning. For the hardest reasoning and long-horizon agent work, frontier models generally still lead. Test on your own data.
What hidden costs does self-hosting have?
Initial engineering (one guide estimates two to four weeks), ongoing operations (roughly 10-20% of a senior engineer), idle GPU time, monitoring, security hardening and model updates.
When should I self-host for privacy rather than cost?
When data cannot leave your environment for legal, contractual or sovereignty reasons, a local or private-cloud model can be the only option regardless of price. Pair it with access controls and logging.