The cheapest LLM API in 2026
"Cheapest LLM API" has an obvious answer and a useful answer. The obvious answer is whichever open model is currently being served at the lowest price per million tokens. The useful answer requires one more step, because price per token is not the same as cost per completed task.
Why the sticker price misleads
Three things break the naive comparison:
- Token efficiency differs. A model at half the price that needs twice the output tokens to reach the same answer costs exactly the same. Reasoning models are the extreme case — their thinking tokens are billed.
- Retries are not free. A cheap model that fails schema validation 20% of the time and gets retried is a cheap model with a 20% surcharge.
- Cache behavior dominates at scale. Prompt caching and semantic caching routinely move the bill more than switching providers does.
The pricing tiers, roughly
- Frontier models — highest per-token price, best on hard reasoning and long-horizon agent work. Worth it for a minority of requests.
- Mid-tier models — an order of magnitude cheaper, and genuinely sufficient for most classification, extraction, summarization, and routine chat.
- Small and open models — cheapest per token, often near-free when self-hosted at volume. Strong on narrow, well-specified tasks.
Live prices move constantly. Rather than repeat numbers that go stale, our model catalog shows current price per token alongside measured latency and quality, and the rankings page sorts by quality-per-dollar on real traffic.
The larger saving is not the provider
Teams that cut LLM bills significantly rarely do it by finding a cheaper vendor. They do it by not sending every request to a premium model. Route the easy 60–80% to a mid-tier model behind a quality floor, cache what repeats, and reserve the expensive model for work that needs it.
That is routing, and it stacks with whatever provider pricing you negotiate. See how to reduce LLM API costs.