The cheapest LLM API, ranked — and why row one is a trap

The ranking

Every model in our catalog, costed on the same workload — a history-carrying chatbot at 1,000 requests/day, with 8% retries, a 40% cache hit rate, 15% infra overhead, and per-model tokenizer calibration applied. Sorted cheapest first.

Top to bottom that is a 66× spread on identical work — Amazon Nova Lite at $10/mo against Claude Sonnet 4.6 at $628/mo.

Why the top of this list is a trap

If you stop reading at row one you will pick Amazon Nova Lite and regret it. The cheapest rows here are small models — 8B-class and “lite” tiers — and they are cheap because they are less capable, not because someone is being generous. Our own recommendation engine refuses to recommend budget-tier models for the coding-assistant and multi-model-router archetypes at all, regardless of price.

The failure mode isn’t that a weak model costs more in retries. It’s that on work it can’t do, retries don’t converge — you get a plausible wrong answer, and the cost lands on whoever has to notice and fix it. A ranking by $/call quietly assumes every row can complete the task. Across a 66× spread, that assumption is false.

Caching reshuffles the order

6 of the 18 models here support no prompt caching at all. On a workload with a large reusable prefix that is a bigger handicap than a higher sticker price. The cheapest model that supports caching is Amazon Nova Lite at $10/mo — already row one, and its lead only widens as the cache hit rate goes up, because the rows chasing it cannot cache at all.

If you need batch as well — offline pipelines, bulk classification — the cheapest model offering both caching and a batch tier is Amazon Nova Lite at $10/mo. Those two capabilities eliminate more of this table than price does.

Where price isn’t the differentiator

Some rows are exact ties, which means the decision is made on something other than cost:

• $147/mo — GPT-5.4 mini and GPT-5.4 mini (Azure)
• $490/mo — GPT-5.4 and GPT-5.4 (Azure)

Identical pricing across a first-party API and its cloud-hosted equivalent is common. Pick on region, compliance posture, billing relationship, and rate limits — the cost model has nothing to say.

How to actually use this table

• Pick the tier first, then the row. Establish whether your task needs frontier reasoning. Then take the cheapest option inside that tier, not the cheapest overall.
• Filter on capabilities before price. Caching, batch, and context window remove options outright. A cheaper model that can’t cache your 8,000-token prefix is not cheaper.
• Check the non-price constraints. Data residency, provider region, and procurement policy disqualify rows regardless of where they rank — and that decision usually isn’t yours alone to make.
• Re-run it on your numbers. This ranking uses one workload shape. Change the input:output ratio and the order shifts, because models price input and output independently.