AllocateThe cheapest LLM APIs in 2026
The cheapest capable LLM APIs on the catalog in 2026 are open-weight models: GPT-OSS 20B at $0.05 per million input tokens, Mistral Small 24B at $0.10, GPT-OSS 120B at $0.15, Llama 4 Scout at $0.18 with a 1M window, and Qwen3 235B at $0.20. The bigger savings usually come from caching and routing, not from the list price.
$0.05 per million input tokens and $0.20 out under Apache 2.0. Entirely capable for classification, extraction, and routine support turns, which is most of what high-volume workloads are.
$0.10 per million input tokens, Apache 2.0, from a European lab. The floor price for a well-documented dense model with a permissive license.
A 120B open model at $0.15 per million input tokens and $0.60 out. The step up in capability from the small models costs cents, not dollars.
$0.18 per million input tokens with a 1M-token window. Nothing else on the catalog offers seven-figure context anywhere near this price.
Provider list prices from the Allocate catalog, checked 2026-07-21.
List price is a third of the story
Output tokens cost more than input on most models, up to 10x on some, so a workload that generates long answers can make a cheap model expensive. Price your real input-output split, not the headline number.
Prompt caching changes the ranking: models with published cached-input rates bill repeated system prompts and tool definitions at a fraction of list. An agent with a stable 3,000-token system prompt and a high cache hit rate can run cheaper on a mid-priced model with caching than on the cheapest one without.
The durable saving is routing: send the routine majority of traffic to a model in this list and reserve expensive models for the steps that need them. The price comparison tool ranks every model on your exact workload, cache rate included.
Common questions
What is the cheapest LLM API right now?
Among capable production models on the catalog, GPT-OSS 20B at $0.05 per million input tokens, followed by Mistral Small 24B at $0.10 and GPT-OSS 120B at $0.15. Tiny models go lower still, from $0.02. Prices move, so put your own volumes through the price comparison tool before budgeting.
Why are open-weight models so much cheaper?
Competition. Anyone can serve an open model, so the price falls toward the real cost of the hardware. Proprietary models are only served by one provider, which keeps a margin in the price. The quality gap is far smaller than the price gap on most workloads.
Is a cheap model good enough for production?
For classification, extraction, routing, and routine support turns, usually yes, and those are most of a real workload’s volume. The failure mode is using a cheap model for judgment-heavy steps; route those to a stronger model instead of upgrading everything.
How much does caching actually save?
Where a cached rate is published it is a deep discount: GPT-5.5 lists cached input at $0.50 against $5, and GLM 5.2 at $0.26 against $1.40. For agents and assistants with stable system prompts, that cuts the input side of the bill by more than half.
Do these prices include fine-tuned models?
Serving a fine-tuned open model costs about the same per token as its base. That is the quiet advantage of open weights: a model trained on your data, at commodity serving prices, with the weights owned by you.
Related
Every model here sits behind one key on Allocate: route by name, meter per route, and swap the model in one click.