allocateThe cheapest LLM APIs in 2026
The cheapest capable LLM APIs in 2026 are open-weight models: Llama 4 70B from $0.07 per million input tokens, Qwen 3.5 at $0.08, DeepSeek V4 at $0.09, and Kimi K2.5 at $0.10. Gemini 3.5 Flash at $0.15 is the cheapest frontier option. The bigger savings usually come from caching and routing, not from the list price.
From $0.07 per million input tokens, and entirely capable for classification, extraction, and routine support turns, which is most of what high-volume workloads are.
$0.08 per million tokens with quality across 100+ languages. For non-English volume there is no cheaper way to hold quality.
Near-frontier reasoning at $0.09 per million tokens, roughly a tenth of what proprietary reasoning models charge for comparable work on most tasks.
$0.15 per million tokens with 310 ms latency and 2M context. When you want a frontier lab behind the model but the workload is high-volume, this is the floor.
List price is a third of the story
Output tokens cost 4 to 5x more than input on most providers, 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: repeated system prompts and tool definitions bill at roughly a tenth of the input rate. An agent with a stable 3,000-token system prompt and high cache hit rate can run cheaper on a mid-priced model than a naive setup on the cheapest one.
The durable saving is routing: send the routine 80% 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, Llama 4 70B at $0.07 per million input tokens, followed by Qwen 3.5 at $0.08 and DeepSeek V4 at $0.09. 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 20 to 40x 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?
Cached input bills at roughly 10% of the normal input rate. For agents and assistants with stable system prompts, 40 to 80% cache hit rates are normal, which can cut 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.