What is the cheapest LLM API right now?

As of 2026-08-23, the cheapest paid text model we track is Mistral Mistral Nemo at $0.019/1M input and $0.03/1M output tokens — about $0.0217 per million blended tokens. Rankings shift with almost every price cut, so check the live table above rather than a blog post from three months ago.

Are the cheapest models good enough for production?

For a lot of traffic, yes — classification, extraction, routing, summarization, and short factual answers run fine on sub-$0.20 models. Where they fall over is long multi-step reasoning, subtle instruction-following, and tool-use reliability. The pattern that works: route everything to a cheap model first and escalate the hard 10–20% to a flagship, which keeps your average cost near the cheap end.

How do I compare prices when input and output rates differ?

Weight them by your actual traffic mix. Most production workloads are input-heavy — context, history, and documents get re-sent every request while answers stay short — so a 3:1 input:output blend is a realistic default, and it is what this page ranks by. If your workload generates long outputs (code, articles), re-rank with output weighted heavier; some models that look cheap flip positions.

Do batch and cached-input discounts change the ranking?

Significantly. Batch APIs typically take ~50% off both rates for async work, and cached input can cut the biggest line item — repeated context — by 75–90% on models that support it. A mid-priced model with strong caching can beat a nominally cheaper one without it on chat-style workloads. Check each candidate's model page for its cached rate before deciding.

The Cheapest LLM APIs in 2026

"Cheapest LLM API" is a moving target — prices get cut mid-quarter, new budget tiers launch monthly, and most rankings you'll find were stale the week they were published. This page is generated from our live price feed (290 paid text models, updated 2026-08-23), ranked by blended cost at a 3:1 input:output ratio — the mix a typical production workload actually bills at. Want the full list? Every model is on the models page with live search.

The 10 cheapest text models right now

#	Model	Input /1M	Output /1M	Blended /1M	Context
1	Mistral Mistral Nemo	$0.019	$0.03	$0.0217	131K
2	Alibaba (Qwen) Qwen3.7 Flash	$0.03	$0.13	$0.055	1M
3	OpenAI gpt-oss-20b	$0.03	$0.13	$0.055	131K
4	Meta Llama 3.1 8B Instruct	$0.05	$0.08	$0.0575	131K
5	Mistral Mistral Small 3	$0.05	$0.08	$0.0575	33K
6	Amazon Nova Micro 1.0	$0.035	$0.14	$0.0613	128K
7	Google Gemma 3 4B	$0.05	$0.1	$0.0625	131K
8	DeepSeek DeepSeek V4 Flash 0423	$0.0517	$0.1033	$0.0646	1M
9	Cohere Command R7B (12-2024)	$0.0375	$0.15	$0.0656	128K
10	OpenAI GPT-5 Nano (batch)	$0.025	$0.2	$0.0688	400K

Two caveats before you ship the #1 pick. First, this floor is mostly small open-weight models hosted cheaply — brilliant for simple high-volume tasks, not for nuanced reasoning. Second, free promo tiers are excluded on purpose: their rate limits make them unusable as production capacity.

Cheapest model from each major vendor

If you need a specific ecosystem — OpenAI-compatible tooling, Anthropic's caching, Google's long context — this is the cheapest way into each one:

Vendor	Cheapest model	Input /1M	Output /1M	Blended /1M	Context
Mistral	Mistral Nemo	$0.019	$0.03	$0.0217	131K
OpenAI	gpt-oss-20b	$0.03	$0.13	$0.055	131K
Alibaba (Qwen)	Qwen3.7 Flash	$0.03	$0.13	$0.055	1M
Meta	Llama 3.1 8B Instruct	$0.05	$0.08	$0.0575	131K
Google	Gemma 3 4B	$0.05	$0.1	$0.0625	131K
DeepSeek	DeepSeek V4 Flash 0423	$0.0517	$0.1033	$0.0646	1M
Z.ai (GLM)	GLM 4.7 Flash	$0.06	$0.4	$0.145	203K
Anthropic	Claude 3 Haiku	$0.25	$1.25	$0.5	200K
Moonshot (Kimi)	Kimi K2.5	$0.45	$2.25	$0.9	262K
xAI	Grok Build 0.1	$1.00	$2.00	$1.25	256K

What the spread means in real money

Priced on our standard support-chatbot workload (1,500 input + 400 output tokens per request, 1,000 requests/day):

Mistral Mistral Nemo (cheapest overall): ~$1.22/month
Alibaba (Qwen) Qwen3.7 Flash (cheap major-vendor tier): ~$2.91/month
OpenAI GPT-5.6 Sol (flagship): ~$210/month

Same workload, 173× spread. That gap is the entire economic case for model routing: even moving half your traffic down-tier pays for the engineering in the first invoice. Run your own numbers in the API cost calculator, or estimate your token counts first with the token counter.

What "cheap" actually costs you

A quality floor: sub-$0.20 models handle routine language work well but fail quietly on multi-step reasoning and strict formats. Measure on your own eval set, not leaderboards.
Weaker tool use: function-calling reliability is usually the first casualty of aggressive distillation — a real problem if you're building agents.
Missing discounts: some budget models lack cached-input pricing or batch endpoints, so their effective price on chat workloads can exceed a pricier model with 90% cache discounts — the prompt caching guide walks that math.
Churn: budget tiers get deprecated fast. Keep the model name in config, not code.

How to actually capture the savings

Pick one cheap candidate from each table above and run your real prompts through it.
Route by difficulty: cheap model first, escalate failures to a flagship — the routing patterns are in how to cut your LLM API bill.
Move async work (summaries, backfills, evals) to batch endpoints for the extra ~50% off — see the batch API guide.
Re-check this page monthly. Prices only move in one direction, and last quarter's pick is rarely still the winner.