TokenSave logo TokenSave

The Cheapest LLM APIs in 2026, Ranked by Real Cost per Request

· October 6, 2026

The Cheapest LLM APIs in 2026, Ranked by Real Cost per Request

If you only need the lowest bill, the cheapest LLM API in October 2026 is GPT-5 nano, followed by GPT-6 Luna and Qwen 3.8 Flash. But the cheapest model is rarely the right one for every job. Here is every major model ranked by what a real request costs, which cheap models are worth using, and where paying more saves money.

🧮 Open the token counter →

How we ranked them

Price per million tokens is misleading on its own, for two reasons:

So we priced one typical request: 2,000 input tokens and 500 output tokens of English, counted the way each model's tokenizer would count it, at official list prices checked October 1, 2026. No caching or batch discounts.

The cheapest LLM APIs

The cheapest LLM APIs: Model, Maker, Input / output per 1M, 1 request, 10,000 requests

Model Maker Input / output per 1M 1 request 10,000 requests
GPT-5 nano OpenAI $0.05 / $0.40 $0.00030 $3.00
GPT-6 Luna OpenAI $0.10 / $0.50 $0.00045 $4.50
Qwen 3.8 Flash Qwen $0.15 / $0.47 $0.00053 $5.35
GPT-4o mini OpenAI $0.15 / $0.60 $0.00060 $6.00
Mistral Small Mistral $0.15 / $0.60 $0.00063 $6.30
DeepSeek V4.1 Flash (peak) DeepSeek $0.30 / $1.20 $0.00120 $12.00
Gemini 3.1 Flash-Lite Google $0.25 / $1.50 $0.00125 $12.50
GPT-5 mini OpenAI $0.25 / $2.00 $0.00150 $15.00
GPT-4.1 mini OpenAI $0.40 / $1.60 $0.00160 $16.00
Mistral Large 3 Mistral $0.50 / $1.50 $0.00184 $18.38
Gemini 3.5 Flash-Lite Google $0.30 / $2.50 $0.00185 $18.50

For comparison, the same 10,000 requests cost $90 on GPT-6 Sol, $124 on Claude Sonnet 5.5, $248 on Claude Opus 5.5 and $450 on GPT-6 Astra. The cheapest model in the table costs less than 1% of what GPT-6 Astra does.

Which cheap model to pick

Which cheap model to pick: GPT-6 Luna is the one to try first. It is the newest small model from OpenAI (September 2026), and only $1.50

When the cheapest model costs more

A cheap model that fails costs you twice: once for the bad answer, and again for the retry on a better model, plus your time. Two patterns save more than picking the cheapest model everywhere:

  1. Route by difficulty. Send everything to a small model first and escalate only the requests it gets wrong (or that you flag as hard) to GPT-6 Sol, Claude Sonnet or Gemini 3.1 Pro. If 80% of requests stay on GPT-6 Luna and 20% go to GPT-6 Sol, the average cost drops by about 75% compared with sending all of them to Sol.
  2. Cut the tokens, not just the price. Prompt caching bills repeated instructions and documents at about a tenth of the input price on most providers, and batch APIs charge about half for work that can wait. See Prompt caching explained and Batch APIs: half-price AI.

Your language changes the ranking

All the costs above are for English. Other languages use more tokens for the same meaning: Korean about 44% more and Japanese about 79% more on GPT, so every price in the table rises by the same share. The ranking stays the same; the bill does not. See the same prompt in 41 languages.

Check your own prompt

Paste a real prompt into the token counter to see its exact token count and cost on every model, or open the OpenAI, Claude or Gemini counter. For a full side-by-side of capability against price, see AI model capability vs price.

Prices change often. Check the provider's pricing page before you commit to a model.

Sources

More articles

Disclosure: TokenSave gets no funding, credits or sponsorship from AI companies. All measurements use our own paid subscriptions.

Paste a prompt to see its tokens and cost on every GPT-6 model, Claude and Gemini.

Open the token counter →