The Cheapest LLM APIs in 2026 (Ranked by Real Cost)

Updated November 4, 2026

If cost is your top priority, small models from several providers now cost cents per million tokens. Here is how they compare.

Cheapest models (input / output per million tokens)

What they cost in practice

For 45 million input and 12 million output tokens a month:

At this scale the differences are a few dollars. Quality, latency and context length matter more than the last cent.

Cheapest is not always best

Small models are strong at classification, extraction, routing and short summaries. They struggle with long reasoning and complex code. A common pattern is to send easy requests to a cheap model and only escalate hard ones, covered in 7 ways to cut your LLM bill.

Why the bottom of the market keeps falling

Small models are getting cheaper every generation. Claude Haiku 4.5 cost $1 / $5, and Haiku 5.5 now costs $0.10 / $0.50 for prompts up to 100,000 tokens. OpenAI cut GPT-6 Luna to half the price of its predecessor. Competition from open-weight models, which anyone can host, keeps pressure on prices.

What small models do well

• Classifying support tickets, emails and reviews
• Extracting fields from invoices, forms or documents
• Routing requests to the right tool or team
• Short summaries and rewrites
• Tagging and moderation at high volume

Where they fall short

Long multi-step reasoning, hard coding problems, subtle writing and tasks that need a lot of world knowledge are still better handled by larger models. If a small model gets the job right only 70% of the time, retries and human review can cost more than a mid-tier model.

A routing example

Our standard workload is 45 million input and 12 million output tokens a month. Sent entirely to Claude Sonnet 5.5, it costs about $210. Now route 80% of the traffic to Claude Haiku 5.5 and keep 20% on Sonnet for the hard cases:

The savings depend on how reliably you can tell easy requests from hard ones, so test the router on real data.

What to check before picking a budget model

• Context and output limits: Amazon Nova Micro, for example, has a 128,000 token context window and a modest maximum output.
• Features: vision, tool calling and structured output are not available on every small model.
• Availability: providers retire older models. Some sources list parts of the Gemini 2.5 family as deprecated, so confirm a model is still offered before building on it.
• Rate limits and region: cheap models can have tighter quotas, and data residency may matter.
• Self-hosting: open-weight models such as Qwen and Ministral can be run yourself, but GPUs cost money. At low and medium volume, the API is usually cheaper.

Frequently asked questions

What is the cheapest LLM API?

By list price, Amazon Nova Micro at $0.035 input and $0.14 output per million tokens, followed closely by Cohere Command R7B.

Is a cheap model good enough for a chatbot?

For simple FAQ-style bots, often yes. For open-ended assistants, test first. Escalating hard questions to a stronger model is a common compromise.

Are free tiers worth it?

They are useful for prototyping, but check the terms, since free usage can be limited and may allow providers to use your data.

Compare them yourself

The free calculator lists all of these models with editable prices, plus a filter to compare one provider at a time.

List prices checked November 4, 2026.

Check the source: official pricing pages: Anthropic · OpenAI · Google Gemini · xAI · Mistral · DeepSeek · Amazon Bedrock · Cohere. See our methodology.