Cheapest LLM APIs in 2026: Every Model Ranked by Real Cost

If you are paying frontier-model prices for every API call, you are almost certainly overpaying. In 2026 a wave of budget models does a genuinely good job on routine work — chat, classification, extraction, summarization — for a tiny fraction of the cost. Here is what the cheapest LLM APIs actually cost, and when they are the right call.

How to read LLM pricing

Every API bills two things per million tokens: the tokens you send (input) and the tokens the model generates (output). Output usually costs several times more than input, so the cheapest model for you depends on whether your workload is input-heavy (long prompts, retrieved documents) or output-heavy (long generated answers). A token is roughly four characters, or about three-quarters of a word. You can run your own workload through our LLM cost calculator.

The cheapest LLM APIs, ranked

Standard published rates as of August 2026, per 1M input / output tokens:

• Mistral Small 4 — $0.15 / $0.60. The cheapest capable hosted option from Mistral, with even cheaper Ministral tiers for high-volume, simple tasks.
• DeepSeek V4 Flash — $0.22 / $0.66. Strong general reasoning at rock-bottom prices, with off-peak discounts that go lower still.
• GPT-5.6 Luna — $0.20 / $1.20. OpenAI's cheapest tier and now the default free model in ChatGPT, after an 80% price cut.
• DeepSeek V4 Pro — $0.66 / $1.98. A step up in quality while staying far below Western frontier pricing.
• Gemini 3.7 Flash — $0.75 / $3.75. Google's price-performance workhorse, with a generous free tier for testing.
• Mistral Large 3 — $0.50 / $1.50. Symmetric, cheap output makes it attractive for generation-heavy jobs.

For comparison, frontier models run far higher — Claude Opus 5 is $5 / $25 and GPT-5 is $1.25 / $10 — which is 10–30× the budget tier on output.

When cheap is good enough

Budget models comfortably handle the majority of production traffic: FAQ and support chat, routing and classification, tagging, data extraction, first-draft writing, and summarization. If a task has a clear right answer or a narrow scope, a cheap model usually nails it.

Where you still want a frontier model: multi-step reasoning, hard coding, long-context synthesis, and anything where a subtle mistake is expensive. The smart pattern is to route the easy majority to a cheap model and escalate only the hard requests — a tiered setup routinely cuts total spend by more than half.

Two ways to save without downgrading

Before you switch models, try trimming input. Long system prompts, chat history, and retrieved context inflate every call; compressing them can cut a bill by a third with no change in model. And use batch and cached-input pricing where available — most providers offer roughly 50% off for asynchronous batch jobs and about 90% off for cached input reads.

Run your exact numbers on the LLM cost calculator, and if you want the interactive picker, describe your use case to the AI Tool Advisor.

Prices are standard, non-cached per-million-token rates as of August 2026 and change frequently — confirm current rates with each provider before budgeting.