How do I compare AI model API prices fairly?

Compare input and output prices separately, because output tokens are billed 2x to 6x more than input on most providers. A model with a low input price can still be expensive if your app generates long answers. The blended column on this page weights input and output 3-to-1 as a neutral baseline, and the monthly-cost column uses your own token volumes so the ranking reflects the bill you would actually pay, not a single sticker number.

Which provider has the cheapest AI models in 2026?

Hosted open-weight models — DeepSeek, Llama 4, Qwen, Mixtral and their resellers such as Together, Fireworks, DeepInfra and Groq — are the cheapest frontier-class options, often under $1 per million tokens. Premium closed models like GPT-5.5 and Claude Opus sit at the top of the range. Filter by category above to see budget, mid and flagship tiers separately, or enter your token mix to rank every model by your real cost.

Compare 300+ AI Model API Prices (2026) — Input/Output Cost

Why a fair AI price comparison needs two numbers

Almost every "AI model pricing" table you'll find online sorts by input price, because it's the smaller, friendlier number to print. But real applications are billed on both directions, and output tokens are routinely priced two to six times higher than input. A retrieval or summarisation app that reads long documents and writes short answers is input-heavy; a chatbot, agent or code generator that produces long completions is output-heavy. The same two models can swap places in the ranking purely based on which way your traffic leans. That's why this table shows input and output separately, adds a blended 3:1 baseline (a common real-world ratio of three input tokens per output token) for quick apples-to-apples sorting, and lets you enter your own volumes to compute the bill you'd actually pay.

How to read the table

Type in the search box to jump to a model or provider, or click the provider and category chips to narrow the field — flagship, mid, budget, reasoning, embedding, vision or open-weight. Click any column header to sort ascending, click again to reverse. The Your $/mo column recalculates instantly from the input and output volumes at the top, so you can size a real workload against every model at once. Context is shown in thousands of tokens (K) — a 1000K context window means one million tokens.

Frontier quality is getting cheap

The gap between premium and budget models has narrowed dramatically. In 2026 you can run a frontier-class open-weight model — Llama 4, Qwen 3, DeepSeek V4 — for a fraction of what GPT-5.5 or Claude Opus cost, and for many production tasks (classification, extraction, routine chat) the quality difference is invisible to end users. Hosted resellers such as Together, Fireworks, DeepInfra, Groq and Novita compete on the same weights, so the same model can appear several times at slightly different prices. The right move is rarely "use the best model everywhere" — it's to route by difficulty, letting a cheap model handle the bulk while a frontier model is reserved for the hard minority of calls.

Three levers below the model price

Once you've picked a model, three mechanisms cut the bill further. Prompt caching bills repeated system prompts and context at a fraction of the normal input rate — size it on the prompt caching savings calculator. Batch APIs give roughly 50% off for jobs that can wait — see the batch API savings calculator. And routing routine calls to a smaller model is often the biggest lever of all — model it on the AI model router cost calculator.

Related tools & guides

Cheapest LLM API finder · LLM token cost calculator · OpenAI vs Claude vs Gemini pricing · Blended LLM price calculator · AI model router cost calculator · All AI APIs