LLM API pricing

How to read this table

Input is what you pay for the tokens you send: the system prompt, the conversation so far, retrieved documents and the user's message. Output is what you pay for the tokens the model writes back. Cached input is the discounted rate for prompt text the vendor has already processed and stored, such as a long system prompt reused across requests. All three are list prices in US dollars per 1 million tokens.

Per-token prices are hard to compare in your head, so the last column prices one ordinary request: 1,000 input tokens and 500 output tokens, roughly a short chat turn with some context. The table is sorted by that figure. It ranges from $0.00025 for GPT-5 nano to $0.45 for o1-pro. If your workload is different, for example long documents in and short answers out, the ranking can change, so look at the input and output columns separately.

The price you actually pay can be lower or higher than the list price. Batch processing, caching, peak hours, long prompts and launch promotions all change it, and each vendor page explains its own rules.

Why LLM API prices differ so much

The 62 models above are all priced per token, yet a request can cost 1,800 times more on one model than on another. These are the main reasons.

Model size and tier

Every vendor sells a ladder of models, from small and fast to large and slow. Bigger models need more GPU memory and compute for every token, and the price follows. Inside a single OpenAI generation the gap is wide: GPT-5.4 nano costs $0.20 per 1M input tokens and GPT-5.4 Pro costs $30.00, 150x more. Names like nano, mini, Flash-Lite, Flash, Haiku, Sonnet and Opus mark these tiers.

Output costs more than input

A model reads your whole prompt in one parallel pass, but writes its answer one token at a time. That makes output tokens more expensive to serve. Across this table, output is priced at 2x to 8.3x the input rate, with a median of 5x. Workloads that generate long answers, code or reports are therefore dominated by the output price.

Reasoning and "Pro" models

Reasoning models think before they answer, and those thinking tokens are billed as output even when you never see them. Pro variants spend even more compute per request. That is why the most expensive rows here are reasoning and Pro models, and why a cheap-looking reasoning model can cost more per task than its per-token price suggests.

Caching discounts

Reusing the same prompt prefix lets the vendor skip work, and most pass part of the saving on. Cached input costs between 2% and 50% of the normal input price across the models that offer it. 8 models, mostly Pro variants, list no cached rate at all. Anthropic also charges extra to write to the cache, so caching only pays off when the same prefix is reused often.

Long prompts cost more on some models

15 of the 62 models switch to higher rates once a prompt passes a threshold, 200K or 272K tokens. Above it, the input price doubles and the output price rises 1.5x to 2x. Anthropic's Claude 4.6 and later models bill their full 1M-token context at standard rates instead.

Tokenizers are not the same

A token is not a fixed unit. Each vendor splits text with its own tokenizer, so the same paragraph becomes a different number of tokens on different models. Anthropic says its tokenizer for Claude 4.7 and later produces about 30% more tokens for the same text. Two models with the same per-token price can still produce different bills for identical prompts.

Time of day and promotional pricing

DeepSeek charges double during peak hours: DeepSeek V4 Pro is $0.66 input off-peak and $1.32 at peak. OpenAI's GPT-5.6 Sol costs $4.00 / $20.00 on a promotion that runs at least through November 21, 2026, and Google sells 3 Gemini Flash models at launch prices through December 31, 2026; Gemini 3.8 Flash goes from $0.75 to $1.50 per 1M input tokens in 2027. Prices in this table are what the vendors charge today, not what they will charge next year.

Newer is not always pricier

Vendors often launch new models at lower prices than the ones they replace. OpenAI's o1 still lists $15.00 input and $60.00 output, while the newer GPT-5.4 costs $2.50 and $15.00. Staying on an old model out of habit can be the expensive choice.

Resellers and open-weight models

Open-weight models can be hosted by anyone, so the same model shows up at different prices on different platforms. For DeepSeek V4 Pro, DeepSeek charges $0.66 input and $1.98 output off-peak, while OpenRouter lists $0.58 and $1.74. Cloud platforms and routers also add their own markups, discounts and regional pricing. This page uses the vendor's own price, because that is the reference the others are measured against.