Show us a process where humans are doing work software should be doing

What does the AI API really cost per month?

Official prices for OpenAI, Anthropic, Google and DeepSeek, applied to your volume, your document sizes, and your language. No email required.

documents processed, conversations, summaries, extractions

prompt plus the document or conversation the model reads

what the model writes back

ModelInput $/MOutput $/MMonthly cost
Gemini 2.5 Flash-LiteGoogle · Budgetcheapest$0.10$0.40$2.24/mo
DeepSeek v4 FlashDeepSeek · Budget$0.14$0.28$2.24/mo
GPT-5.4 nanoOpenAI · Budget$0.20$1.25$5.92/mo
DeepSeek v4 ProDeepSeek · Mid tier$0.435$0.87$6.96/mo
Gemini 3.1 Flash-LiteGoogle · Budget$0.25$1.50$7.20/mo
GPT-5.4 miniOpenAI · Budget$0.75$4.50$22/mo
Claude Haiku 4.5Anthropic · Budget$1.00$5.00$26/mo
Gemini 3.6 FlashGoogle · Mid tier$1.50$7.50$38/mo
Gemini 3.1 ProGoogle · Flagship

Up to 200K context

$2.00$12.00$58/mo
Claude Sonnet 5Anthropic · Mid tier$2.00$10.00$67/mo
GPT-5.4OpenAI · Mid tier$2.50$15.00$72/mo
GPT-5.5OpenAI · Flagship$5.00$30.00$144/mo
Claude Opus 5Anthropic · Mid tier$5.00$25.00$166/mo
Claude Fable 5Anthropic · Flagship$10.00$50.00$333/mo

Official standard-tier rates, each read from the provider's own pricing page on the date shown: OpenAI 2026-07-08, Anthropic 2026-08-24, Google 2026-08-03, DeepSeek 2026-07-08. Claude Fable 5, Opus 5 and Sonnet 5 use a newer tokenizer that produces roughly 30% more tokens for the same text (per Anthropic's pricing page); the estimate accounts for it.

Where these numbers come from

Every price here was read from the provider's own official page, standard tier, no batch discounts: OpenAI (developers.openai.com), Anthropic (platform.claude.com), Google (ai.google.dev) and DeepSeek (api-docs.deepseek.com). Providers are re-checked on different dates rather than all at once, so each one carries the date its page was last read: OpenAI 2026-07-08, Anthropic 2026-08-24, Google 2026-08-03, DeepSeek 2026-07-08. Nothing on this table is older than 2026-07-08. NVIDIA Nemotron is excluded because build.nvidia.com publishes no per-token price: its hosted catalog is a free prototyping tier and production runs self-hosted under NVIDIA AI Enterprise.

Word-to-token conversion uses ~1.33 tokens per English word (the 4-characters heuristic both OpenAI and Anthropic publish) and ~1.6 for Spanish, based on tokenization fairness research. Prices change: always confirm with the provider before budgeting. And if what you actually need is the cost of implementing AI end to end, that is a different number: see the full implementation pricing guide, the AI agent cost breakdown, and how to budget the year-one project.

Common questions

How much does the AI (LLM) API cost per month for a company?

It depends on three numbers: tasks per month, words the model reads per task, and words it writes. As a reference, 10,000 monthly tasks reading 600 words and writing 200 in Spanish cost from a few dollars per month on budget models (DeepSeek v4 Flash, Gemini Flash-Lite) to a few hundred on flagships. The calculator above computes it with official prices.

Why does Spanish cost more tokens than English?

Tokenizers were trained mostly on English, so Spanish splits into more pieces: roughly 1.5 to 1.7 tokens per word versus about 1.33 in English (Petrov et al., NeurIPS 2023). The same document in Spanish can cost 15 to 25 percent more.

Is the API the main cost of implementing AI?

Almost never. For a production agent, model tokens are usually a small fraction of the total: the real costs are integration engineering, evaluation, and day-two operation. Our full pricing guide breaks down what an implementation actually costs.

Which AI model is cheapest in 2026?

Among official APIs, DeepSeek v4 Flash ($0.14 in / $0.28 out per million tokens) and Gemini 2.5 Flash-Lite ($0.10 / $0.40) are the cheapest published tiers as of August 2026. Cheapest is not always best: accuracy on your real cases decides the true cost.

Turn the number into a decision

The ranges on this page are the market. A 10-day validation on one of your workflows turns them into a fixed-scope quote — or tells you to wait.