How much does an LLM cost: 2026 token pricing guide
Real 2026 prices for Claude, GPT-5, Gemini, and DeepSeek per million tokens, how an API bill is calculated, and how to save up to 90% by choosing the model well.
Direct answer: In 2026, prices per million tokens range from 0.15 USD (GPT-4o mini) to 5 USD input and 25 USD output (Claude Opus 4.5). The monthly bill is calculated by multiplying your input and output tokens by those rates; for a typical support agent (60M input tokens per month) the cost ranges between 18 and 300 USD depending on the model chosen.
How an AI API bill is calculated
Providers charge two rates per million tokens: one for input (what you send: prompt, documents, history) and a higher one for output (what the model generates). One token equals about three quarters of a word in English. The formula: (input tokens / 1,000,000) × input price + (output tokens / 1,000,000) × output price.
The typical mistake is underestimating input: in RAG or agent applications, every request drags documents and context along, so input usually weighs 4-10 times more than output. That's why our calculator separates both figures and projects the real monthly cost.
Real prices per million tokens (2026)
High end: Claude Opus 4.5 (5/25 USD input/output), Claude Sonnet 4.5 (3/15), GPT-5 (1.25/10), Gemini 2.5 Pro (1.25/10), o3 (2/8). Mid-range and budget: Claude Haiku 4.5 (1/5), Gemini 2.5 Flash (0.30/2.50), GPT-5 mini (0.25/2), DeepSeek V3 (0.27/1.10), and GPT-4o mini (0.15/0.60).
The golden rule of saving: use the big model for design and hard cases, and the mini for 80% of volume (classifying, summarizing, formatting). That single change cuts bills by 60 to 90% with no perceptible quality loss on routine tasks.
- Claude Opus 4.5: 5 / 25 USD — maximum reasoning
- Claude Sonnet 4.5: 3 / 15 USD — professional use
- GPT-5: 1.25 / 10 USD — balanced high end
- Gemini 2.5 Flash: 0.30 / 2.50 USD — volume
- GPT-4o mini: 0.15 / 0.60 USD — minimum cost
7 levers to pay less (without losing quality)
First, complexity-based routing: mini model by default, scaling to the big one only on failure. Second, prompt caching: Anthropic and OpenAI discount repeated tokens by up to 90%. Third, trimming history: don't resend full conversations. Fourth, explicitly requested short answers. Fifth, document compression in RAG. Sixth, batch API for non-urgent tasks (50% discount at OpenAI). Seventh, monitoring from day one: no measurement, no optimization.
Before launching any agent or product, simulate the cost with real usage data. Our Token Cost Calculator includes this guide's exact prices, compares the 13 main models, and tells you how much you'd overspend by choosing wrong.
Frequently asked questions
What exactly is a token?
The smallest unit of text the model processes: whole words or fragments. In English, 1000 tokens equal roughly 750 words. API prices are always charged per million tokens.
Is a subscription (20 USD/month) or the API better?
For personal and professional interactive use, the subscription almost always comes out better: 20 USD equals millions of tokens per month. The API suits automated products (agents, apps) where each request runs with no human in front.
Do LLM prices go down or up?
Historically they fall: the cost per unit of intelligence drops an order of magnitude every ~18 months. High-end models rise at launch and their mini equivalents undercut them months later. Review the calculator every quarter.