LLM API Pricing Explained: What You’ll Actually Pay (October 2026)

Prices checked October 23, 2026, from each provider’s official pricing page (see Sources). Prices change often; confirm them before you commit to a budget.

LLM API pricing confuses almost everyone at first — per-token billing, separate input and output rates, subscriptions versus pay-as-you-go, and prices that range from cents to dollars for seemingly similar models. This guide demystifies all of it, shows you real 2026 prices, and gives you a simple way to estimate and slash your bill before it surprises you.

How per-token billing works

Nearly every LLM API charges by the token, not by the request or the hour. Get this one concept and the rest follows. A token is a piece of text — roughly ¾ of a word — and you’re billed for every token in your prompt and every token the model generates back. Providers quote prices per million tokens, which sounds large but adds up fast in a busy app.

Pricing, step by step

Real prices, October 2026 (per million tokens)

A snapshot of the range across the three biggest providers plus a budget option. Note how far apart the tiers are, and how output costs several times more than input:

The headline: the most capable models can cost about 100× more per output token than the smallest ones ($50 vs $0.50 per million). That’s why model choice is the single biggest driver of your bill, and why paying top-tier prices for simple tasks is the most common way teams overspend. To see what your own workload would cost on 46 models, use our free AI API cost calculator.

Estimate your bill with one figure

The quickest way to estimate a bill is to multiply your expected monthly requests by the average input and output tokens per request, then by the model’s per-million rates. Our cost calculator does this for you, including caching and batch discounts. For agent-style workloads, see how much it costs to run an AI agent.

Hidden costs and pricing gotchas

The per-token rate is only the headline. Several real costs hide behind it, and missing them is how a “cheap” model ends up expensive:

• Reasoning tokens. Some advanced models generate hidden “thinking” tokens before their answer — and you pay for those too. A model that reasons more can cost more per request than its rate suggests.
• Context re-sending. In multi-turn conversations or agents, the entire history is usually re-sent on every call, so input costs grow with conversation length. This is why long agent runs get expensive fast.
• Retries and failures. Failed or retried calls still consume tokens. Build in error handling so a loop doesn’t quietly run up a bill.
• Tiered and committed pricing. Some providers offer volume discounts or cheaper committed-use rates — worth it at scale, but they lock you in, so model your real volume first.
• Fine-tuning surcharges. A fine-tuned model often costs more per token to run than the base model, on top of the one-time training cost.

Subscription or API? A worked example

People often ask whether to just pay for a $20/month subscription or use the API. The answer depends entirely on who’s consuming. If you personally are chatting with the model a few dozen times a day, the flat subscription is almost always cheaper and simpler — you’d struggle to burn $20 of API tokens at human typing speed. But if you’re building an app that makes thousands of calls on behalf of many users, there’s no subscription that covers it; you need the API, and your cost scales with usage. A useful rule: subscriptions price a person, APIs price a workload. Many builders use both — a personal subscription for their own work and the API for what they ship. Once you know which side of that line you’re on, the pricing model practically chooses itself, and you can focus on the levers that actually move your bill: model choice and token discipline.

How to cut costs (a checklist)

• Right-size the model. Don’t pay frontier prices for classification or simple generation — a budget model handles it.
• Trim tokens. Shorter prompts and capped output lengths directly lower the bill; output is the expensive side.
• Cache repeated context. Many providers discount cached input heavily — huge for apps that resend the same system prompt.
• Batch non-urgent work. Anthropic, OpenAI and Google all offer batch processing at about half the standard price.
• Route by complexity. Cheap models for the bulk, flagships only for the hard 10%. (See Claude vs ChatGPT vs Gemini.)
• Consider self-hosting open-weight models at very high volume.

Frequently asked questions

Sources

1. Anthropic: API pricing
2. OpenAI: API pricing
3. Google AI for Developers: Gemini API pricing
4. DeepSeek: API pricing