LLM API pricing comparison 2026: Claude vs GPT vs Grok vs DeepSeek

Every LLM API is priced per token, but the numbers vary so widely that comparing them by eye is hard. The most expensive model in this comparison charges more than a hundred times as much per output token as the cheapest. Some models are discounted heavily through gateways and some aren't discounted at all. And the figure that matters, what your workload costs each month, depends on the mix of input and output you send.

All prices are per million tokens at the time of writing, taken from the GreenTokens catalogue, which lists each model's official price next to ours. Prices change often, so check the models page for current figures before you decide.

How LLM API pricing works

• Input tokens are everything you send: system prompt, conversation history, documents and tool results.
• Output tokens are what the model writes back. They're usually five times the input price or more, which is why output-heavy work gets expensive.
• Cached input is repeated context the provider has already processed, billed at a small fraction of the input price. See what prompt caching is.

Flagship models

The most capable, most expensive tier. Official list prices first, then GreenTokens prices:

Mid-tier models

The workhorses most products run on: strong enough for most tasks, at a fraction of the flagship price.

Budget models

The cheapest LLM API options, for high-volume and simpler tasks:

Budget models are mostly sold at their list price, because there's little room left to discount. Claude Haiku 4.5 is the exception, at 50% off.

What a real workload costs

Per-token prices are easier to judge as a monthly bill. Take a typical product workload: 100,000 requests a month, each with 2,000 input tokens and 500 output tokens. Here's what it costs on each model:

The range is striking: the same workload costs anywhere from under $30 to $4,500 a month depending on the model. Two things drive most of that spread. The first is the model tier, which moves the bill by one or two orders of magnitude. The second is the discount, which halves it or better on the flagship and mid-tier models.

Cached input changes the maths

If your prompts repeat, such as a long system prompt, a shared document or an agent's growing context, cached input pricing can matter more than the headline rate. Cache reads at GreenTokens prices:

For an agent that resends 50,000 tokens of context on every turn, reading that context from cache instead of paying full input price is the difference between a manageable bill and an alarming one.

Which is the cheapest LLM API?

By raw token price, GLM-5.3-Flash, GPT-6 Luna and DeepSeek V4 Flash are the cheapest LLM APIs in this comparison. But the cheapest API for you is the lowest-priced model that handles your task well, bought at the best price available. In practice that usually means:

1. Test a budget model first on your real inputs; many tasks don't need more.
2. Move up to a mid-tier model only where the budget one falls short.
3. Keep flagship models for the hardest requests, and pay a discounted price for them.
4. Structure prompts so the repeated parts are cached.

We haven't compared quality here; the right model depends on your task, and the best test is your own data. For deeper dives by family, see the cheap Claude API and cheapest OpenAI API model guides. What the numbers do show is that the model tier and the price you pay for it are the two biggest levers on an LLM bill, and both are easy to change when every model sits behind one API.

Common questions

How often do LLM API prices change?

Often. New models usually launch at a lower price than the ones they replace, and older models are sometimes repriced or retired. Treat any comparison, including this one, as a snapshot, and check live prices before committing a large workload.

Why is output so much more expensive than input?

The model reads your whole input in one pass, but generates output one token at a time, which takes far more computation per token. That's why trimming output, and caching repeated input, are the two quickest ways to lower a bill.

Compare them yourself with one key

GreenTokens gives you every model in this comparison through one OpenAI- and Anthropic-compatible API, with a prepaid balance and no subscription. Switching from Claude Sonnet 5 to GPT-6 Sol or GLM-5.3 to compare them is a one-word change in your request, and the models page lists every model's official price beside ours.