The same model can cost different amounts because each host sets its own price per token, and the tokens a model burns on the same job differ by model family and by reasoning behavior. TaskCents tracks 1,283 host prices across 326 models on OpenRouter, so many models appear at several hosts. Compare the host, the list price and the tokens used per task, not just the headline rate.
Why one model can have several prices
A model's price is set by whoever serves it. The maker may sell it directly, and other companies may host the same model at their own rates. The number of hosts varies a lot. DeepSeek V4.1 Flash is served by 32 hosts and GLM 5.3 Flash by 33. Claude Sonnet 5.5 is served by 8 endpoints (Anthropic, Amazon Bedrock, Azure, Google).
More hosts means more chances for prices, terms and speeds to differ. When a comparison page shows one price for a model, check which host it refers to. Each model page has a provider table showing the host-level prices.
Tokens per task matter as much as price per token
A low price per million tokens does not guarantee a low bill. Reasoning tokens are billed as output tokens, so a model that thinks at length costs more even at a low rate.
My chat answer data shows the spread. GPT-6 Luna uses 639 reasoning tokens and 540 output tokens per answer, and lists at $0.10 input and $0.50 output per 1M tokens. Claude Sonnet 5.5 uses 0 reasoning tokens and 1,256 output tokens, and lists at $2.00 and $10.00. The result is 0.06¢ per task for GPT-6 Luna and 1.3¢ for Claude Sonnet 5.5, a gap of more than 20 times.
| Model | List price (input / output per 1M) | Chat answer cost per task | Basis |
|---|---|---|---|
| GPT-6 Luna (OpenAI) | $0.10 / $0.50 | 0.06¢ | measured |
| Claude Haiku 5.5 (Anthropic) | $0.10 / $0.50 | 0.07¢ | measured |
| Qwen3.8 Flash (Alibaba Qwen) | $0.15 / $0.47 | 0.08¢ | measured |
| Claude Sonnet 5.5 (Anthropic) | $2.00 / $10.00 | 1.3¢ | measured |
| Claude Opus 5.5 (Anthropic) | $4.00 / $20.00 | 2.5¢ | estimated |
The median across all 326 models is 0.24¢ per chat answer. Claude Fable 5.1 is the most expensive in my chat data at 6.3¢, though that figure is estimated.
Reasoning can also run the other way. DeepSeek V4.1 Flash lists at $0.30 input and $1.20 output, yet it uses 5,461 reasoning tokens on the chat task and lands at 0.76¢, more than GPT-6 Luna's 0.06¢.
Tokenizers change the input count
For the same English document, Anthropic's tokenizer produces 1.39 times as many tokens as OpenAI's (OpenAI = 1.00). Google is at 0.98 and MiniMax at 0.99. An identical prompt can therefore cost more on one family's models before any output is counted.
My figures come from one run per task at default settings, through OpenRouter, with output capped at 16,000 tokens. Some rows are estimated rather than measured. Treat estimates as a guide. See the methodology for details, and the chat answer task page for the full ranking.
Context limits are not usable context
The context window is the maximum input a model accepts, and it is separate from price. Listed limits in my data run from 192,000 tokens for Command A+ to 1,050,000 for several OpenAI models. Grok 4.7 lists 500,000 and Nemotron 3 Ultra 262,144.
Two points matter:
- A large window does not mean your task needs it. My test inputs are a 10,023-word PDF (J. S. Mill's On Liberty), a 497-line code file (CPython textwrap.py) and a 1,017-word translation of The Prince to Spanish. None need the largest windows.
- Long inputs are billed. A model with a high input rate can cost more on a long document than one with a lower rate, whatever the window size.
The agent scenario shows this. It starts with an 8,000-token system prompt, runs 50 steps, adds 1,500 tokens of context per step and returns 300 visible output tokens per step. Input grows every step, so the input rate, and the cached input rate if your prompts repeat, drives much of the total.
Quantization and uptime
Hosts may serve the same model differently. Some may run a compressed (quantized) version to save compute, and uptime and speed can vary between hosts. My my data does not include quantization details or uptime figures, so I give no numbers for either. Check each host's documentation and test with your own prompts before committing.
How to compare providers
- Find every host that serves the model. DeepSeek V4.1 Flash, for example, has 32 to check.
- Compare input, output and cached input prices.
- Measure tokens per task, including reasoning tokens.
- Check the context window only if your inputs are long.
- Run your own test set before switching.
To estimate your own bill, use the cost calculator. To see how a model performs on a task, start with its page, such as Claude Sonnet 5.5 or GPT-6 Luna. Consumer AI subscriptions are compared on the plans page.
FAQ
Why does the same model show different prices on different sites?
Each host sets its own rates and terms, and may serve the model differently. Check the host name next to any price. Prices here are list prices as of 2026-10-11.
Is a cheaper price per token always the cheaper option?
No. Tokens used per task, including reasoning tokens billed as output, can outweigh a lower rate. DeepSeek V4.1 Flash has a $1.20 output price but costs 0.76¢ per chat answer because it uses 5,461 reasoning tokens, against 0.06¢ for GPT-6 Luna.
Which numbers here are measured and which are estimated?
The chat answer table labels each row. Measured models were run once per task at default settings through OpenRouter. Estimated rows were not directly run, so test the models you care about yourself.