Data note
LLM API Prices per Million Tokens: 10 Models Side by Side
Ten frontier and budget models priced per million tokens, pulled from OpenRouter's public model API on 2026-10-25, with the output-to-input ratio each one charges.
On OpenRouter's published rates from 2026-10-25, Claude Sonnet 4.5 costs $3.00 per million input tokens and $15.00 per million output, GPT-5 costs $1.25 and $10.00, Gemini 2.5 Flash costs $0.30 and $2.50, and Llama 3.3 70B costs $0.10 and $0.32. Output tokens cost 3x to 8x more than input tokens on every model in this set.
- GPT-5: $1.25 per million input tokens, $10.00 per million output, 400,000 token context (checked 2026-10-25 via OpenRouter's model API).
- Claude Sonnet 4.5: $3.00 input, $15.00 output, 1,000,000 token context (checked 2026-10-25).
- Gemini 2.5 Flash: $0.30 input, $2.50 output, 1,048,576 token context (checked 2026-10-25).
- DeepSeek Chat: $0.257 input, $1.029 output, 163,840 token context (checked 2026-10-25).
- Llama 3.3 70B on OpenRouter: $0.10 input, $0.32 output, 131,072 token context (checked 2026-10-25).
- OpenRouter's model API listed 440 models with a positive price on 2026-10-25.
1 primary source · If this is out of date, tell us and we will re-measure.
On this page
Model prices change faster than any other number on this site, which is exactly why a dated table beats a blog post from last spring.
These rates were pulled from OpenRouter's public model API on 2026-10-25. The API returns per-token prices for every model it routes; the table below converts them to dollars per million tokens, which is the unit people actually reason in.
Ten models, priced per million tokens
| Model | Input / 1M | Output / 1M | Output ÷ input | Context |
|---|---|---|---|---|
| Llama 3.3 70B | $0.10 | $0.32 | 3.2x | 131,072 |
| GPT-4o mini | $0.15 | $0.60 | 4.0x | 128,000 |
| DeepSeek Chat | $0.257 | $1.029 | 4.0x | 163,840 |
| GPT-5 mini | $0.25 | $2.00 | 8.0x | 400,000 |
| Gemini 2.5 Flash | $0.30 | $2.50 | 8.3x | 1,048,576 |
| Claude Haiku 4.5 | $1.00 | $5.00 | 5.0x | 200,000 |
| GPT-5 | $1.25 | $10.00 | 8.0x | 400,000 |
| Gemini 2.5 Pro | $1.25 | $10.00 | 8.0x | 1,048,576 |
| Mistral Large | $2.00 | $6.00 | 3.0x | 128,000 |
| Claude Sonnet 4.5 | $3.00 | $15.00 | 5.0x | 1,000,000 |
What a real workload costs
Prices per million tokens are hard to feel. A support assistant that handles 20,000 conversations a month, each with 2,000 input tokens of context and 400 tokens of answer, moves 40 million input and 8 million output tokens.
| Model | Monthly cost for that workload |
|---|---|
| Llama 3.3 70B | $6.56 |
| Gemini 2.5 Flash | $32.00 |
| GPT-5 mini | $26.00 |
| GPT-5 | $130.00 |
| Claude Sonnet 4.5 | $240.00 |
The spread between the cheapest and most expensive option is 36x for identical token counts. That is the whole decision: on most chat products, the model that is 5% better does not justify being 10x more expensive, and on code or long-document work the cheap model fails often enough that retries erase the saving.
How we measured it
We requested OpenRouter's /api/v1/models endpoint on 2026-10-25, parsed the pricing.prompt and pricing.completion fields as dollars per token, and multiplied by one million. Models with a zero or missing price were excluded, leaving 440 priced models. The ten in the table were chosen to cover one model per vendor at both a frontier and a budget tier; nothing was measured by running prompts.
What changes in practice
The output-to-input ratio varies more than the headline price. Mistral Large charges 3x for output; GPT-5 and Gemini charge 8x. A workload that reads long documents and writes short answers is far cheaper than its input price suggests, while a workload that generates long reports pays almost entirely at the output rate. Comparing models by input price alone gets the ranking wrong for exactly those cases.
Caching and batch discounts can move a bill by half again, and neither is visible in a per-token table — you have to read each provider's own pricing page for those. The numbers here are the list rate before any of that.
Bottom line
For high-volume chat where occasional wrong answers are survivable, the $0.10–$0.30 input tier is 8x to 12x cheaper than the frontier tier and good enough. For anything that has to be right the first time — code, extraction, legal text — the $1.25–$3.00 tier earns its price back in retries avoided. The output rate, not the input rate, is what to model your bill on.
Frequently asked questions
Why is output more expensive than input?
Output tokens are generated one at a time by the model and dominate GPU time; input tokens are processed in parallel. On the models listed here output costs between 3x and 8x the input rate, so a workload that generates long answers pays mostly for output.
Are these the vendors' own list prices?
They are the per-token rates OpenRouter advertises for each model on 2026-10-25, which track the provider's list price for the same model. OpenRouter's API exposes them as prompt and completion price fields per token.
What does a 10 million input / 2 million output workload cost?
On GPT-5 it is $32.50: 10 x $1.25 plus 2 x $10.00. On Gemini 2.5 Flash the same workload is $8.00, and on Llama 3.3 70B it is $1.64.
Does a larger context window cost more?
Not per token on these plans. Claude Sonnet 4.5 lists a 1,000,000 token context and Gemini 2.5 Flash lists 1,048,576, both at the same per-token prices as their shorter requests.
Sources
- OpenRouter models APIChecked 2026-10-25
Cite this page
Nadia Voss (2026). “LLM API Prices per Million Tokens: 10 Models Side by Side”. Actual Database. https://actualdatabase.com/openrouter-llm-prices-per-million-tokens/