LLM API pricing comparison
Five LLM APIs side by side, in US dollars per 1M tokens: what one-api.ai debits, what another public reseller posts, and what the model's maker lists where we use their API.
Early access: billing is not open yet. Prices below are what a call will debit when it opens. Request access →
Prices per 1M tokens
| Model | Input | Cached input | Output | hiapi in / out | Official in / cached / out | Chatbot month |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | $0.328 | $0.007 | $1.314 | $0.44 / $1.32 | – / – / – | $8.862 |
| DeepSeek V4 Pro | $1.445 | $0.048 | $4.335 | $1.50 / $4.50 | – / – / – | $34.68 |
| GLM 5.3 | $2.21 | $0.41 | $6.95 | $4.40 / $15.40 | – / – / – | $54.00 |
| Claude Haiku 5.5 | $0.121 | $0.012 | $0.605 | – / – | $0.10 / $0.01 / $0.50 | $3.63 |
| Grok 4.7 | $2.421 | $0.605 | $7.263 | – / – | $2.00 / $0.50 / $6.00 | $58.104 |
The first three columns are one-api.ai prices for the standard context tier. hiapi prices were collected by hand on 2026-10-21; official prices were checked on 2026-10-31 and are shown only for models we would route to the maker's own API. A dash means no price is posted. The last column prices a support chatbot workload: 10,000 requests of 1,500 input and 300 output tokens, at the posted rate, which is what billing debits.
Long-context tiers
Two models charge more once a prompt crosses a size threshold, and the higher rate then applies to the whole request:
- Claude Haiku 5.5, prompts over 100,000 tokens: Input $0.605, Cached input $0.061, Output $3.026, Cache write $0.757 per 1M tokens. In the standard tier, a cache write costs $0.151.
- Grok 4.7, at or above 200k input tokens: Input $4.842, Cached input $1.211, Output $14.526 per 1M tokens.
How to choose
Lowest cost: DeepSeek V4.1 Flash is the cheapest per token by a wide margin and suits routing, tagging, extraction and short chat. Harder tasks: DeepSeek V4 Pro costs a few times more and is the better default for code and multi-step reasoning. Chinese-language products: GLM 5.3 is trained for Chinese and English, though its cached input saves less than DeepSeek's.
Comparison only: Claude Haiku 5.5 and Grok 4.7 are listed so you can compare their prices with the models in early access, but access to them is not sold today. Their pages explain where their long-context tiers start and link to Anthropic's and xAI's own price lists.
Per-token price is only half of the bill. Output tokens cost three to five times input tokens on every model here, so the output share of a workload drives most of its cost. The model pages work through three workloads each so you can see the effect before you commit.
Where the numbers come from
Every one-api.ai price is derived from the Cloudflare list price for that model, plus Cloudflare's 5% credit fee and our posted margin; the pricing page shows the list, floor and retail rows for each tier. Sources are linked on each model page with the date they were checked. If a source changes, the page and its date change with it.
Last updated 2026-10-31.