2026 年 11 月 AI 模型 API 價格:每百萬 token 由 $0.02 至 $180
AI model API prices in November 2026: from $0.02 to $180 per million tokens
AI-Sovereign 比較 373 個 AI 模型 API 的價格,截至 2026 年 11 月 22 日,每百萬 token 輸入價格低至 $0.017,輸出價格最高為 $180。
Picking an AI model used to be a question of quality. In late 2026 it is just as much a question of price, because the gap between the cheapest and the most expensive model you can call by API is now enormous.
We track 373 AI model APIs in the AI-Sovereign directory, with input and output prices per million tokens and the context size of each one, refreshed every day. This report looks at what the numbers say on 22 November 2026.
The short version
- The cheapest capable models cost $0.02 to $0.10 per million input tokens. At that price, a million short chatbot replies cost a few dollars.
- The most expensive model on the list, GPT-5.5 Pro, costs $30 in and $180 out per million tokens. That is 6,000 times the output price of the cheapest model on the list ($0.03 for Mistral Nemo).
- A one-million-token context window is no longer rare. OpenAI, Google, Qwen, DeepSeek, Z.ai and NVIDIA all sell models that read about a million tokens at once, and two Grok models read two million.
- Only 3 of the 373 APIs are listed as free.
The cheapest models from the big labs
These are the lowest prices we see from established labs, for models with at least 128K tokens of context:
| Model | Lab | Input / 1M tokens | Output / 1M tokens | Context |
|---|---|---|---|---|
| Llama 3.1 8B Instruct | Meta | $0.02 | $0.04 | 131K |
| Mistral Nemo | Mistral | $0.029 | $0.03 | 131K |
| Gemma 3 4B | $0.04 | $0.05 | 131K | |
| gpt-oss-20b | OpenAI | $0.017 | $0.085 | 131K |
| GLM 5.3 Flash | Z.ai | $0.04 | $0.078 | 1M |
| Qwen3.7 Flash | Qwen | $0.03 | $0.13 | 1M |
| DeepSeek V4.1 Flash | DeepSeek | $0.05 | $0.20 | 1M |
| Nemotron 3.5 Lightning | NVIDIA | $0.07 | $0.20 | 1M |
Two things stand out. First, the cheapest end of the market is dominated by open-weight models (Llama, Mistral, Gemma, gpt-oss), which many providers compete to serve. Second, the "flash" tier now comes with a million-token context: GLM 5.3 Flash, Qwen3.7 Flash and DeepSeek V4.1 Flash read a whole codebase or a stack of contracts for a few cents.
The most expensive models
| Model | Lab | Input / 1M tokens | Output / 1M tokens |
|---|---|---|---|
| GPT-5.5 Pro | OpenAI | $30 | $180 |
| GPT-5.4 Pro | OpenAI | $30 | $180 |
| GPT-5.2 Pro | OpenAI | $21 | $168 |
| GPT-5 Pro | OpenAI | $15 | $120 |
| o3 Pro | OpenAI | $20 | $80 |
| Claude Opus 4.1 | Anthropic | $15 | $75 |
The top of the price list belongs to the "Pro" reasoning models, which spend far more compute on every answer. They make sense for hard, high-value questions where one good answer is worth dollars. For anything that runs at volume, they are the wrong tool.
What the newest flagships cost
The newest frontier models from the big labs cluster in two price bands:
| Model | Lab | Input / 1M | Output / 1M | Context |
|---|---|---|---|---|
| GPT-6.2 Astra | OpenAI | $8 | $40 | 1.05M |
| Claude Fable 5.2 | Anthropic | $8 | $40 | 1M |
| Gemini 4 Neon | $3 | $15 | 1M | |
| GPT-6.2 Sol | OpenAI | $2 | $10 | 1.05M |
| Gemini 4 Hydrogen | $1.50 | $7.50 | 1M | |
| Qwen4 Max | Qwen | $1.43 | $4.28 | 1M |
| GPT-6.1 Terra | OpenAI | $1 | $5 | 1.05M |
The top flagships sit at $8 in and $40 out. One step down, at $1 to $3 in, every lab now sells a model with a million-token context, which is where most production workloads will land.
The biggest context windows
| Model | Lab | Context | Input / 1M | Output / 1M |
|---|---|---|---|---|
| Grok 4.20 | xAI | 2M | $1.25 | $2.50 |
| Llama 4 Scout | Meta | 1.31M | $0.10 | $0.30 |
| GPT-6.2 Astra | OpenAI | 1.05M | $8 | $40 |
Llama 4 Scout is the surprise here: 1.3 million tokens of context for $0.10 in and $0.30 out.
How to read these numbers
- Price is per million tokens. A token is roughly three quarters of an English word, so a million tokens is about 750,000 words.
- Output usually costs more than input, often four to six times more. A chatbot that writes long answers pays mostly for output; a tool that reads long documents and returns a short summary pays mostly for input.
- The cheapest model is rarely the cheapest system. A model that needs three attempts costs more than a better one that gets it right the first time. Test your real prompts on two or three price tiers before you commit.
Methodology
Prices and context sizes come from OpenRouter's public model list, which shows the price of each model across the providers that serve it. We refresh them every day. See every model, sorted by price or context, on the AI Model APIs page, and read how AI-Sovereign ranks projects.
Originally published on AI-Sovereign, the daily ranking of AI models, APIs, agents and open-source projects.
來源:DEV Community · dev.to