Official API
What a million tokens actually costs
Everything is put into the same unit before comparing: input, output and cache-hit prices are listed separately. CNY vendors are not converted to USD and USD vendors are not converted to CNY, so exchange rates can't inflate the gap.
| Model | Vendor | Input ↓ | Output ↓ | Cache hit | Context | Positioning |
|---|---|---|---|---|---|---|
| Claude Fable 5.2 | Anthropic | $8 | $40 | $0.2 | 1M | Flagship |
| Claude Opus 5.5 | Anthropic | $4 | $20 | $0.2 | 1M | Flagship |
| Claude Sonnet 5.5 | Anthropic | $2 | $10 | $0.1 | 1M | Balanced |
| Claude Haiku 5.5 | Anthropic | $0.1 | $0.5 | $0.01 | 200K | Lightweight |
| deepseek-v4-pro (peak) | DeepSeek | $1.32 | $3.96 | $0.044 | 1M | Flagship |
| deepseek-v4-pro (off-peak) | DeepSeek | $0.66 | $1.98 | $0.022 | 1M | Flagship |
| deepseek-flash (peak) | DeepSeek | $0.3 | $1.2 | $0.006 | 1M | Lightweight |
| deepseek-flash (off-peak) | DeepSeek | $0.15 | $0.6 | $0.003 | 1M | Lightweight |
| gemini-3.1-pro-preview | $2 | $12 | $1 | 1M | Previous flagship | |
| gemini-4-carbon Promo until 2026-11-30, then 4 / 20 from 12-01 | $2 | $10 | $0.1 | 1M | Flagship | |
| gemini-4-argon | $2 | $10 | $0.1 | 1M | Flagship | |
| gemini-4-hydrogen | $1.5 | $7.5 | — | 1M | Balanced | |
| gemini-3.8-flash | $0.75 | $3.75 | — | 1M | Lightweight | |
| gemini-2.5-flash | $0.3 | $2.5 | $0.15 | 1M | Legacy lightweight | |
| kimi-k3.1-pro-highspeed | Moonshot | ¥26 | ¥136 | ¥1.3 | 256K | Flagship, fast |
| kimi-k3 | Moonshot | ¥20 | ¥100 | ¥2 | 256K | General |
| kimi-k3.1-pro | Moonshot | ¥13 | ¥68 | ¥1.3 | 256K | Flagship |
| gpt-6-astra | OpenAI | $10 | $50 | $1 | 272K | Flagship |
| gpt-5.6-sol | OpenAI | $4 | $20 | $0.4 | 272K | Previous flagship |
| gpt-5.4 | OpenAI | $2.5 | $15 | $0.25 | 272K | Previous gen |
| gpt-6.1-sol | OpenAI | $2 | $10 | $0.1 | 272K | Flagship |
| gpt-4.1 | OpenAI | $2 | $8 | $0.5 | 1M | Legacy |
| gpt-6.1-terra | OpenAI | $1 | $5 | $0.05 | 272K | Balanced |
| gpt-4o-mini | OpenAI | $0.15 | $0.6 | $0.075 | 128K | Legacy lightweight |
| gpt-6-luna | OpenAI | $0.1 | $0.5 | $0.01 | 272K | Lightweight |
| GLM-5.3 | Zhipu GLM | ¥8 | ¥28 | ¥2 | 1M | Flagship |
| GLM-4.7 | Zhipu GLM | ¥2 | ¥8 | ¥0.4 | 200K | Legacy |
| GLM-5.3-Flash | Zhipu GLM | ¥0.8 | ¥2.8 | ¥0.23 | 1M | Lightweight |
Unit: official list price per 1M tokens, in the currency the vendor uses on its own pricing page. Captured 2026-11-15.
Four things people miss when comparing
1. Input and output are not the same price
Output usually costs 2–5× input. Pick a model on input price alone and long-answer workloads will be badly underestimated; conversely, if you only classify or extract and outputs are short, output price barely moves the bill.
2. Thinking tokens are billed as output
Gemini's pricing page says it outright: "Output price (including thinking tokens)". The longer a reasoning model thinks, the more the output column charges — often just a small-print line in the docs.
3. Long context costs more
OpenAI uses 272K as the cut-off: beyond it, input, output and cache all jump a tier. DeepSeek prices by time of day instead — off-peak is exactly half of peak.
4. Cache-hit pricing can differ 20×
gpt-6-astra: $10 input vs $1 cache hit. deepseek-v4-pro: $1.32 input vs $0.044 cache hit. If your requests share a long prefix (system prompt, knowledge base, long document), your cache-hit rate shapes the bill more than model choice.
A note on scope.This page lists official list prices: no tax, no proxy margin, no discounts. Promotional prices carry their own expiry — when a row has a date, check whether the promo is still live today.