Model Pricing Overview

AI Gateway lets you call virtually all mainstream large models on the market (Claude, GPT, Gemini, Qwen, DeepSeek, GLM, Kimi, Doubao, and more) through a single interface and a single bill — no need to register, integrate, or top up accounts on each provider's platform separately. The table below lists the official list prices for each model, organized by currency zone — domestic (CNY) and international (USD) — with a separate table per vendor for easy comparison.

Billing in Three Minutes

Billing dimension descriptions

Dimension	Description
Input	Unit price per prompt token in the user's request
Output	Unit price per token of model-generated content
Explicit cache write	Unit price for writing a prompt into the context cache for the first time
Explicit cache write · 5m / 1h	Anthropic's two cache write price tiers
Cache hit	Discounted price for reusing a cached prompt on subsequent requests

Large models are not billed per call — they are billed by the volume of text processed, measured in units called tokens.

💡 Tip: How to estimate cost? Example: using a model for Q&A with 1,000 characters of input and 2,000 characters of output ≈ 1,000 input tokens + 2,000 output tokens. If the input price is ¥2/million and the output price is ¥8/million, the cost is approximately 1000÷1,000,000×2 + 2000÷1,000,000×8 = ¥0.018.

What Is Caching and Why Does It Save Money?

If you repeatedly send the same opening content (for example, the same long system prompt or the same knowledge base document every time), the model can "remember" that content and reuse it directly on subsequent calls without recomputing it — this is caching. The cached portion is priced far below the normal input price, resulting in significant savings.

There are two types of caching:

Cache type	Plain-language explanation	Billing characteristics
Explicit cache (manual)	You actively tell the model "store this segment." Like renting a locker: storing it incurs a one-time write/create fee (slightly above the input price); each subsequent retrieval incurs a low hit fee; long-term storage may also incur a storage fee.	Write fee + hit fee (+ storage fee)
Implicit cache (automatic)	The system automatically detects repeated prefixes and caches them for you — no action required. Like a store automatically giving a discount to regular customers: no write fee; you simply enjoy the lower price on a cache hit.	Hit fee only, no write fee

⚠️ Note: Not every provider supports both types! This directly determines which cache columns appear in the pricing tables:

Provider	Explicit cache	Implicit cache	Notes
Alibaba Cloud · Qwen	✅ Supported	✅ Supported	Both types, highest flexibility
OpenAI · GPT	❌ Not supported	✅ Supported	Automatic implicit cache only, no manual operation needed
Anthropic · Claude	✅ Supported	❌ Not supported	Manual explicit cache only, write available in 5-minute / 1-hour tiers
Google · Gemini	✅ Supported	❌ Not supported	Explicit cache, additional hourly storage fee
DeepSeek	Partial	Partial	v3.2 supports both; r1 / v3.1 implicit only; v4 series not yet available
Zhipu · GLM	Partial	✅ Supported	GLM-5.1 supports both; others implicit only
Moonshot · Kimi	✅ Supported	✅ Supported	Both types
MiniMax	❌ Not supported	✅ Supported	Implicit only
ByteDance · Doubao	✅ Supported	✅ Supported	Explicit incurs storage fee; implicit takes effect in batch mode

When reading the tables: a "Explicit·Write/Create" column = explicit cache supported; an "Implicit·Hit" column = implicit cache supported; — or a missing column = that model does not support that cache type.

Domestic Zone (CNY · ¥ / million tokens)

Alibaba Cloud · Qwen Series

> Positioning: China's all-around model, covering general conversation, coding, vision, speech, and multimodal, with a context window up to 1 million tokens. Both explicit and implicit caching are supported.

Model	Context window	Input	Output	Explicit·Create	Explicit·Hit	Implicit·Hit
qwen3.6-max-preview (strongest)	0–128K	9	54	11.25	0.9	—
	128K–256K	15	90	18.75	1.5	—
qwen3.6-plus (flagship general)	0–256K	2	12	2.5	0.2	—
	256K–1M	8	48	10	0.8	—
qwen3.6-flash (fast & low-cost)	0–256K	1.2	7.2	1.5	0.12	—
	256K–1M	4.8	28.8	6	0.48	—
qwen3.5-plus	0–128K	0.8	4.8	1	0.08	0.16
	128K–256K	2	12	2.5	0.2	0.4
	256K–1M	4	24	5	0.4	0.8
qwen3.5-flash	0–128K	0.2	2	0.25	0.02	—
	128K–256K	0.8	8	1	0.08	—
	256K–1M	1.2	12	1.5	0.12	—
qwen3-max	0–32K	2.5	10	3.125	0.25	0.5
	32K–128K	4	16	5	0.4	0.8
	128K–256K	7	28	8.75	0.7	1.4
qwen3-coder-plus (coding)	0–32K	4	16	5	0.4	0.8
	32K–128K	6	24	7.5	0.6	1.2
	128K–256K	10	40	12.5	1	2
	256K–1M	20	200	25	2	4

DeepSeek Series

> Positioning: Known for exceptional cost-effectiveness and strong reasoning capabilities, suitable for budget-sensitive scenarios that still demand quality. Cache support varies by model version (see table below; missing columns indicate no support).

Model	Input	Output	Explicit·Create	Explicit·Hit	Implicit·Hit
deepseek-v4-pro (flagship)	12	24	—	—	2.4
deepseek-v4-flash (fast)	1	2	—	—	0.2
deepseek-v3.2	2	3	2.5	0.2	0.4
deepseek-r1 (deep reasoning)	4	16	—	—	0.8

Zhipu · GLM Series

> Positioning: Balanced domestic general-purpose model; GLM-5 series is the new flagship generation. Most versions support implicit cache only; GLM-5.1 additionally supports explicit cache.

Model	Context window	Input	Output	Explicit·Create	Explicit·Hit	Implicit·Hit
glm-5.1 (flagship)	0–32K	6	24	7.5	0.6	1.2
	32K–200K	8	28	10	0.8	1.6
glm-5	0–32K	4	18	—	—	0.8
	32K–198K	6	22	—	—	1.2
glm-4.7	0–32K	3	14	—	—	0.6
	32K–166K	4	16	—	—	0.8

Moonshot · Kimi Series

> Positioning: Excels at understanding and processing ultra-long text. Both explicit and implicit caching are supported.

Model	Input	Output	Explicit·Create	Explicit·Hit	Implicit·Hit
kimi-k2.6	6.5	27	8.125	0.65	1.3
kimi-k2.5	4	21	5	0.4	0.8

MiniMax Series

> Positioning: Cost-effective general-purpose model. Implicit cache only (system discounts automatically, no action required).

Model	Input	Output	Implicit·Hit
MiniMax-M2.7	2.1	8.4	0.42
MiniMax-M2.5	2.1	8.4	0.42

Volcano Engine (Doubao)

> Positioning: China's high-value all-in-one suite, covering text, vision, video, images, and 3D. Explicit cache (additional storage fee of ¥0.017/million token·hour) and implicit cache (effective in batch mode) are both supported. The table below shows standard "online inference" prices; Doubao also offers approximately 50% off for batch inference.

Model	Context window	Input	Output	Explicit·Hit
doubao-seed-2.0-pro (flagship)	[0, 32K]	3.2	16	0.64
	(32K, 128K]	4.8	24	0.96
	(128K, 256K]	9.6	48	1.92
doubao-seed-2.0-code (coding)	[0, 32K]	3.2	16	0.64
	(32K, 128K]	4.8	24	0.96
	(128K, 256K]	9.6	48	1.92
doubao-seed-2.0-lite	[0, 32K]	0.6	3.6	0.12
	(32K, 128K]	0.9	5.4	0.18
	(128K, 256K]	1.8	10.8	0.36
doubao-seed-2.0-mini (cheapest)	[0, 32K]	0.2	2	0.04
	(32K, 128K]	0.4	4	0.08
	(128K, 256K]	0.8	8	0.16
doubao-seed-1.6	[0, 32K]	0.8	2 / 8 ※	0.16
	(32K, 128K]	1.2	16	0.16
	(128K, 256K]	2.4	24	0.16
doubao-seed-1.6-flash (fast)	[0, 32K]	0.15	1.5	0.03
	(32K, 128K]	0.3	3	0.03
	(128K, 256K]	0.6	6	0.03
doubao-seed-1.6-vision (vision)	[0, 32K]	0.8	8	0.16
	(32K, 128K]	1.2	16	0.16
	(128K, 256K]	2.4	24	0.16
doubao-1.5-pro-32k	—	0.8	2	0.16
doubao-1.5-lite-32k	—	0.3	0.6	0.06