One API key.
Every top Chinese model.

Drop-in OpenAI compatibility for DeepSeek, Qwen and more. Streaming, tool calls, per-token billing. Switch models by changing one string.

No subscription. Pay only for what you use. Keys are live in 30 seconds.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.tokenforge.dev/v1",
    api_key="sk-live-...",
)

stream = client.chat.completions.create(
    model="deepseek-chat",
    messages=[{"role": "user", "content": "Explain vector search in 2 sentences."}],
    stream=True,
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

Models & pricing

Per 1M tokens · USD
ModelContextInputCached inputOutput
deepseek-chat
通用对话与创作,性价比最高
1,000,000$0.250
peak $0.425
$0.008
peak $0.014
$1.45
peak $2.46
deepseek-reasoner
深度推理,适合复杂任务与代码重构
1,000,000$1.10
peak $1.87
$0.040
peak $0.068
$4.75
peak $8.07
qwen-plus
通义千问,国内调用延迟低
131,072$0.400$0.100$1.20

Billing is metered on every request. Cached input is charged only when the provider reports a prompt cache hit — otherwise you pay the normal input rate.

OpenAI-compatible

Point any OpenAI SDK at our base URL. No code rewrite, no new abstractions.

Automatic failover

If a provider degrades, requests are retried on a healthy channel of the same model.

Prepaid, no surprises

Credits never auto-recharge. Set a key-level quota and you cannot overspend.

Usage you can audit

Every request logged: tokens, cached tokens, cost, latency. Export to CSV.

FAQ

How does billing work?

We meter prompt tokens, cached prompt tokens and completion tokens separately, at the rates above. Cost is deducted from your prepaid balance per request. You can see every line item in the dashboard.

What happens when my balance runs out?

Requests return HTTP 402. Nothing is charged, nothing is queued. Top up and the same key keeps working.

Do you store my prompts?

Prompts are not persisted by default. We store request metadata (model, token counts, latency, status) for billing and abuse review.

Can I use this in production?

Yes. We publish rate limits per key and support streaming, tool calls and JSON mode across all listed models.

Which regions are supported?

The service is available worldwide except where restricted by law. Accounts registered from unsupported regions are declined at signup.