DeepSeek-class models are among the cheapest frontier-quality models you can run. The only friction is access: the official channels want a Chinese phone number, and top-ups go through Alipay or WeChat Pay. NovaToken removes that step — one OpenAI-compatible endpoint, card payments in USD, first call in under a minute.
https://api.novatoken.dev/v1. Your prompts are not stored, credits never expire.
DeepSeek-class models are extremely popular with developers outside China — the models are open-weight, cheap, and strong at reasoning and code. But three things block most international users:
NovaToken is a reseller gateway built specifically for this gap: we run the models on our own infrastructure and partner inference capacity, bill you in USD with a normal card, and expose everything through the standard OpenAI API shape you already use.
base_url to https://api.novatoken.dev/v1 and use your new key.All served through the same endpoint, same key, same per-token billing:
| Model | Best for | Price / 1M tokens |
|---|---|---|
| DeepSeek-V4-Flash | Reasoning, long documents (1M context) | $0.30 in / $1.20 out |
| Qwen2.5-7B-Instruct | Fast chat, classification, tool calling | $0.10 in / $0.30 out |
| GLM-5.3-Flash | Bilingual tasks, agents | $0.15 in / $0.50 out |
| Kimi-K2.7-Code | Code, agentic workflows (262K context) | $1.20 in / $5.00 out |
DeepSeek-V4-Flash also supports prompt caching: input tokens that hit the upstream cache are billed at $0.006 / 1M instead of $0.30. Nothing to configure — the discount applies whenever the provider reports a cache hit, which is typical for long system prompts, agents and coding assistants.
$5. That's roughly 25 million input tokens or 11 million output tokens on DeepSeek-V4-Flash — enough to build and test a real application. Credits never expire, there is no subscription, and you can top up again any time in $5 / $20 / $50 increments.
Buy $5 of credits Create free account
Yes. NovaToken only needs an email address and a password. Nothing is sent to your phone, no real-name verification, no local ID required.
Major credit and debit cards worldwide, in USD, processed by our payment provider (Creem). PayPal-style wallets are not needed — a normal card works.
You keep using the standard openai SDK family (Python, Node, Go, Rust…) — every language that speaks the OpenAI API works. Chat completions, streaming SSE, function/tool calling and JSON mode are all supported.
No. We keep timestamps and token counts for billing only, purged on a rolling basis. Request bodies are not retained.
DeepSeek-V4-Flash is a reasoning model: it thinks before answering, and thinking tokens are billed as output (industry standard). With stream: true the first bytes still arrive quickly. If you want zero thinking delay, use Qwen2.5-7B-Instruct.
Top up again in $5 / $20 / $50 increments — credits are added within a minute. Nothing about your keys, code or history changes.