AI Model Credit Rates: the same model costs different amounts on different platforms

This is the authoritative rate table for Soniva's AI chat and translation (token-based billing). Text-to-speech is billed per character — that's a separate table: Full Credit Rate Table.

⚠️ This page is generated directly from product code (web/scripts/gen-ai-model-rates.mjs). The rates you see in the app and the credits deducted on your bill come from the same data.

---

One thing to understand first: the same model name can have three prices

A model like DeepSeek-V4-Pro can be called from three places:

1. DeepSeek's own API;
2. the copy hosted on Volcengine Ark;
3. the copy hosted on Alibaba Cloud Model Studio.

Same model — but three independent pricing entities. Each sets and changes its own prices. So the tables below are grouped by platform, each with its own source and verification date.

A real example: GLM-4.7 costs 2 in / 8 out from Zhipu directly, and 3 in / 14 out when hosted on Alibaba Cloud Model Studio. Same model, nearly double the price.

So check which platform's entry you picked. In the app, hosted models sit in a "Models hosted on this platform" section, and every row shows its credit rate inline.

---

How the rate is calculated

credits per 10k chars = (input price × input tokens + output price × output tokens) ÷ 7.2 × 1000 × 1.15

• 1.15 is Soniva's only operating multiplier, covering payment processing fees, upstream retries and FX movement. It applies identically to every brand and every model — no platform is marked up individually.
• 1 USD = 7.2 CNY, the same FX basis as the TTS table.
• 1 token ≈ 1.5 characters: prices are quoted per token, "per 10k characters" is the unit you can actually feel. 1.5 is the conservative Chinese-text ratio, so English content typically costs less than the table shows.
• Per 10k, not per 1k: the TTS table uses 1,000 characters, but AI models are far cheaper — at that scale the cheapest and most expensive models both collapse to 1–4 credits, hiding a 50× real price gap.
• Output estimated at 0.6× input: the table shows the typical cost of feeding in 10,000 characters. Nobody knows in advance how long the reply will be.

🚨 So this number is typical, not a bill. Actual charges settle against the token usage returned by the upstream provider. Asking for a long piece of writing (output ≫ input) costs more; a one-line follow-up in a long conversation (input ≫ output) costs less.

---

DeepSeek (first-party)

Source: https://api-docs.deepseek.com/zh-cn/quick_start/pricing · basis: peak-hour, cache-miss price (the more expensive end) · last verified 2026-10-15

Alibaba Cloud Model Studio

Source: https://help.aliyun.com/zh/model-studio/model-pricing · basis: North China 2 (Beijing) list price, lowest tier; no promo or off-peak discounts · last verified 2026-09-21

Zhipu GLM

Source: https://open.bigmodel.cn/pricing · basis: list price, excluding limited-time discounts · last verified 2026-09-05

Volcengine Ark

Source: https://www.volcengine.com/docs/82379 · basis: lowest tier of the range (short-context tier) · last verified 2026-09-21

OpenRouter

Source: https://openrouter.ai/models · basis: OpenRouter USD list price × 7.2 × 1.055 (its 5.5% top-up fee is folded into the price) · last verified 2026-10-15

---

When credits are not charged

• Bring your own API key (BYOK): enter your own DeepSeek / Alibaba Cloud / Volcengine / Zhipu key in Settings and requests go through your own account. The upstream provider bills you directly and Soniva charges no credits. Free allowances those platforms hand out (e.g. 1M tokens per model on Model Studio) are yours to keep.
• Built-in system speech: never touches a cloud model. See the Full Credit Rate Table.

Will this table change?

Yes. When an upstream platform changes its prices, Soniva updates the catalog in code and regenerates this page, with the verification date updated. If the 1.15 multiplier ever changes, it will be announced in advance; credits you already own are never altered — only how many get deducted per 10,000 characters.

---

Related

• Full Credit Rate Table (text-to-speech, per character)
• What credits are and how they work
• Reading your usage ledger
• Choosing a plan