Chat and Tutor conversations bill chat credits. The cost depends on which model answers your message — not on how many tokens you typed, though longer conversations carry more history into each request.

Credit range (0.5 – 12)

All text models are priced on a compressed 0.5–12 scale:

Reasoning depth — many models offer a reasoning-level picker (None → Max). Higher depth can improve hard problems but does not change the per-message credit — you always pay the flat model rate in the picker. Legacy separate "Medium" model rows (e.g. GPT-5.2 Medium) were folded into one model + a reasoning picker.

Zack (auto-router) is priced at 2.5 credits per message.

Retired models

Older keys (GPT-4.x, o3, o4-mini, Grok 4) are no longer in the picker. Old conversations may resolve to successors such as GPT-5.6 Terra or Grok 4.3 — switch to a current model when prompted.

Default vs named models

There are no separate Chat-Thinking / Chat-Fast / Chat-Chinese modes — use Zack, a named Chinese-capable model (e.g. DeepSeek / Qwen), or raise Reasoning when you need deeper thinking.

Agent tools (also chat credits)

Agent tools run during a chat message and add to the message's chat credit total. They are tracked separately in transaction history but draw from the same chat pool.

Other agent tools (flashcards, notes, mind maps, slides, etc.) follow the underlying model's per-message rate for the generation step.

How the per-message cost is calculated (technical)

For each model, Zense AI:

1. Reads OpenRouter price per million input/output tokens.
2. Applies the representative profile (12k in / 1k out per message).
3. Targets ~25% margin against the Pro-plan credit value (0.0495 HKD/credit).
4. Compresses the raw cost onto the 0.5–12 table so everyday models stay affordable and flagships are subsidized by cheaper tiers.

You are billed the final compressed credit, not the raw token math.

Tokens you will see in the UI

Even though billing is per message, each turn records token usage for debugging and analytics:

• promptTokens — system prompt + history + attachments + tool context sent to the model.
• completionTokens — the assistant's reply.
• reasoningTokens — extra tokens on reasoning models (GPT-5.6, thinking modes, etc.).

A long thread increases promptTokens on later messages, which is why the pricing formula assumes a heavy input profile — but your credit charge stays the flat model rate unless you also invoke paid agent tools.

Tips to manage chat spend

• Use Zack or 0.5-credit models for simple questions.
• Reserve flagship models (9–12 credits) for work that needs top quality.
• Web search (10 credits) adds up quickly — ask Chat to search only when needed.
• School admins: set per-user chat caps and monitor the shared pool in School Management → Users.

Full model-by-model table lives in the engineering docs (chat-pricing.md). Rates may be retuned when OpenRouter prices change — check the model picker for live costs.

> Full model-by-model table lives in the engineering docs (chat-pricing.md). Rates may be retuned when OpenRouter prices change — check the model picker for live costs.