How tokens are counted & credits are charged — per model

This document explains, for the Rayu‑hosted (paid) models, exactly how token usage is turned into a credit charge, and the per‑model rate for every model currently offered.

TL;DR

1 credit = 1,000,000 tokens (the reference rate).

Every model has a rate = credits per 1,000,000 tokens (its creditMultiplier).

You are charged: credits = tokens_used × rate ÷ 1,000,000, where tokens_used is the actual input + output tokens the model reported.

A failed or unavailable request costs 0 credits.

For the underlying mechanism (reserve/settle, cache‑aware billing, the Redis counter), see credits-and-limits.md.

1. The charge formula

Charging happens in the gateway, off the model's real usage — not an estimate. For the active hosted models (served via Ollama Cloud), billing is flat: input tokens and output tokens are charged at the same rate, and there is no prompt‑cache discount, so:

billable_tokens = (input_tokens + output_tokens) × rate
credits_charged = billable_tokens ÷ 1,000,000
              = tokens_used × rate ÷ 1,000,000

rate is the model's credits‑per‑1M‑tokens value (the creditMultiplier). Because 1 credit = 1,000,000 tokens, the rate reads directly as “credits per million tokens”.

rate = 1.0 → 1,000,000 tokens costs 1 credit
rate = 2.5 → 1,000,000 tokens costs 2.5 credits
rate = 0.75 → 1,000,000 tokens costs 0.75 credits

Only tokens the provider actually reports are counted; the display in /usage shows the real tokens used, not a rounded‑up number.

2. Per‑model rates

Rates below are the current defaults. All of them are admin‑editable in the dashboard (Models → per‑model creditMultiplier); nothing here is hard‑coded into the charge.

Model (id)	Served via	Rate — credits / 1M tokens	Context window	Status
DeepSeek V4 Flash (`deepseek-v4-flash`)	DeepSeek API	0.33	1M	Active
DeepSeek V4 Pro (`deepseek-v4-pro`)	DeepSeek API	1.0	1M	Active
GLM‑5.2 (`glm-5.2`)	Ollama Cloud	2.5	1M	Active
Kimi K2.7 (`kimi-k2.7`)	Ollama Cloud	2.5	256K	Active
MiniMax M3 (`minimax-m3`)	Ollama Cloud	2.5	1M	Active
Llama 4 (`llama-4`)	Ollama Cloud	1.0	1M	Active
GPT‑OSS 120B (`gpt-oss-120b`)	Ollama Cloud	0.75	128K	Active
Qwen3.5 397B (`qwen3.5-397b`)	Ollama Cloud	0.75	256K	Active
Qwen3.5 122B (`qwen3.5-122b`)	Ollama Cloud	0.75	256K	Active
LongCat 2.0 (`longcat-2`)	LongCat	0.5	1M	Disabled

Notes:

DeepSeek (official API) and Ollama Cloud are both active right now (LongCat's provider row has enabled = false in Admin → Providers). A disabled provider's models are hidden from users and any request to them is refused (with no charge). DeepSeek V4 Flash/Pro are served DIRECT through DeepSeek's own Anthropic-compatible API (provider deepseek); the remaining hosted models are served through Ollama Cloud.
MiniMax M3 and Llama 4 rates were not individually specified and use a sensible default (2.5 and 1.0) — adjust in the admin dashboard as needed.
Context window is what the CLI reports for the model (used for context management). If a model is served by an upstream with a smaller real window, set a per‑model context override to avoid overflow.

Rate tiers at a glance

0.33 / 1M — DeepSeek V4 Flash (cheapest)
0.75 / 1M — GPT‑OSS 120B, Qwen3.5 397B, Qwen3.5 122B
1.0 / 1M — DeepSeek V4 Pro (reference), Llama 4
2.5 / 1M — GLM‑5.2, Kimi K2.7, MiniMax M3

3. Worked examples

Assume the reference rate (1 credit = 1,000,000 tokens).

Model	Tokens used (in + out)	Charge
DeepSeek V4 Pro (1.0)	1,000,000	1.00 credit
DeepSeek V4 Pro (1.0)	200,000	0.20 credit
DeepSeek V4 Flash (0.33)	1,000,000	0.33 credit
GPT‑OSS 120B (0.75)	1,000,000	0.75 credit
GLM‑5.2 (2.5)	1,000,000	2.50 credits
GLM‑5.2 (2.5)	40,000 (a small turn)	0.10 credit

A trivial message (e.g. a few thousand tokens) costs a tiny fraction of a credit — it is not rounded up to a whole credit.

4. What a plan buys

A plan grants a credit allowance per billing period (creditsPerPeriod). Because 1 credit = 1,000,000 tokens, the allowance converts to tokens per model by dividing by the model's rate.

Example — a plan with 50 credits / period:

If you only used…	You'd get about…
DeepSeek V4 Pro (1.0)	50,000,000 tokens (50 credits ÷ 1.0)
DeepSeek V4 Flash (0.33)	~151,000,000 tokens (50 ÷ 0.33)
GPT‑OSS 120B (0.75)	~66,000,000 tokens (50 ÷ 0.75)
GLM‑5.2 (2.5)	20,000,000 tokens (50 ÷ 2.5)

Credits deplete over the period and reset at renewal. When the balance is exhausted, further hosted requests return a credit‑limit error (or draw from top‑up credits if the plan enables top‑up). Plan credit amounts and prices are admin‑configured — see credits-and-limits.md.

5. Edge cases (what you are NOT charged for)

Failed / errored request — if the upstream fails, usage settles to the real amount (0 for a failed turn), so a failed request costs 0 credits.
Disabled provider — a request for a model whose provider is turned off returns 503 before any reserve, so it charges 0 credits and does not consume a daily turn.
Rate‑limited key — the gateway rotates/fails over across the provider's keys automatically; you are only charged for the request that actually succeeds, at the model's rate.

6. Where the numbers live (admin‑editable)

What	Where
Per‑model rate (`creditMultiplier`) + prices	Admin → Models (`hosted_models` table)
`1 credit = 1,000,000 tokens` baseline (`baselineCreditsPer1M`)	Admin → Credit Settings (`app_settings`)
Plan credit allowance (`creditsPerPeriod`)	Admin → Plans
Which providers are active	Admin → Providers (`providers.enabled`)
Where/how a provider is called (base URL, wire format, auth)	Admin → Providers (`providers` table)
Per‑model capabilities (thinking, image input)	Admin → Models (`supportsReasoning` / `supportsImage`)

The gateway computes every charge from these values at request time — changing a rate in the dashboard changes the charge with no code change or redeploy of the CLI. The CLI is display‑only; the gateway is the single source of truth for billing.