Concepts · Weighted Tokens

How Anoman bills subscription quota.

A single formula maps raw provider tokens onto your monthly plan quota. Premium models cost more weighted tokens; regional routing costs less; cache + batch costs nothing.

Why weighted tokens

One quota, every model

Raw provider tokens are a bad billing unit. A token of GPT-5 output costs ~40× a token of Gemma 2 9B output. If we billed raw tokens, your monthly quota would mean wildly different things depending on which model you chose, and you’d either need a separate quota per tier or a pricing model that’s impossible to budget.

Weighted tokens normalize this. Every raw token gets multiplied by the model’s tier multiplier, so 1,000 raw tokens of GPT-5 deducts a lot more from your quota than 1,000 raw tokens of Gemma. You can move freely between models without renegotiating quota.

The formula

weighted_tokens = raw_tokens
                  × provider_multiplier   (model's data-processing region)
                  × tier_multiplier       (Pivot v2 model tier)
                  × quota_discount        (cache + batch savings)

Four factors multiply together. The model’s tier_multiplier is the largest lever; routing decisions (provider_multiplier, quota_discount) reduce it. Cache hits zero it out entirely.

Tier multipliers

Pivot v2 brackets

Tier	Multiplier	Example models
budget	1	qwen-2.5-7b, llama-3.1-8b, gemma-2-9b
mid	4	claude-sonnet-4-6, gpt-4o, deepseek-v3, gemini-1.5-pro
premium	17	claude-opus-4-7, gpt-5.5, gpt-4-turbo
ultra	42	Future flagship tier (reserved)

Multipliers are chosen to approximate the cost ratio between providers. Premium ≈ 17 × budget; ultra ≈ 42 × budget. Each model’s tier is visible on its detail page.

Provider multipliers

Regional incentives

Provider type	Multiplier	Why
cloud_direct	1.0	Default — model runs in upstream’s home region
bedrock	0.5	Bedrock-regional routing — incentive to stay regional
self_hosted	0.5	Models hosted in our Indonesia data center
local_id	0.3	Indonesia-resident open-source models (sahabat-ai variants)

The same model can be available through multiple provider types. Picking a regional variant cuts your weighted-token spend.

Quota discounts

Cache + batch reduce quota cost

Semantic cache hit — quota_discount = 0%. Zero quota deducted; the cached response is served. No upstream provider call is made.
Provider cache hit — quota_discount = 90%. We still call the upstream but only pay (and bill) 10% of normal. Anthropic + OpenAI + Google all offer prompt caching for the system prompt prefix.
Batch routing — quota_discount = 50%. Quota deducted at half rate when you opt-in to batch via the header.
Realtime + no cache — quota_discount = 100%. Full weighted-token cost.

Worked examples

Sanity-check the math

Pro plan quota:           20,000,000 weighted tokens / month
Request:                  1,000 raw tokens to claude-sonnet-4-6
                          (mid tier, cloud-direct US)

provider_multiplier:      1.0   (cloud_direct)
tier_multiplier:          4     (mid)
quota_discount:           100%  (realtime, no cache hit)

weighted_tokens = 1000 × 1.0 × 4 × 1.0 = 4,000

Available budget:         20,000,000 / 4,000 = 5,000 such requests/month

Where to see weighted tokens

Every response — _anoman.weighted_tokens field on the chat completion body.
Traces — per-request weighted_tokens column in the dashboard Traces table.
Usage page — current-month weighted-token spend rolled up by model + region.
Model detail page — each model’s detail page shows the provider × tier multiplier so you can predict spend before calling.

See your current weighted-token spend.

Usage dashboard shows MTD + 7d + 30d trends.

Open Usage All concepts