How do limits work?
Each paid plan includes one monthly dollar budget for models, paid tools, voice and active Cloud Computer work. Usage is charged at twice the provider cost: $1 of provider cost uses $2 of Notis credit. Idle cloud time and retries caused by a Notis fault are not charged to you. Your billing cycle, included usage and On-Demand usage are in Settings → Plan, and the breakdown of each request and its cost is in Settings → Usage. In a Manager thread, the usage indicator also shows the thread’s completed cost and how much context the latest completed run used. Each paid plan includes a different monthly amount: see Plans & trial. On a team, every seat’s amount goes into one shared pool that any teammate can use. Notis manages model routing automatically based on the task. Prices below are shown per 1M tokens and reflect the current Notis rates after markup.When you use your own Claude or ChatGPT subscription, that agent’s model usage stays on that subscription. When Notis hands a task to Codex or Claude in your Cloud Computer, the model bill lands on the subscription you already pay for, and the rates below do not apply. Notis orchestration, paid tools and active Cloud Computer work still use your Notis credit. There is no separate hours allowance.
Auto mode
By default, Notis runs on Auto (recommended). Instead of a fixed model, Notis’s smart router reads each turn and picks the intelligence level — model and reasoning effort together — that fits the task: light, inexpensive work for a quick reply, and a stronger model for a hard, multi-step task. Because the choice is made per turn, a single conversation can move up and down levels as the work changes, so you only pay for the capability each step actually needs.Intelligence levels
You can pin an intelligence level — per conversation in the Desktop and Web app, as a per-channel default in Channels, or per automation in the automation editor — instead of letting Auto choose. Each level sets the model and its reasoning effort together:
The multiplier compares each level’s regular input and output rates with Medium, before cached-input discounts, which differ by model. Reasoning can also change how many tokens a message uses, so actual billing always uses the real token counts and the rates per level below.
Processing tiers (priority)
Notis chooses the processing tier automatically, but on a plan with priority access you can pin one per thread or per channel alongside the intelligence level (it shows as Standard / Fast / Economy in the picker):
If a model does not support Economy or Fast, Notis uses its Standard rate instead.
Intelligence and speed preferences, including automation settings, apply to the agents doing your work. The coordinator that organizes each conversation always uses Low intelligence and Fast speed; these preferences do not change it.
Rates per level
Rates are per 1M tokens on the Standard tier. Economy halves them and Fast doubles them.
For unusually large requests with more than 272,000 input tokens in total
(including cached input), the whole request uses long-context rates: input and cached
input are charged at 2× the table rate, and output at 1.5×. A request with exactly
272,000 input tokens still uses the regular rates above.
Tool rates
Specialized tools have their own rates. Web search and Deep Research are charged on the tokens they use, with no separate per-search or per-report price. Economy halves their rates and Fast doubles them.
All ChatGPT Image models share these rates, and text output is not charged. What changes the cost of one image is explained in Images.
Voice Mode also uses the twice-provider-cost rate, including call transport when applicable. Its cost depends on the duration and usage reported by the provider, not a separate minutes allowance. Other image, audio and video tools use their current provider units and rates; check your usage breakdown for the amount charged.
Pay-per-use tools are counted the same way. Each one has its own price per use, shown on its page, and it comes out of the same monthly usage.