Token use per task: find the jobs that use up your AI plan

noHuman Team 4.4 counts the tokens behind each task your noHumans report. Open Work beside a chat to see a count on the task cards and the tasks that used the most tokens in the last 7 days. Each noHuman can check its own numbers with the team:token-usage tool, and a leader can check its whole team.

Status: new in noHuman Team 4.4, out for Mac on 25 October 2026, with the Windows build to follow. Details are in the 4.4.0 release notes and the 4.4 roundup.

Which tasks use the most tokens?

Open a noHuman's chat and select Work, next to Profile. Below Done, the section Token use · last 7 days ranks up to five tasks by tokens used, each with a bar. A line above the list gives the week's total, the number of model calls and the share reused from cache.

Each task card carries its own count beside its status and age. Hold the pointer over the count for the breakdown: input, output, tokens reused from cache with their share, the largest prompt a single call sent, and how many calls the app "counted by timing".

A leader's Work adds a My team list. Each direct report gets a row with its latest task and its "tokens this week", so you can spot the expensive teammate before you open its panel. If the Work panel is new to you, start with the Work panel guide and the conversation docs.

The mobile companion shows the same numbers in its Work view, which opens from the Work button in each chat: tokens per task, the 7-day list and the team rows, without the hover breakdown. Its iOS and Android builds are prepared for store release.

How does noHuman Team know which task a model call was for?

From the noHuman's own work reports. A noHuman names its task in each report, and the app ties the model calls of that turn to that task. Work a teammate handed over with a delegation id counts the same way. The app reminds each noHuman that naming the task in its report ties that turn's tokens to it.

Calls in a turn that named no task go to the task the noHuman last reported as working or waiting, with the mark "counted by timing". Calls made while no task was open land on a separate line, "Not tied to a task". A quick question of yours that never became a task would end up there.

The split depends on the reports. The Work panel is the noHuman's own account of its work, so a noHuman that skips reports gives you a rough split. If the app counted most of a task's calls by timing, team:token-usage warns that the split between tasks is approximate.

Is a token count the same as my bill?

No. The count adds up input, output, cache reads and cache writes. A cache read is conversation the provider reused instead of processing it again. The panel's note reads: "Reused cache tokens cost far less than new ones, but count toward subscription limits." A task with a high cache share can still eat your plan's window.

For money, open Activity → Usage, where a note says the costs "combine reported amounts and price-table estimates; they are not your subscription bill." Your AI provider bills you for its subscription or usage, as the pricing FAQ explains. Use token counts to compare one task with another. For a broader budget, see The True Cost of Running AI Agents.

Does Activity → Usage include my ChatGPT noHumans?

Yes, from 4.4. noHumans on ChatGPT and other models run through the opencode engine, whose telemetry names each call but carries no token counts. Before 4.4, Usage left those calls out, so you missed the ChatGPT side of a mixed team.

The app now copies exact input, output, reasoning and cache tokens from each noHuman's own opencode records. It syncs every few minutes and again when you open Usage. Claude Code reports its own calls, and the app skips those steps in the opencode records, so no call counts twice.

Can a noHuman check its own token use?

Yes. Each noHuman has the read-only team:token-usage tool. It returns tokens per task over 1 to 14 days, 7 by default, with calls, input, output, cache, the average and largest prompt per call, and the models used. A leader can read any noHuman under it, or ask for "all" to cover its whole branch.

The tool adds hints drawn from the numbers:

• If a task's calls averaged 150,000 tokens of conversation or more, it suggests compacting or resetting that noHuman before unrelated work, or handing the task to one with a fresh session.
• If a task made 40 calls or more that wrote under 150 tokens each on average, it points to polling, a waiting loop or a repeated check.

The tool's instructions also mention a smaller model for a simple task and splitting a sprawling one. The hints state facts; the leader decides what to change.

To use it, ask your top leader which tasks used the most tokens this week. It can read the whole team in one call, then compact, reset or move a teammate to another model with its leader tools. You can do the same by hand. Make room in Profile compacts a long conversation (context management FAQ), and the model selector beside Send switches models (models FAQ). To match models to tasks, read How to Choose the Right AI Model for Every Task.

What happens when a noHuman hits its plan's usage limit?

It pauses until the provider's reset time and queues new messages. Its status reads "waiting for quota until" that time. In 4.4, switching that noHuman to another model ends the pause at once. The queued messages go out, and the chat says it resumed and how many messages it has to deliver.

A noHuman that restarts on a different model, as a provider switch does, starts without the old pause. If the new model has hit its cap too, the noHuman waits for that provider's own reset, and the queue stays intact.

On 24 October I moved my Executive Producer noHuman to another provider during one of these pauses. It kept waiting out the old provider's window, with every queued message stuck behind it. Version 4.4 ends that wait.

Which earlier features keep token use down?

Three changes from 4.1.7 and 4.3.0 limit background work and runaway chains:

• Helper model (4.1.7). In Team settings → AI models → Helper model, you choose one small, fast model for the Archivist, heartbeat checks and loop checks, apart from each noHuman's own model. Automatic uses the fastest small model your Claude or ChatGPT subscription offers, and Test now shows whether it answers. Heartbeat checks can use tokens with no visible message, as the heartbeat FAQ notes.
• Loop guard (4.1.7). A message several hops into a chain of noHuman-to-noHuman messages, with no real work in between, waits for Continue instead of starting another turn. Since 4.3.0 the pair's leader makes that call, and the question comes to you if no leader stands over both. With Check before pausing on, the helper model reads the last few messages first and lets real progress carry on.
• Coding helper limits (4.3.0). Up to 3 coding helpers run per noHuman, with a queue, and each stops after a long silence or an hour.

The 4.1.7 and 4.3.0 release notes cover these in full.

Quick answers

How far back do the numbers go?

Task cards count the last 14 days, the Token use list covers 7, and team:token-usage takes 1 to 14. The app keeps its raw usage records for 14 days by default.

Does a high token count mean a task went badly?

No. Tokens measure how much the model read and wrote. To judge the result, open the task's timeline and the files it produced.

Can I see token use on my phone?

Yes, in the mobile companion's Work view. The phone needs your computer running with the team on it. See the mobile guide.

What else is new in 4.4?

You can choose how long an approval lasts, and noHumans can put charts, diagrams and forms in their replies.

Install noHuman Team to see which tasks use up your AI plan. The monthly subscription starts with a 7-day trial; payment details are required at checkout. Your AI provider charges separately.