How to Track and Control Team Token Spend

What is token spend

Token spend is the number of tokens a team consumes with AI, and what those tokens cost. A token is a unit of text, such as a word, part of a word, or a punctuation mark, that an AI model reads or writes. In English, 1 token is roughly 4 characters. Tokens aren't standardized across models or providers, so the same text can produce a different count in different tools.

Usage reports usually split tokens into a few types:

• Input tokens are everything the model reads for a request. That includes the query, any attached text, and conversation history the model reuses from earlier in the same thread.
• Output tokens are the response the model writes. They often cost more than input tokens. In Perplexity's API pricing, some models charge six times as much for output.
• Reasoning tokens are text a model generates internally to work through a problem before answering. They don't appear in the answer, so a longer visible response doesn't necessarily mean a model used more tokens overall.

Teams using AI tools on a subscription rarely see a per-token charge. Tokens still shape what they get, since heavier tasks use up credits or reach usage limits sooner. Teams that build their own tools on an AI model typically pay per token instead, under usage-based pricing.

How to track token spend

Tracking token spend on a subscription plan

Subscription tools usually don't show token counts.. Instead, many plans use credits, prepaid units that a plan draws down as work gets done. Credits aren't the same as tokens, since one credit can cover text, images, or other work. But tasks that use more tokens usually use more credits too, so they’re the best guide to token spend on a subscription. Users can track their own credit usage, and some plans will also include an admin dashboard that shows credit use for the whole team.

In Perplexity, for example, regular searches don't use credits. Only Perplexity Computer, the assistant that handles multi-step tasks, uses them. People who use Computer can see how many credits they've used and what they used them on. Each Computer task also shows its own credit total in the "Usage" menu of its conversation.

On Enterprise plans, administrators can see how many people use Perplexity and how many queries they run each week in the Insights dashboard (available to organizations with 50 or more seats, or at least one Enterprise Max member).

Tracking token spend in API usage reports

The API provider's usage reports show token counts directly, usually split by type of token. Perplexity's API Platform console, for example, lets users track metrics like input, output, and reasoning token spend on the Billing page.

Why token spend is growing

Agents run many steps per task, and each step is a separate request to a model. One task can add up to hundreds of thousands or even millions of tokens.

Reasoning models are also becoming more common. A reasoning model writes out its thinking before it answers, and every word of that counts as tokens. So the same request usually uses more tokens on a reasoning model than on a standard one, even though the answer on screen looks the same length.

Finally, as more people and workflows use AI in general, token usage adds up. Some of that growth also comes from tokenmaxxing, where engineers run agents mainly to raise their personal token counts and rank higher on internal token leaderboards.

How to reduce token spend

Most high token spend comes from a few kinds of work that use more tokens than they appear to.

Long conversations and agent tasks

Every time a model replies, it usually rereads the whole conversation so far. Each new request includes all previous conversation history plus the current user message. So a short question late in a long thread uses more tokens than the same question in a new one. Agent tasks work the same way, because each step adds its results to the history the next step rereads.

What to do about it

• Start a new conversation when the topic changes. Open it with a two- or three-line summary of anything from the old thread the model still needs.
• Split long agent tasks into smaller ones. Run each part as its own task, and pass along only the results the next part needs.
• Check whether the team's agent tools use context compaction. Compaction summarizes older parts of a task automatically so the model rereads less. If a tool offers it as a setting, turn it on.

The same documents processed again and again

When someone attaches the same long file, such as a policy handbook or a product catalog, to many separate queries, the model reads the whole file each time. Each read counts as input tokens.

What to do about it

• Train users to copy the relevant section of a handbook or catalog into the request instead of attaching the whole file.
• Make a short version of documents the team uses often. A one-page summary of a long policy can stand in for the full document in routine queries.

Retries and loops

When an answer isn't what someone needed and they run the task again, every attempt uses tokens. The same happens when an agent gets stuck repeating the same step.

What to do about it

• Spell out the task, the result, and the sources in the first request. For example: "Summarize the main reasons customers cancelled in Q3, using only the attached survey results, in five bullet points."
• Check the first few steps of an agent task. If the agent repeats the same step or search, stop it and restart with a narrower request.

Large models on simple tasks

More capable models usually cost more per token, so routine work such as summaries or reformatting can use up a team's budget faster than it needs to.

What to do about it

• Use model routing, which matches each task to a model that can handle it. A short rewrite or summary can go to a smaller, cheaper model, and the most capable model is kept for multi-step reasoning or analysis. Perplexity's API takes this approach with its presets, which assign a smaller model to a quick lookup and a larger one to exhaustive analysis.

Spikes after new workflows

A new automated workflow, or a tool rolled out to a whole team, can raise usage overnight. A spike isn't a problem in itself, but it's worth knowing which work caused it.

What to do about it

• Test new workflows with a small group first. Check their usage for a week or two before rolling the workflow out to the whole team.
• Set a monthly budget for each team. Base it on a typical month's usage, so a jump stands out.
• Turn off auto-refill. With auto-refill off, work stops when credits run out, instead of more credits being bought automatically.

How to optimize token spend in Perplexity

For everyday use

Computer uses credits, but it runs only when you choose it, either by selecting the Computer icon, working in a Project which defaults to Computer, or using Comet Work mode.

To go back to regular search, select the Perplexity icon to switch back to Ask mode. To see credit usage, users can open the tracker within the Perplexity interface.

Teams that work in Projects, Perplexity's shared workspaces, can set each Project's default mode to Search or Computer. A Project used mostly for quick research can default to Search, so no one starts a Computer task by accident.

For Enterprise administrators

• Know what each seat includes. Enterprise Pro users receive a monthly allocation of 500 credits, and Enterprise Max users receive 15,000 credits. These expire at the end of each monthly cycle and do not roll over.
• Set a default usage limit first. A usage limit caps how many extra credits a member can draw from the shared pool once their own allocation runs out.
• Create named limits for different kinds of users. Admins can set up limits such as "Standard," "Power user," and "Contractor", and assign them to user groups or individual members. Groups can come from Perplexity or sync from the company's identity provider. If someone belongs to more than one group, the highest limit applies.
• Use personal overrides for exceptions. When one person needs a different cap from their group, a personal override takes precedence over any group or shared limit. That's useful for a short project with heavy research, without raising the cap for a whole team.
• Put a ceiling on auto-refill. Auto-refill tops up the shared pool automatically. It comes with a configurable monthly reload limit, so spending can't grow without a set maximum.
• Decide how members ask for more. The Allow members to request more credits setting lets people who hit their limit request more from an admin, rather than being stopped outright.
• Review usage every month. The Organization credits page shows credits granted, used, and remaining for each member, along with group limits and the pool balance. For organizations with 50 or more seats, or at least one Enterprise Max member, the Insights dashboard adds monthly active users and weekly queries over time. For more detail, audit logs record task execution, connector usage, credit consumption, and credit grants.

For teams building on the API

• Give each tool or team its own API key. Perplexity can show billing for each key separately, and separate keys make it clear which tool is driving usage.
• Check the billing page for which kind of token is growing. It shows input, output, and reasoning tokens over time. Rising input tokens usually mean a tool is sending more text than it needs to. Rising reasoning tokens mean tasks are running on models that think through problems at length.
• Use the Search API when a tool only needs search results. It returns web results without a model writing an answer, and it charges per request with no additional token-based charges.
• Start agent tools from a preset. Perplexity's presets match the size of the task to the size of the model and keep simple requests off the most expensive models

Local-first work

Some work doesn't need a paid cloud model at all. Portable Computer is a version of Perplexity's agent that runs on a person's own computer, using an AI model stored on that machine. Because the model runs locally, routine work doesn't use paid tokens. It reaches out to the web, connected apps, or a stronger cloud model only when a task needs it, and asks the user first.

Explore how Perplexity Computer handles multi-step work across the web, files, and connected tools.

FAQs

What's the difference between token spend and AI costs?

Token spend measures how much AI work a team does, counted in tokens. AI costs cover everything a team pays to use AI, including subscriptions, extra credits, and setup. Token spend drives the usage-based part of that bill.

Do subscription users pay per token?

Usually not directly. Subscription tools track usage in messages, queries, or credits, and tasks that use more tokens run those down faster. On Perplexity, regular searches don't use credits at all; only Computer tasks do.

Why do AI agents use so many tokens?

An agent completes a task in many steps, and each step is a separate request that rereads the work so far. A long task therefore uses more tokens with every step it takes.

Why does token usage spike suddenly?

Most spikes follow a change in how a team works, such as a new automated workflow, a tool rolled out to more people, or a long agent task. Comparing the date of the spike with recent launches usually shows the cause. Testing new workflows with a small group first helps prevent spikes.