5 AI Token-Usage Metrics Every Company Needs To Know in 2027

Most companies we talk to can tell you what they pay for Claude. Far fewer can tell you who actually uses it, what one unit of work costs, or which team turns those tokens into something the business values. That gap widens the moment Claude moves past chat seats into Claude Code, Cowork, and custom agents running on the API.

This guide covers the five AI token-usage metrics we think every company should track in 2027. For each one you get a formula, where Anthropic exposes the data, a starting-point signal for healthy versus warning, and the action the number should drive. If you are an operating partner looking across a portfolio, every metric rolls up by portfolio company, and we flag where the data gets uneven.

Key Takeaways

• Paid seats overstate adoption. Measure monthly active users against assigned seats in every company before you buy more licenses.
• Cost per active user shows who drives spend. Break it out by team and check spend concentration, because a small group of heavy users can account for most of the bill.
• Model mix is the biggest price lever you control. On Anthropic's current price list, Opus 5.5 costs twice what Sonnet 5.5 does per token, and Haiku 5.5 costs one-twentieth of Sonnet 5.5.
• Caching and batch are the cheapest fixes on the API. Cache reads on Opus 5.5 and Sonnet 5.5 bill at 5% of the base input price, and batch jobs bill at half price.
• Cost per completed workflow is the ROI number. Tokens are an input. What an operating partner or CFO can act on is the cost of one finished piece of work.

Why AI Token-Usage Metrics Matter in 2027

AI spend grew faster than the reporting around it. KPMG's Q2 2026 AI Quarterly Pulse, a survey of 204 US leaders at companies with $1 billion or more in revenue, found that 66% have monitoring dashboards but only 26% have full, real-time visibility into what their AI systems cost to run. Only 36% have direct token or usage controls in place. Meanwhile the State of FinOps 2026 survey found 98% of its 1,192 respondents now manage AI spend, up from 31% two years earlier.

So almost everyone is managing it. Very few can see it clearly. That is the problem these five metrics solve.

What a token actually is

A token is the unit Claude reads and writes. It is a chunk of text, usually part of a word. Anthropic's glossary puts one token at roughly 3.5 English characters on older models, and notes that Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text.

Every request has two sides. Input tokens are everything you send: your prompt, attached files, instructions, and the conversation so far. Output tokens are everything Claude writes back, including its extended thinking. On every current model in Anthropic's pricing table, output costs five times as much per token as input. A long conversation also gets more expensive per message, because Claude rereads the whole history each turn.

That is the whole cost model. Which model you use, how much you send, how much comes back, and how much of what you send repeats.

The mental model: Use, Leak, Return

We sort every token report into three questions.

Use. Who is actually using Claude, and how deeply? Metrics 1 and 2 answer this.

Leak. Where is money going out without producing anything? Metrics 3 and 4 answer this.

Return. What does a finished piece of work cost, and is it worth it? Metric 5 answers this, with metric 2 pointing at which teams to look at first.

If a usage dashboard cannot answer all three, it is a bill with charts on it.

Where Anthropic exposes the data

Which source you use depends on how each company buys Claude.

• Claude Team and Enterprise admins get an analytics dashboard and a spend report CSV. Anthropic's usage analytics guide lists views for active members against assigned seats, adoption level, product stickiness, spend concentration, and spend by model. The spend report has one row per user per model, with token counts and net and gross spend.
• Claude Enterprise organizations can pull the same data through the Claude Enterprise Analytics API. It covers per-user daily activity, active-user summaries, and project, skill, and connector usage. Cost and usage reports apply to usage-based Enterprise plans. On seat-based plans they show usage credits only.
• Companies on the Claude Console use the Usage and Cost Admin API. It breaks tokens down by model, workspace, API key, and service tier, and splits uncached input, cache writes, cache reads, and output.
• Claude Code has its own Claude Code Analytics API for Console organizations, with per-user sessions, commits, pull requests, tokens, and estimated cost by model. OpenTelemetry export works on any plan, including Claude through Amazon Bedrock, and lets you tag every data point with a team or cost center.

The 5 AI Token-Usage Metrics at a Glance

1. Active Seat Rate

Active seat rate is the share of paid Claude seats that someone actually used in a period. Anthropic defines an active user as anyone who sent at least one chat message, ran a Claude Code session with tool use or git activity, or ran a Cowork session with tool use or message activity that day.

Seats are a fixed cost. An idle seat is the simplest leak there is, and it compounds across a portfolio. If five portfolio companies each bought Claude Enterprise during the same budget cycle, active seat rate tells you in one number which ones rolled it out and which ones bought licenses and sent a memo.

Look past the headline number too. The dashboard splits activity by product, so you can see whether a company uses Claude only for chat or has moved into Claude Code and Cowork. Daily or weekly actives divided by monthly actives tells you whether usage is a habit or a monthly experiment.

Formula: monthly active users ÷ assigned seats. Track weekly active users ÷ monthly active users alongside it as a stickiness check.

Where to get it: the "Who's using Claude" view in the Team and Enterprise analytics dashboard, or the activity summaries in the Enterprise Analytics API, which return daily, weekly, and monthly active users and seat counts.

Healthy vs. warning: our starting point, which is a PressW suggestion with no industry benchmark behind it: if fewer than about two-thirds of a company's seats are active for two months running, treat it as a rollout problem before you treat it as a budget problem.

How to get started:

1. Pull active members and assigned seats for each company, split by group and product.
2. List every user with no activity in the last 30 days.
3. Reclaim or reassign those seats before the next renewal.
4. Put the lowest-adoption teams through hands-on training built around their own work.

Watch out for: one prompt a month counts as active. A team can post a healthy monthly number while nobody uses Claude for real work. The stickiness ratio catches this. The tech is relatively easy nowadays. It's the people side.

2. Cost per Active User, by Team

Cost per active user is everything a team spends on Claude divided by the number of people on that team who used it. Count seat cost, usage-credit spend above the seat allowance, and any API or Claude Code spend you can attribute to the team.

This is the metric that tells you which teams drive spend, and whether that spend is concentrated. Anthropic's analytics dashboard includes a spend concentration view. When a handful of heavy users drive the bill, that can be your best news or your worst. It depends on what those users produce.

For engineering teams, Anthropic publishes a useful reference point. Its Claude Code cost guide says enterprise deployments average around $13 per developer per active day and $150 to $250 per developer per month, with 90% of users under $30 per active day. Use it as a sanity check for engineering teams. The same page notes that unexpectedly high spend usually traces back to long sessions that were never cleared, or to Opus left as the default model.

Formula: team spend ÷ team active users. For concentration, calculate the share of total spend that comes from the top 10% of users.

Where to get it: the spend report CSV on Team and Enterprise plans, the per-user cost reports in the Enterprise Analytics API on usage-based plans, and the Usage and Cost API grouped by workspace for Console organizations. Set up one workspace per team so the grouping means something. For Claude Code on any plan, add a team.id resource attribute to the OpenTelemetry export.

Healthy vs. warning: our starting point: a team whose cost per active user runs several times the company median with no matching output is a warning. A team with high cost and high output is a playbook to copy.

How to get started:

1. Map every user to a team, using groups in claude.ai or workspaces in the Console.
2. Calculate cost per active user per team for the last full month.
3. Interview the top three spenders on what they do differently.
4. Share what works, and set group caps for spend that produces nothing.

Watch out for: on seat-based plans, usage inside the seat allowance is not metered in dollars. Cost per active user there is mostly seat price divided by actives. Compare usage intensity through the dashboard's usage views instead of dollars.

3. Model Mix

Model mix is the share of spend that goes to each model family: Fable, Opus, Sonnet, and Haiku. It is the single biggest price lever most companies control, and it moves without anyone signing a purchase order.

The spread is wide. On Anthropic's pricing page today, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Sonnet 5.5 costs $2 and $10. Haiku 5.5 costs $0.10 and $0.50 for prompts up to 100,000 tokens. Fable 5.1 costs $10 and $50.

Here is what that means for one routine task with 20,000 input tokens and 2,000 output tokens, before any thinking tokens. On Opus 5.5 it costs about $0.12. On Sonnet 5.5 it costs about $0.06. On Haiku 5.5 it costs about $0.003. Run that task 10,000 times a month and the gap is $1,200 versus $600 versus $30.

Anthropic's own Claude Enterprise consumption guide ranks the families by token intensity. It calls Sonnet the everyday option and reserves Haiku for high-volume, lightweight automation. It also tells admins to audit which models people actually select, because heavy Opus usage suggests the guidance is not landing.

Formula: spend on each model family ÷ total spend.

Where to get it: the spend-by-model view in the analytics dashboard and the model columns in the spend report. On the Console, the Usage API groups tokens by model, and the Cost API returns model in its parsed description fields. The Claude Code Analytics API and OpenTelemetry both break out cost by model.

Healthy vs. warning: warning when high-volume, repetitive work runs on Opus or Fable. Think classification, extraction, routing, or summarizing the same report type every day. Healthy when premium models sit mostly on the work that needs judgment.

How to fix it:

1. Set an organization default model, and pin role defaults where your plan allows it.
2. Lower the effort level for simple tasks. Thinking tokens bill as output.
3. In API workflows, route each step to the cheapest model that passes your evals.
4. In Claude Code, run simple subagents on Haiku.

Watch out for: a cheaper model that fails and retries can cost more than a premium model that finishes on the first pass. Check model mix against metric 5 before you force anything down a tier.

4. Cache Read Share

Cache read share is the share of your input tokens that Claude served from its prompt cache instead of processing from scratch. When the same long instructions, policy documents, or codebase context go out with every request, prompt caching lets Claude reuse them at a steep discount.

The discount is large. Per Anthropic's prompt caching docs, a cache read costs 0.1 times the base input price, and only 0.05 times on Opus 5.5 and Sonnet 5.5. Writing to the cache costs 1.25 times base input for the default five-minute cache, or 2 times for a one-hour cache.

A hypothetical example on Sonnet 5.5. Say a workflow sends the same 50,000-token instruction set 100 times, each call within five minutes of the last. Uncached, that input costs $10.00. Cached, it costs one write at about $0.13 plus 99 reads at half a cent each, about $0.62 in total.

Batch processing is the companion lever. The Message Batches API cuts the price of input and output tokens by 50% for work that can wait, and most batches finish in under an hour. Overnight document processing, evaluations, and bulk enrichment belong there.

Formula: cache-read tokens ÷ (cache-read tokens + cache-write tokens + uncached input tokens). Track batch share of eligible API spend next to it.

Where to get it: the Usage and Cost API reports uncached input, cache creation, and cache read tokens, and you can filter by service tier to see batch traffic. The Claude Code Analytics API reports cache reads and writes per user per model. Inside Claude Code, the /usage command shows what share of a session's input came from cache.

Healthy vs. warning: warning when an API workflow resends the same long context on every call and its cache read share stays low. Your baseline depends on the workload, so set one per workflow and watch the trend.

How to fix it:

1. Put stable content first in every prompt, and variable content last.
2. Add cache breakpoints after the stable sections.
3. Use the one-hour cache for workflows with gaps longer than five minutes.
4. Move every job that can wait an hour to the Batch API.

Watch out for: cache writes cost more than plain input. A cache that gets written and never read raises your bill. Watch cache-write tokens next to reads. On seat-based chat plans Anthropic handles caching for you, so this metric matters most for API workloads and Claude Code.

5. Cost per Completed Workflow

Cost per completed workflow is the total token cost of a workflow divided by the number of runs that finished successfully. "Completed" is a business definition. An invoice processed. A pull request merged. A diligence memo drafted and approved. Retries and failed runs count in the cost.

This is the metric that connects tokens to return. Anthropic's consumption guide works an illustrative example of the logic: $0.90 of cost per run against an estimated $20 of value. Your numbers will differ. The point is that a cost per run you can set next to the value of the work, or the cost of a person doing it, is the only spend number a CFO can actually defend.

For engineering, the Claude Code analytics dashboard on Team and Enterprise plans counts merged pull requests that contain Claude Code-assisted code once you connect GitHub. Spend divided by those PRs is a rough cost per merged change. The Team and Enterprise dashboard also reports skill cost per use, which is a ready-made version of this metric for skills.

Formula: total token cost of the workflow, including retries and failures ÷ successful completions.

Where to get it: give each production workflow its own workspace or API key, then filter the Usage and Cost API by it. For Claude Code, the OpenTelemetry export tags cost by skill and agent name. Count completions from your own systems: the ERP, the ticketing tool, the CRM, or GitHub.

Healthy vs. warning: a rising cost per completion with flat volume signals retries, growing context, or a model change. A falling cost per completion means the workflow is maturing.

How to get started:

1. Pick your three most expensive workflows by spend.
2. Write down what "completed" means for each one.
3. Calculate cost per completion for the last full month.
4. Scale the ones well below their value, fix the ones where retries dominate, and retire the ones nobody uses.

Watch out for: counting starts instead of completions. If a workflow fails 30% of the time, those failed runs are part of the real unit cost.

Metrics We Left Off the List

A few numbers come up often and did not make the five.

Output-to-input ratio. Output costs five times input, so a workflow that writes far more than it reads deserves a look. We treat it as a diagnostic inside metrics 3 and 5.

Tool acceptance rate. Claude Code reports how often developers accept its edits. It is a quality signal for engineering, and Anthropic's Claude Code Analytics API calculates it per tool.

Total tokens. Raw totals rise with adoption. On their own they mostly measure that people use the product.

How to Roll Out AI Token-Usage Tracking Across a Portfolio

A 90-day plan works for a single company or a full portfolio. For a portfolio, run it in each company and roll the five numbers into one monthly view for the operating partner.

Days 1 to 30: inventory and baseline

1. List every Claude surface each company uses: Team or Enterprise seats, Console organizations, Claude Code, and Claude through Amazon Bedrock or Google Cloud.
2. On Enterprise, have the primary owner create an Analytics API key. On the Console, have an admin create an Admin API key.
3. Map people to teams through claude.ai groups, Console workspaces, or OpenTelemetry resource attributes.
4. Record a baseline for all five metrics.

Days 31 to 60: guardrails

1. Set default models and effort levels by role.
2. Set group spend caps and alerts. The consumption guide describes organization alerts at 75% and 90%.
3. Set workspace spend limits in the Console.
4. Fix caching on the top three API workloads by spend, and move non-urgent jobs to batch.
5. Reclaim idle seats.

Days 61 to 90: tie spend to outcomes

1. Define "completed" for every production workflow.
2. Report cost per completed workflow monthly.
3. Hold a monthly review with each company's lead and the operating partner.

Two data notes before you build. Enterprise Analytics API data starts on 1 January 2026, and cost figures can change until about seven days after a month closes, so use closed months for anything finance will rely on. Usage through Amazon Bedrock does not appear in Anthropic's analytics APIs, so companies on Bedrock need OpenTelemetry, a gateway, or their cloud billing console.

Where PressW Fits

Tarun describes PressW this way: "We're kind of like an IT MSP, but for AI." Tracking these five metrics month after month, across every company, is exactly that kind of work.

Token & Usage Analytics and Token Optimization are part of our Managed AI Services. We pull the usage and cost data from every Claude surface a company runs, report it through Argos, our platform for transparent reporting, usage analytics, and adoption metrics, and then act on it. That means model routing, caching, batch moves, and seat right-sizing through Claude License Management. Our forward-deployed engineers are 100% Anthropic-certified, and we are an Official Anthropic Partner with SOC 2 Type II certification.

Keep in Mind: we're a boutique. If you need one global program across dozens of geographies on day one, a large consultancy may fit that shape better.

FAQ

What is a token in AI, and why does it matter for cost?

A token is a small chunk of text, usually part of a word. Claude bills per token for everything it reads and writes. On current Claude models, output tokens cost five times as much as input tokens. Long conversations cost more per message because Claude rereads the history each turn.

How do I see Claude usage by user in Claude Enterprise?

Admins can open the analytics dashboard in claude.ai or export the spend report CSV, which lists usage per user and per model. For automated reporting, the primary owner can create an Analytics API key and use the Claude Enterprise Analytics API. Per-user cost reports apply to usage-based Enterprise plans.

How much does Claude Code cost per developer?

Anthropic says enterprise deployments average around $13 per developer per active day and $150 to $250 per developer per month. It also says 90% of users stay under $30 per active day. Your cost depends on model choice, codebase size, and habits like clearing sessions between tasks.

How much does prompt caching save on Claude?

Cache reads cost 10% of the base input price on most models, and 5% on Opus 5.5 and Sonnet 5.5. Cache writes cost more than normal input, so caching saves money only when the cached content gets reused. Workflows that resend the same long instructions benefit the most.

How should a private equity firm track AI spend across portfolio companies?

Track the same five metrics in every company and roll them into one monthly view. Start with active seat rate and cost per active user, because they need the least setup. Then add model mix, cache read share, and cost per completed workflow as each company moves onto the API or Claude Code.

What is a good Claude seat adoption rate?

No published industry benchmark exists for Claude seat adoption. Our starting point is that a company with fewer than about two-thirds of seats active for two months running has a rollout problem. Set your own baseline and track it monthly.

Final Thoughts on AI Token-Usage Metrics

We do not think anyone has token economics fully figured out yet, including us. Anthropic's prices, tokenizers, and analytics tools keep moving. Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text, and Sonnet 5's introductory price became its standard price. Any threshold in this guide will need adjusting. The five AI token-usage metrics will still be the right questions: who uses it, where it leaks, and what it returns.

Start with one company and one month of data. Expect the first report to be the most useful one, because it is the first time anyone sees the whole picture. If you want help building that report, or running it across a portfolio, reach out to our team.

Last reviewed: 23 October 2026. Product, pricing, and analytics details are drawn from public sources as of this date and are linked throughout. Anthropic updates its pricing and admin tools often, so check the linked pages before you build on them.