Using LevelFour
AI
What your company spends on AI tools, who uses them, and which Claude Team seats to change. Split the bill by area from a spreadsheet of email and area.
Open AI in the dashboard sidebar. It appears once an AI vendor reports usage to LevelFour: Claude Code and Cowork, Codex or DeepSeek. Vendor, at the top of the page, switches between Claude, Codex and DeepSeek.
Every vendor shows what its tools report: usage value, active people, sessions, models, and each person's tokens and sessions. Usage value prices that usage at API list rates. It is an estimate, not a bill.
On a Claude Team plan, two files you export from claude.ai add what telemetry cannot see: the seat each person holds, Chat, the people who never used their seat, and the bill itself. Upload them with Upload files, as Claude Team files describes.
The tabs
| Tab | What it shows |
|---|---|
| Overview | Health: people with failed requests, rate limited, idle seats, people behind on version, MCP servers failing and people without managed settings. Then seats, the AI bill and savings, seat use, models, seats against active and engaged people, reliability by day, this month's bill and its estimate, the top people and the current risks |
| Cost | The bill, usage value, the month's estimate and the bill per active person, then Cost over time, Cost by area, models with what a million tokens cost on each, tokens and the cache, how use spreads and who drives it |
| People | Everyone with a seat or reporting usage, with their area, seat, product, tokens, extra usage, failed requests, Claude Code version, last active day and the recommendation for their seat |
| Outcome | Pull requests, commits, lines changed and usage value per pull request, then Output over time, output by area, lines of code, edit decisions by language, usage per pull request by person, and repositories |
| Fleet | Failed requests by what the API answered, the people they hit most, response times by model, tool success, Claude Code versions, MCP servers, managed settings, tool approvals and permission modes, plugins and skills, where requests come from, effort, terminals and operating systems |
Without the Claude Team files, Overview and Cost show what telemetry reports: usage value, active people, sessions and cost per session, and a line at the top says what the files add.
Every chart opens the next level down
- Legends: click a name in any legend to hide or show it, and the totals beside the chart follow.
- Chart controls: switch any chart over time between Daily and Cumulative, and between bar and line.
- Split by: on Cost over time, pick Show (usage value or the bill) and Split by (product, model, area, seat tier, seats or extra usage, or person). Output over time does the same for pull requests, commits and lines, split by area, person or repository. The table under each chart reads the same split.
- Drilling down: click a row to open its next level down: an area opens its people, a person opens their page, and a session opens its detail. Click a day on a chart to narrow the table to that day. Each filter shows as a chip above the chart, and removing the chip removes the filter.
A person and their sessions
On People, filter by state (failed requests, rate limited, behind on version, no managed settings, idle seat, upgrade or downgrade), area, seat, product and recommendation, and click Export for a CSV of the list. Click a person to open their page:
- Overview: what their seat costs, tokens per week by product, failed requests by day, their models, and their tools and MCP servers with how many calls failed. Review, when a recommendation names the seat, opens it in Savings.
- Sessions: each session with its product, model, repository, tokens, usage value and how it ended: with no failures, recovered after one, or stopped on errors. Click a session to open its detail: tokens by type, requests and failures every 10 minutes, and the tools it called.
LevelFour never keeps what a session said. A session from before November 10, 2026 shows its totals and no detail, because the detail was not kept yet.
Claude Team files
Anthropic offers Team plans no API for seats or spend, so LevelFour reads two CSV files from claude.ai. Upload a fresh pair every week: the page header says when the last files came and when the next upload is due.
| File | What it adds | Who can export it |
|---|---|---|
| Members | Who holds which seat, including people who never used it | An Admin or above |
| Spend report | Tokens and extra usage per person, in every product | An Owner or Primary Owner |
Export the Members file
In claude.ai, go to Organization settings > Members and click Export CSV.
Export the spend report
In claude.ai, go to Settings > Analytics. Under How much is Claude costing?, click Export spend report, pick MTD, Last Month, Last 90 Days or Custom, then click Download.
Upload both
In LevelFour, open AI and click Upload files. Upload the Members file, then the spend report, and set the days the report covers. MTD runs from the first of the month through yesterday.
Check what LevelFour found, then save
Check what we found lists the seats, the people in the spend report, the extra usage, the products and what the report covers. Pick your Billing cycle, Monthly or Annual, and click Save.
The bill is the seats plus extra usage. LevelFour prices each seat at Anthropic's US list price for your billing cycle: Standard at $25 a month on monthly billing or $20 on annual, Premium at $125 or $100. Extra usage is the net spend the report names for each person.
A spend report that covers only the usage above the seats cannot tell a seat nobody used from one used in Chat inside its allowance. On such a report, LevelFour raises no Remove seat, reads Claude Code and Cowork telemetry for the token rule, and says so on the page.
Seat recommendations
LevelFour checks every seat against three rules after each upload, and raises a recommendation in Savings for each seat that meets one. A seat meets one rule at most, removal first.
| Recommendation | When | Saves a month, up to |
|---|---|---|
| Remove seat | No use in any product for more than two weeks in a row | The seat's price |
| Upgrade to Premium | A Standard seat's extra usage in the last 30 days passed the price gap between Premium and Standard: $100 on monthly billing, $80 on annual | The extra usage minus the gap |
| Evaluate downgrade | A Premium seat used fewer than 200,000 tokens in the last 7 days | The price gap |
Each rule is one recommendation in Savings, with one row per person under it and Anthropic's own steps under Manual Setup. LevelFour changes nothing in your Claude organization.
When you upload the next Members file, LevelFour compares each open recommendation with it. A seat now on Premium, a seat now on Standard, or a person no longer in the list marks that recommendation as applied, with the evidence in Status & Logs. Anthropic bills a downgrade from the plan's next renewal. Removing a member frees the seat but does not lower the plan's seat count, so the bill drops only once you lower the count too.
Cost by area
Split the bill by cost center from a spreadsheet you already keep.
- Click Upload spreadsheet at the top of the page.
- Upload a CSV with two columns,
emailandarea, of 1 MB at most, and click Review. - Check who matched shows how many rows matched a person, the people left without an area and the cost a month they leave unassigned, and the rows that match nobody in your organization. Nothing is saved until you click Save areas.
Saving creates a virtual tag called Area, with one value per area matching its people by email on Anthropic, OpenAI and DeepSeek. Boards, Breakdown and Tags then split the same bill by area, and so do the Cost, People and Outcome tabs here. Upload the spreadsheet again whenever people change teams: the new file replaces the old one. A spreadsheet can hold up to 100 areas.
Costs and Usage
One page per connected provider, with spend so far this month, spend over time, and a breakdown you can group by service, account, tag or virtual tag.
Savings Recommendations
Review the savings LevelFour finds on your connected accounts, decide on each one, and choose which resources it covers before it runs.