AI Cost Drilldown for VS Code

See where your AI spend goes, from a month down to a single tool call, for Claude Code, GitHub Copilot CLI and VS Code Copilot Chat, inside VS Code.

The screens are in the tour: run AI Cost: Show Tour from the Command Palette, or open Help → Get Started → AI Cost Drilldown. They are made-up sample sessions, so nothing in them comes from anyone's real logs.

Contents

• The screens
  • Overview · The sidebar · A month
  • A session · A prompt · A model call · A tool call
• Finding things: filter, sort, sync and back
• Advice
• Copilot tracking and cost tips
• Reading the numbers
• Settings and commands
• Your data and privacy
• What it cannot do

The screens

Inside VS Code the same screens are in the tour: run AI Cost: Show Tour from the Command Palette, or open Help → Get Started → AI Cost Drilldown. The extension's own page in the Extensions view shows this text without the images, because VS Code only loads images there from public https addresses.

Overview

Selecting the AI Cost icon opens the overview in an editor tab. It starts with cost by month, stacked by tool: Claude Code is blue, Copilot CLI orange, VS Code chat green, and a tool keeps its colour on every screen.

• Tiles: total cost, today, number of sessions, and how much of the input came from the prompt cache.
• Cost by month as a chart and as a table (cost per tool, sessions and prompts).
• Where the tokens went: cache reads, cache writes, new input and output, with each one's share.
• By source, by model and the most expensive prompts.
• By month · By day · 7d · 30d · 90d switches the range. By day shows the cost of every day in the loaded history; the 90-day range is drawn by week, and a month shows its days.
• The legend under the chart is the tool selector (Claude Code, Copilot CLI, VS Code chat). Click a tool to leave it out and click it again to bring it back; the last one cannot be switched off. It is the same filter as the Sessions view's Filter → Source, so the chart, tiles, tokens, sources, models, prompts, the Sessions tree and Advice all follow.
• ↻ Sync sessions reads the logs again and shows when it last did.

Almost everything on the page is clickable. A source or model bar filters the Sessions view to it, a month opens that month, a day or week filters to it, and a prompt opens its detail. Hover any bar for its numbers.

The sidebar

The AI Cost icon shows four panels. The first is a compact copy of the overview; its Open the full overview › link opens the tab.

The Sessions title bar has Overview, Sort, Filter, Clear filter (while one is active) and Sync buttons.

A month

Click a month, in the chart or the table, and the whole overview switches to it: its tiles, its cost per day, its sources, models and prompts. A breadcrumb at the top (All months › September 2026) takes you back, and Filter sessions to this month applies the month to the Sessions view.

Copilot premium requests left

Under the tiles, a bar shows how many of your plan's premium requests are left this month: used, left, the reset date, and where the month will end at the current pace, flagged when that is over what the plan includes. The numbers are the ones Copilot CLI writes into its own session log, so they are your plan's real allowance, with nothing to configure. They update when Copilot CLI makes a call; the panel is hidden on plans with unlimited requests, and says so when the last reading is from a month that has ended. The status bar carries it as well (AI today $1.20 · Copilot 6,832 left (89.9%)) and turns to a warning colour when the month is on pace to go over. The panel also sets a daily budget: what is left, spread over the weekdays remaining until the reset (Monday to Friday, today included while it is a weekday), with today's use against it. The status bar shows it as 3/16/day, today's requests over the budget. Set your own number with Daily budget (premium requests) in the Coach panel (empty means automatic, the calculation above), or from a terminal with cost coach budget 50 (cost coach budget auto goes back to automatic; with no number it shows the current budget); the status bar turns to a warning colour when today goes over it. Today's use comes from Copilot CLI sessions. Claude Code has no monthly request quota in its logs, so only Copilot is shown.

Two lines under the budget turn the pace into something to act on:

• What the room left today buys, by model. A model's multiplier is what makes a request cost 1, 3 or 0.33, so the panel says how many prompts the rest of today's budget is on each of the models you use most (17 on claude-sonnet-5.5 (1 a prompt, from your 3 prompts) · 5 on claude-opus-5.5 · unlimited on gpt-5-mini). The per-prompt figure is the median of your own single-model prompts over 30 days, with at least three, so it is what your prompts cost, not a table. With no history yet it shows the usual 0.33x, 1x and 3x tiers instead and points you to /model.
• How the last 7 days ran. The month's pace is an average since the reset, which is slow to notice a change. When the last 7 days' pace would end the month at a clearly different number, the panel says so, and whether that is over what your plan includes. This machine's prompts set the rate and the account's reading sets where it starts, so use on another machine makes it an underestimate.

The model guard

When Copilot's premium requests are on pace to run out (at the last 7 days' pace, or today's use has reached the daily budget), the extension says so once, names the models that cost you the most requests a prompt, and names the cheaper one you already use most (claude-opus-5.5 (3 a prompt) … claude-sonnet-5.5 costs you about 1 a prompt: use it for routine work). The status bar goes to its warning colour and its tooltip carries the same line. It stays on until the month is projected under 95% of the plan, so it does not flicker at the line, and a new billing period starts clean.

It advises and nothing more: it does not hide or switch any model. VS Code keeps which models are hidden in its own profile storage and offers no setting, command or extension API to change it, so hiding is done by you, with the eye icon in Chat: Manage Language Models. An organization that must enforce a list does it in GitHub's Copilot model policy. The notification's button opens the model picker. In the Coach panel: the Model guard switch (on) and the Model guard: costly from limit (2: the requests a prompt from which a model counts as costly).

A session

Click a session in the tree, or a bar in the overview. The panel shows its tool, folder and branch, a cost-per-prompt chart, where its tokens went, a table of models, and every prompt with tokens in and out, cache hit and cost. Click a column header to sort the table; click a row or a bar to open that prompt.

A prompt

A prompt shows its cost, model calls, cache hit and tool calls, the token split, the cost of each model call, a Model calls table with the tools each call used, and every tool call it made.

A breadcrumb with a Back button runs along the top of every detail panel: ‹ Back | Overview › billing-service › Prompt 6. Each part is a link, and Back steps through what you opened, in order.

A model call

A call shows what it cost, its tokens in and out, and the tools it made. Below that it shows what went in (what was new since the previous call: your prompt, or the results of the previous call's tools) and what came out (the reply, the reasoning where it is recorded, and each tool request).

The rest of a call's input was the earlier conversation, almost all of it served from the cache; the token bar shows how much.

A tool call

Click any tool in a Tool calls table to see its input (the full arguments) and its result, with its status, result size and time. This works for both Claude Code and Copilot CLI, including prompts that have no token data.

Finding things: filter, sort, sync and back

• Filter (the funnel in the Sessions title bar) narrows the tree by source, day or date range, models, minimum prompt cost and text (a word in the prompt, session title, session ID, folder or branch). Clicking a chart adds to the filter too, so you can go source → model → day. While a filter is active, a message at the top of the tree says what it is and how many sessions and prompts match, and the funnel next to it clears it. The Advice panel and the overview follow the filter.
• Sort orders sessions, the prompts in them and each prompt's calls together, by time, cost, tokens in, tokens out, cache hit or number of calls. Choose the same key again to reverse it. Anything without a value for the key goes last.
• Sync sessions reads the logs again. It also runs by itself every minute; the manual button shows progress, says what changed (synced 108 sessions (+2), 1177 prompts (+8)) and the overview shows when it last synced.
• Back and the breadcrumb work in every detail panel, as described above.

Advice

The Advice panel measures your own sessions and lists where spend could be cut. Each finding says what was measured, an estimated saving in dollars, and the prompts it comes from (click one to open it).

Savings overlap (a fresh session also removes cache expiries), so they do not add up. Findings that would save less than a few cents are left out.

Token-saving tools that were not used

Under the measured findings, with a tools icon, the panel lists the habits that would have prevented them and were not used. They claim no saving of their own, because what they cost is already in the findings above.

Claude Code's log does not say when a context was compacted, so its sessions are not judged on that.

Copilot tracking and cost tips

Without any setup, Copilot sessions already show credits and dollars per prompt, and prompt tokens, read from Copilot CLI's session log and from VS Code's chat storage. Copilot tracking adds two things that need to run inside Copilot itself:

• Per-call tokens. Copilot's log has credits but no tokens. A small Copilot CLI extension records the tokens of every model call as it happens, which gives the token split, the model-calls table and exact call costs. Prompts it did not see show "No token data": that cannot be recovered afterwards.
• Cost tips. A prompt hook shows one line when a prompt is about to cost more than it needs to, in Copilot CLI and in VS Code's chat.

To turn it on, tick Copilot tracking in the Coach panel, or run AI Cost: Set Up Copilot Tracking. VS Code also offers it once if it finds Copilot on your machine. It asks first and lists exactly what it will write:

Copilot CLI loads extensions only in experimental mode: run /experimental on once, then start a new session. AI Cost: Remove Copilot Tracking deletes the three pieces and keeps your token log and settings. After an extension update the copies are refreshed by themselves.

The cost tips and the five switches

The Coach panel has five switches, saved where the hook reads them (~/.ai-cost-drilldown/config.json):

A switch shows "inactive" when the one it depends on is off, and "set by …" when an environment variable (AI_COST_COACH, AI_COST_COACH_STRICT, AI_COST_COACH_EXPIRY, AI_COST_COACH_REWRITE, AI_COST_MODEL_GUARD) decides it.

Every held prompt's banner says why it was held, gives a Tip on what to change (and a fill-in template for a thin prompt), and says how to run it anyway: send the same prompt again. When a prompt trips several rules at once, one banner lists them all (Also:), so sending it again never runs it past a reason you were not shown. With strict mode off, or in VS Code chat while the extension is not running, each of these is a one-line tip instead: at most one tip per prompt, and each kind once per session. The checks are patterns on the prompt text and the session's own log, run locally; they catch prompts that are plainly thin, not ones that are subtly unclear. Questions, answers ("yes", "continue"), follow-ups that lean on the conversation ("still the same error", "yes, fix the alias bug first") and anything naming a file, identifier, ticket, link, number or quoted text are never flagged as thin, and /compact, /clear and /new are never held. The cost rule needs the token logger that Copilot tracking installs, because that log says how big the session's context is; without it you get the plain cache tip.

If the hook does not run

VS Code and Copilot start the hook with a thin PATH, so a bare node command is not found (the hook fails with node: command not found and every prompt runs unchecked). Setting up Copilot tracking therefore writes the absolute path to your node into the hook file, and an earlier setup that wrote a bare node is repaired the next time the extension starts. AI Cost: Check Copilot Tracking runs the hook with that same thin PATH and tells you whether it works; it is also run right after setup and a few seconds after VS Code starts, and a failure shows an error. Copilot records a failed hook in the session's events.jsonl (hook.end with success: false and the error), which is where to look if prompts are not being held.

A hook that starts but then fails inside lets the prompt run, as a hook must, and writes the error to ~/.ai-cost-drilldown/hook-errors.jsonl. The Copilot tracking row in the Coach panel then says how many prompts ran unchecked in the last day, and AI Cost: Check Copilot Tracking shows the last error.

On Windows nothing documents which shell starts a hook, and they disagree: "C:/…/node.exe" "C:/…/cost.js" hook is a command in cmd.exe and Git Bash but a parse error in PowerShell. So there the hook file names one small file, ~/.ai-cost-drilldown/runtime/hook.cmd, which holds node's full path. Written with forward slashes and no quotes it runs from cmd.exe, PowerShell and Git Bash alike; if your home folder has a space in its name it is run through cmd /d /c "…", which cmd.exe and PowerShell run and Git Bash does not. AI Cost: Check Copilot Tracking starts the hook from cmd.exe and from PowerShell and reports the one that fails. A setup made by 0.2.1 or earlier is rewritten the next time the extension starts.

Strict mode in VS Code chat

VS Code stops a prompt that a hook blocks, but it never shows the hook's reason: the prompt would just vanish. So the hook blocks in VS Code chat only while this extension is running to show why. The extension tells the hook it is there (a small heartbeat file, refreshed every 15 seconds), and each held prompt appears as a notification in the VS Code window you are using:

• Why opens the full banner (the reason, the tip, what to do) with the prompt.
• Send anyway sends the same prompt to the chat again, which runs it. It is the same as sending it again in Copilot CLI.
• Copy prompt copies it, so you can reword it first.

If no VS Code window with the extension is open, or aiCost.notifyHeldPrompts is off, the hook gives VS Code chat a tip instead, as before, so a prompt is never stopped silently. A notification is shown in the focused window only, and only for a minute: a hold you did not see in time is dropped, and sending the same prompt again still runs it.

Cost limit, results and pause

Under the switches the Coach panel has:

• Coach results opens what the coach did over the last 30 days: tips shown, prompts held, and whether you then changed the prompt or sent it anyway, with an estimate of the cost the changes avoided.

  Held is a prompt strict mode stopped. Changed is a different prompt afterwards, or a hold that no longer applied (for example after /compact). Sent anyway is the same prompt sent again. A cold cache is different, because rewording the prompt does not warm it: a different prompt in the same session is the same hold, still open. It becomes Changed when the cache is dealt with (/compact, /clear, /new) or the session is left alone for a day (the work went to a new session), and Sent anyway when a prompt is sent again on the cold cache. The avoided figure counts only held prompts you changed, from list prices: for a cold cache the extra cost of re-sending the context at full price, for a repeat the cost of the earlier run. It is an estimate, not a measured saving.

• Cost limit, Long session (tokens), Long session (prompts), Daily budget (premium requests), Big paste and Model guard: costly from are the thresholds above. Click one to change it; leave the box empty to put the default back.

• Pause strict mode stops prompts being held for 30 minutes, 1 hour, 4 hours or until tomorrow, and resumes by itself. Tips still show while it is paused.

What the coach learns

The coach's rules start from fixed lines. Two things in it are then corrected from what happens on your machine, and Coach results has a What the coach learned section that says what was corrected and from how much evidence.

• The cache lifetime. Five minutes for Claude models and thirty for GPT are what is documented, not necessarily what your plan does. With the token logger on, the hook looks at the idle gaps between a session's model calls and whether the call after each gap read its context from the cache or paid for it again, and picks the lifetime (5, 10, 30 or 60 minutes) that explains them. It overrules the documented figure only with five or more gaps that tell the choices apart and a lifetime that explains at least four in five of them. It is worked out once a day and kept in ~/.ai-cost-drilldown/cache-ttl.json.
• Cold-cache warnings, checked. When you send a prompt anyway after a cold-cache hold, the hook looks at the next model call in the token log: did it start cold, as warned? Coach results counts how many warnings were right and how many were false alarms.

Both need the token logger from Copilot tracking; without it the documented lifetimes are used and nothing is checked.

Moving from the Copilot CLI plugin

Earlier versions were installed as a Copilot CLI plugin (copilot plugin install). The extension replaces it. If the plugin is still installed, setting up Copilot tracking offers to remove it (it runs copilot plugin uninstall ai-cost-drilldown), and until it is gone the plugin's own hook answers prompts, so a prompt is never handled twice. The /cost command typed inside Copilot is no longer provided: the reports are in VS Code.

Reading the numbers

Costs are estimates. Claude Code's cost is each model's list price applied to the tokens in its transcript, scaled to the session total Claude Code recorded when it has one (a ~ marks an estimate). Copilot's cost per prompt is the AI credits Copilot recorded, at $0.01 per credit; each model call is priced at GitHub's published token prices (lib/copilot-prices.json), so the calls of a prompt may not add up to its credits to the cent. A model without a price shows "—". Subscription usage is not billed per token: these are list-price figures for comparing where the usage went.

Cache. Every model call re-sends the whole conversation, normally from a cheap cache. Read from cache is input served from it, written to cache is input stored for next time at a higher price, and new input is neither. A high cache hit means a conversation is being reused well; a prompt that starts with a miss pays full price for all of it.

Sources. Claude Code is read from ~/.claude/projects/, Copilot CLI from ~/.copilot/session-state/, and VS Code Copilot Chat from VS Code's own storage.

Settings and commands

Commands (Command Palette, AI Cost: …): Overview, Sync Sessions, Filter, Clear Filter, Sort, Set Up Copilot Tracking, Remove Copilot Tracking.

Your data and privacy

The extension reads session logs that already exist on your machine, which include your prompt text, folders and branch names. It sends nothing anywhere and needs no account or key. The one exception is Strict model, which sends the held prompt, the session's last few prompts, its last answer, the result of its last failed tool call and the branch name to a small model through your own copilot command.

It writes nothing to those logs. Its own files are in ~/.ai-cost-drilldown/ (a cache of what it read, the coach settings, saved reports, and the token log when tracking is on); AI_COST_HOME changes that folder. Copilot tracking is the only thing that writes outside it, and only after you agree.

What it cannot do

• Recover tokens it did not record. Per-call Copilot tokens exist only for calls made while tracking was on.
• Per-call detail for VS Code chat. VS Code stores one total per request, so its sessions stop at the prompt.
• Show Claude's reasoning. Claude Code's transcripts keep only a signature for it, so the Thinking box is empty.
• Go back past what the tools keep. Claude Code deletes its own transcripts after about 30 days unless cleanupPeriodDays in ~/.claude/settings.json is raised, and only from then on.
• Bill-accurate totals. It reports what the tools logged on this machine: not other machines, and not what is finally invoiced.
• Show Copilot's full replies. Copilot CLI's event log keeps tool requests and a short preview of each reply.
• Coach Claude Code. Claude Code sessions are measured, but the coach (tips and strict mode) acts on Copilot sessions only: Copilot CLI, and Copilot sessions running inside VS Code.
• Hold a prompt it was never shown. The coach acts through a hook. If the hook cannot run, Copilot lets the prompt through, which is why AI Cost: Check Copilot Tracking exists.

Checked in a live session: strict mode holding a prompt in a Copilot session running inside VS Code, with the notification and Send anyway. On Windows the hook is started from cmd.exe and PowerShell by the tests, on every pull request. Not yet checked: the classic Copilot Chat panel, a prompt held by the hook in a live Copilot session on Windows, and the token logger across Copilot CLI updates (Copilot's extensions are experimental).