Why it saves

Where the tokens go, and the four levers that move the bill

Every call to a frontier model re-sends the whole conversation. A long session pays for its own history again and again. This page shows what we measured on a real account, what Workboard's rules found, and what a change in habits is worth at the scale of a company.

96% of tokens are context being re-read

Thirty days, one account, 89,928 API calls, about $7,760 at API list prices. Cache reads were 15.9 billion of 16.5 billion tokens. On average each call re-read about 177,000 tokens of context. Output, the part people think of as the answer, was under half a percent.

15.88BCache reads, 96.1%

0.54BCache writes, 3.3%

0.079BOutput, with thinking, 0.5%

0.013BFresh input, 0.1%

cost ≈ calls × context per call × price per token

Cache reads cost about a tenth of fresh input per token, which is why they are 96% of tokens but a smaller share of dollars. They are still the largest single cost.

What each lever moves

Context per call: instructions, tool and skill listings, file reads, tool output, history. 10% less context is about 10% fewer cache-read tokens.
Calls: loops, retries, failing tools, permission round trips and polling.
Price per token: the model, the effort level, and whether the cache is warm. A cold cache writes at about 1.25 times input; a warm one reads at a tenth.

What the rules found

Eleven rules priced about $1,420 of avoidable spend in 30 days

Most of it came from oversized context. Estimates overlap where one call breaks two rules, so treat the total as a guide. Loops, the same call repeated with nothing changed, are a twelfth finding shown on the Savings page.

Rule	What it detects	Findings	Estimated	Advice it gives
Big context	Calls whose context is over 200,000 tokens	391	$835	Start a fresh session or agent with a short handoff
Cache lifetime	A subagent idle past its five-minute cache, then writing it again	8	$380	Run long waits as a detached job with a watcher; end the agent
Effort	High effort on short tool-only turns of a heavy model	662	$192	Run mostly-tool work at lower effort
Unused MCP	An MCP server loaded and never called	91	$13	Switch it off where it is not used
Idle compaction	Auto-compaction with no prompt in the 30 minutes before	23	$3	End or clear a session you have left
Compactions	Two or more compactions in one session	5	$2	Smaller briefs, or compact earlier
Cold switch	Model switched mid-session with a cold cache	1	$0.65	Pick the model at the start, or switch in a fresh session
Permission wait	Over five minutes waiting on permission prompts	13	-	Allow the commands you always approve
Tool failures	A tool failing over 10% of its calls	6	-	Check how it is called: path, missing tool, flaky command
Cache misses	Over five cache misses in a session, by cause	8	-	Keep tools, model and system prompt fixed in a session
Retries	API errors over 5% of a session's calls	1	-	Look at the errors; overload eases off a busy hour

The levers

Four groups of changes; the first two save the most

1 · Smaller context

Where the measured money is

Fresh sessions with a handoff past 150,000 to 200,000 tokens
Lean always-loaded instructions
Partial file reads
Tool output to files, summaries back
Fewer tool, skill and MCP listings
Subagents for searches and scans

2 · Warm cache

The cheapest to adopt

No sleeping agents: detached jobs with a status file
Stable model, tools and system prompt within a session
Resume within the cache lifetime
Change MCP servers between sessions, never during

3 · Right price per token

A per-task choice

Effort matched to the work: high for design and debugging, low for mechanical runs
Model chosen per task at session start
Fast mode only where waiting is the expensive part

4 · Fewer wasted calls

Hygiene that also saves time

Break loops at three
Fix failing commands once
Allow-list routine commands
Batch independent calls
No polling from agents

A terminal window with lines of build output and progress bars

Tools, skills, MCP servers and agents

Every enabled tool is paid for on every call

Each MCP server, plugin and skill adds its tool schemas or description to every call. An invoked skill's whole body stays in context for the rest of the session and is re-attached after compaction. An agent saves context when it returns a conclusion and ends; it costs the most when it waits, grows or gets resumed. A call at 600,000 tokens of context costs about 15 times a call at 40,000, and a watcher agent that wakes every few minutes pays that on every wake.

Workboard shows the MCP load against calls, the largest tool outputs, agent types by cost and context, and the loops, so the fix is a setting or a habit rather than a guess.

What it is worth to you

The recommendations are worth more than the seats

Enter your annual frontier model spend and the number of developers behind it. The avoidable share defaults to the 18% our rules priced on a real account over 30 days; move it to whatever you find credible.

Annual token spend, USD API-equivalent, across all coding tools Developers with a seat people, since machines take no seat Avoidable share

Avoidable spend a year$9,000,000

Workboard seats a year, at $40 a seat a month$960,000

Net saving a year$8,040,000

Saving for each license dollar9.4x

Illustrative. Measured on one account over 30 days; rules overlap, and the share you reach depends on how far your team adopts the changes. Self-hosted installs are priced separately.

On the roadmap

Four more rules and three features, from our own review

Large tool resultA single tool result over about 20,000 tokens: a whole log, a whole file.
Repeat readThe same file read again with no edit between.
Heavy instructionsAlways-loaded instructions over about 5,000 tokens, priced as calls times size.
Listing weightSkill and MCP listings that load on every call, by size.
Context per call trendThe average context per call by day, with the 200,000 line. It is the number that moves cost most.
Handoff from a Savings rowOne click copies a short handoff for the session flagged for big context.
Org policy packA managed settings file an admin pushes, with the saving tracked after. The seed of TokenGuard.

See your own numbers before you decide

Tell us how many developers use Claude Code and where you would run Workboard. A person at Cognitivi reads every message, replies by email and walks you through the page on your team's data.

Request a demo See pricing