Claude API Cost — Plain-English Report
What you're actually paying for, who's spending it, and where the money is going. Covers all of July 2026 and August 8–17, 2026. Figures are actual billed USD.
July Total
$51,868
full month
Aug 8–17 Total
$8,055
10 days
Aug Run-Rate
$24,166
projected full month
Avg / Day (Aug)
$806
≈ $5,639/wk
?The one thing to understand first
Every bill line is just tokens × price. Tokens come in 4 flavors, and they cost wildly different amounts. The flavor that's eating your budget is "cache write."
80% of July's entire bill ($41,439 of $51,868) was "cache write." That's the cost of re-sending the same big context (system prompts, memory files, project docs) over and over so the model can "remember" it for 5 minutes. It is the single biggest lever you have to cut cost.
- input_cache_write_5m — the expensive one (80%)
- Writing a fresh copy of your context into a 5-minute cache. Charged at ~1.25× normal input. Happens whenever a session's big preamble isn't already cached (new session, long gap, or context that keeps changing).
- input_cache_read — the cheap reuse (16%)
- Reading from that cache on the next message. ~10× cheaper than a write. High read vs. write is GOOD — it means caching is working.
- output — the model's replies (4%)
- Text the model generates. Surprisingly small here — you are NOT paying mainly for answers, you're paying for context.
- input_no_cache — uncached input (<1%)
- Negligible. Fine.
$Where July's money went — by token type
| Token type | Cost | Share | |
|---|
| input_cache_write_5m | $41,439 | 79.9% |
|
| input_cache_read | $8,400 | 16.2% |
|
| output | $2,002 | 3.9% |
|
| input_no_cache | $27 | 0.1% |
|
Read this as: for every $1 of useful "cache read" reuse, you spent about $5 re-writing context. Lots of short-lived or fresh sessions = lots of cache writes. Fewer, longer, steadier sessions = the writes get amortized into cheap reads.
#Who spent it — by API key / workspace (July)
| Workspace / key | Cost | Share | |
|---|
| UMS Management | $14,482 | 27.9% |
|
| UMS/VGP ADMIN KEY | $12,481 | 24.1% |
|
| VGP HR | $12,175 | 23.5% |
|
| VGP SALES | $3,923 | 7.6% |
|
| UMS ANALYTICS | $3,347 | 6.5% |
|
| VGP MANAGEMENT | $1,998 | 3.9% |
|
| UMS INVENTORY | $1,137 | 2.2% |
|
| UMS MARKETING | $762 | 1.5% |
|
| UMS ACCOUNTING | $498 | 1.0% |
|
| UMS E-Commerce | $267 | 0.5% |
|
| UMS SALES | $253 | 0.5% |
|
| UMS PURCHASING | $162 | 0.3% |
|
| UMS DISPATCH | $128 | 0.2% |
|
| UMS CREDIT REPAIR | $126 | 0.2% |
|
| UMS WAREHOUSE | $89 | 0.2% |
|
| VGP PURCHASING | $29 | 0.1% |
|
| VGP INVENTORY | $8 | 0.0% |
|
| api key | $2 | 0.0% |
|
"UMS/VGP ADMIN KEY" is a shared key not tied to one department — worth splitting out so each team owns its spend.
UMSUMS vs VGP vs Shared (July)
| Group | Cost | Share | |
|---|
| UMS | $21,255 | 41.0% |
|
| VGP | $18,132 | 35.0% |
|
| Shared Admin | $12,481 | 24.1% |
|
AugTop spenders (Aug 8–17)
| Workspace / key | Cost | Share | |
|---|
| VGP HR | $2,861 | 35.5% |
|
| UMS Management | $1,487 | 18.5% |
|
| VGP MANAGEMENT | $1,134 | 14.1% |
|
| UMS ANALYTICS | $663 | 8.2% |
|
| VGP SALES | $589 | 7.3% |
|
| UMS MARKETING | $574 | 7.1% |
|
| UMS SALES | $238 | 3.0% |
|
| UMS E-Commerce | $182 | 2.3% |
|
| UMS INVENTORY | $158 | 2.0% |
|
| UMS ACCOUNTING | $83 | 1.0% |
|
| UMS PURCHASING | $37 | 0.5% |
|
| UMS CREDIT REPAIR | $21 | 0.3% |
|
| UMS WAREHOUSE | $15 | 0.2% |
|
| UMS DISPATCH | $10 | 0.1% |
|
| VGP PURCHASING | $3 | 0.0% |
|
| console | $0 | 0.0% |
|
~Daily spend
July (each bar = one day; hover for the amount)
August 8–17
July averaged $1,673/day. August 8–17 is averaging $806/day — below July pace, which projects to about $24,166 for the full month if the pace holds.
!What this means & how to cut it
- Model: ~99.9% of spend is Claude Opus 4.8 — the priciest model. Anything routine (summaries, memory writes, simple lookups) run on Sonnet instead; Sonnet was <$40 all July for a reason.
- Cache writes are the whole game. They're 80% of the bill. They spike from: fresh sessions, long idle gaps (the 5-min cache expires), and ever-changing context (bloated MEMORY.md / project files that invalidate the cache every turn). Keep preambles lean and stable.
- Biggest line items: UMS Management, the shared Admin key, and VGP HR together are ~75% of July. Those are where a 20% reduction actually moves the number.
- Shared key: "UMS/VGP ADMIN KEY" ($12,481 in July) isn't attributed to a team. Give each department its own key so costs are traceable.
Rule of thumb: to cut this bill, don't focus on shorter answers (output is only 4%). Focus on keeping context small & stable and reusing sessions so you pay for one cache write and many cheap reads — and move non-critical work off Opus.
Source files: claude_api_cost_2026_07_01_to_2026_07_31.csv & ..._2026_08_01_to_2026_08_17.csv (filtered Aug 8–17). Totals are the sum of the cost_usd column. Generated 2026-08-17 for EJ.