Over a 76-day window, a 52-person org in our sample report spent $39,382 across 295,152 requests on Cursor, the Anthropic API, and Claude Team combined — $25,953 of that broken out by model family. Two numbers inside that $25,953 point in opposite directions, and the gap between them is the whole argument.
| Model family | Share of requests | Share of attributable dollars |
|---|---|---|
| Claude Sonnet | 36.5% | 29.4% |
| Claude Opus | 33.0% | 42.4% |
The model everyone uses most isn't the model most of the money goes to
Claude Sonnet is this org's most-used model family, and not by a wide margin — 36.5% of every model-attributed request (requests where a specific model, not the "auto" router, handled the call) against Claude Opus's 33.0%. But it accounts for only 29.4% of the $25,953 attributable to a model family. Claude Opus, just 3.5 percentage points behind on requests, accounts for 42.4% of that same $25,953 — the single largest dollar share of any model family, on slightly fewer requests than Sonnet.
That gap is the tell. A model that's cheap enough to reach for on routine work by default (Sonnet) is carrying most of the request volume for a modest fraction of the bill. A model priced for hard reasoning (Opus) is carrying the largest dollar share despite being used slightly less. Somewhere in that 42%, some of that work didn't need Opus — it just defaulted to it.
One real example
The report's own coaching draft for one person — call them Person B, as the report does — puts a number on it: 94% of Person B's $1,451 in Cursor spend for the period was Opus. Re-priced at Sonnet, the same work is estimated at roughly $820 less for that period — about 60% less. Re-priced at Haiku, roughly $1,093 less. The report's own assessment isn't "stop using Opus" — it's specific: "Opus is the right call for hard reasoning; Sonnet handles routine work just as well." Person B is clearly a power user getting real work done — the fix isn't less usage, it's matching the model to which kind of request it is.
What it adds up to across the org
The same tier-shift analysis run across everyone in the report — every person whose usage skews toward a premium model for work that doesn't need it — estimates $6,053 in savings for the period, roughly 15% of total spend, spread across 33 tier-shift recommendations covering 31 of the 52 people (a few people show up on more than one vendor). None of it requires using AI less. It requires picking Composer or Sonnet for everyday edits, saving Opus for the changes that are actually hard, and reaching for the top-tier model deliberately, not by default.
One caveat the report itself states on every one of these recommendations, worth repeating here: savings are a list-price estimate — usage re-priced at published per-token rates, which differ from seat-allocated billing. Treat $820 and $6,053 as "roughly this much."
Any single Opus call over Sonnet is a few cents' difference — genuinely too small to notice, and not worth the friction of switching models mid-task. That's exactly why the habit survives: no individual decision ever looks wrong. What compounds is 52 people making that same "doesn't matter" call thousands of times over a quarter. The unit economics only look bad in aggregate, which is precisely the view a per-request receipt never gives you.
The actual rule
Match the model to the task, not to habit: a fast, cheap model for everyday edits and routine questions; the expensive one reserved for work that's genuinely hard — architecture decisions, gnarly bugs, anything where getting it right the first time is worth more than the token cost. "Always reach for the best" is a reasonable rule when you're one person paying out of pocket. At fleet scale, it's a 15% tax nobody voted for.