Save what your employees spend on Claude.
Not every send needs the most advanced model. After messages are in the thread, Lightpass suggests a lighter model for the next send when that is enough. Same work, without compromising quality.
Request a two-week pilot
In the chat
Every conversation. A suggestion for the next send.
- The thread is already thereAfter a message is in the thread, we classify it. Every chat. Every send that is already on screen.
- A lighter next send, when that is enoughIf the work did not need the most advanced model, we suggest a lighter one for the next send.
- They still pickThe person sending still chooses the model for the next send.
Typical overkill
A marketing email does not need the most advanced model.
People reach for the top model out of habit. On a same-family top-to-mid swap, token cost is about 60% less at Anthropic list prices, August 2026. Same tokens, lower rate. That 60% is this pair only. Example math, not a customer.
Marketing email
A campaign draft on the top model instead of the mid model.
About 60% less token cost on that send.
List price, top to mid. Example, not a customer.
Summary or status update
A recap on the top model instead of the mid model.
About 60% less token cost on that send.
List price, top to mid. Example, not a customer.
Rewrite or tone pass
A tone pass on the top model instead of the mid model.
About 60% less token cost on that send.
List price, top to mid. Example, not a customer.
Light coding
A small helper, a regex, a one-file script on the top model instead of the mid model.
About 60% less token cost on that send.
List price, top to mid. Example, not a customer.
How we suggest
Three things. This order.
We only put dollars on a model swap in the same family, same tokenizer, at list. Effort and extra thinking can still be worth taking. We do not turn those into a dollar on this page.
- 1 Effort Claude starts High. We suggest lower when the thread did not need it. The lightest model has no effort knob, only extended thinking.
- 2 Model A lighter model in the same family when the work matches. This is the only line we dollarize, and only at list for that pair.
- 3 Thinking Extra thinking uses more tokens. We treat that as about 1.5×, as an assumption. We do not put dollars on it.
If a lighter model keeps needing more back-and-forth, we stop suggesting it. Ten or more samples, and it is not holding. That vetoes the suggestion. It does not invent a new one. We do not “learn preferences.” We do not get smarter every message.
Where it shows up
The bill that can move is extra-usage.
Lighter next sends on the work that does not need the top model. Potential if they take the suggestion. Not saved. The number that can show up on the invoice is extra-usage, and sometimes the seat.
- Lighter seatsCan mean a cheaper seat if they do not need the high plan.
- Stay inside the poolCan mean they stay inside included usage, so they do not buy extra credits or upgrade.
- Extra-usage invoiceCan mean a lower invoice if they already pay for extra usage.
- More of the same workThe same Claude budget covers more of the same work. They still pick the model.
Your plan
Extra-usage at your seats.
Claude seats, and dollars per seat per month. If extra-usage falls, this is the annual line.
If extra-usage falls
$72,000 / year
Now
$120,000
After
$48,000
Seats × dollars × 12. If the extra-usage invoice actually falls. They still pick the model. Example, not a customer. Not saved.
The dashboard
One view across every employee.
- OrgPotential if people take the next-send suggestion. Aggregate. Not a saved number.
- ActionsWhich kinds of work that spend sits on. Categories and totals. Not the conversations.
- PeopleWho that potential sits on, on which work, with which model, if they take the next-send suggestion.
Prompt text is classified once and not stored. Managers see categories and totals, never the conversation. When Claude ships a new model or changes prices, we retune.
The pilot
Two weeks, one team. Zip install on claude.ai.
Request a two-week pilot