On metered API?
Biggest easy win
Move most people to flat seats, Team or Enterprise. For daily users, a flat seat costs far less than paying per token.
Power users who hit limits get a higher tier or stay on metered.
How to optimize your AI spend
A practical decision tree for seats, coding tools, and the AI your products call. Based on AI spend across 20 tech companies and current model pricing.
Estimate my savings
Four questions. Under a minute. Nothing to share.
Start with the AI usage most relevant to your team.
Choose one path. Each check gives you something concrete to look for in your usage data.
People using AI
Chat, seats, everyday work
Paying for seats nobody uses?
Pull active usage and cut the dormant licenses. Most teams are carrying real dead weight.
Everyone defaulting to the priciest model?
Set a cheaper default for routine drafting and summaries. Keep the top model for genuinely hard work.
Overlapping AI tools across teams?
Inventory every AI subscription by team, consolidate onto one vendor, and cancel the duplicates.
No per person visibility or spend caps?
Turn on per user reporting and hard monthly caps, so one heavy user cannot move the whole total.
Coding tools
Claude Code, Cursor, Copilot
Devs on pure metered API?
Save 35 to 60%
Move daily users onto seats. Steady daily use runs $150 to $250 per developer per month metered, against $100 for a premium seat.
Your heaviest users will still hit their allowance. Turn on extra usage with a per person limit so they keep working without a surprise bill.
Same model for every task?
Save 40 to 70%
Route easy work to cheaper models. Claude Code does this natively through its planning and subagent model settings, and routers like Weave report 40 to 70% on coding agents at Robinhood, PostHog and Reducto.
No per developer caps?
Set them, per tool. Uber burned its entire annual AI budget in four months, with some engineers running $500 to $2,000 a month, then capped at $1,500 per tool.
Code calling AI
Products, internal tools, automations
Large reused prompt prefix, not cached?
Save up to 90%
Largest opportunity to reduce spend in this category
Cache the system prompt and reference documents you send every time. Repeat reads cost a tenth of full price. Identical outputs.
Frontier model doing simple work?
Route classification, extraction, and tagging to a smaller model. Use the cheap model for the easy jobs, and test quality on real traffic before you commit.
Non urgent jobs running live?
Save around 50%
Send reports, backfills, and evaluation runs to the Batch API. Half price when the answer can wait up to a day.
Bloated context on the part that changes every call?
Cut dead conversation history and boilerplate, send only the fields the model actually needs. Caching makes the repeated part cheap, this makes the variable part smaller.
Paying a premium for region locked routing?
Save around 10%
Regional and multi region endpoints cost about 10% more than the global default. Check whether your compliance rules actually require it.
Cut the bill this month. Build something better next.
If your AI spend climbs every month, scattered across models and teams, owned by no one.
We will help you break the cycle.