Keep a way back.
Preserve originals for recovery. Keep the current model when another route fails the requirements.
For agents in your apps and workflows
Your AI traffic.
Less expensive.That’s the job.
Every repeated prompt, oversized model, and retry adds to your bill. Caveman brings caching, compression, and routing together to help your agents cost less to run.
Platform and Router are in private development. Explore the approach below.
Your app or workflow↓
CavemanOptimization layer
01
Reuse repeated context
Cache02
Send less to the model
Compress03
Choose a suitable model
Route↓
Your allowed models
↗Quality checks before wider rollout.Hosted · in development
Customer support · research · internal workflowsLower cost starts with the whole run.
01 / The problem
You pay for the work.
And everything around it.
One user request can trigger many model calls. The bill includes everything the agent reads, writes, repeats, and retries before it delivers an answer.
01
Repeated input
The same instructions and background information get sent again.
02
Expensive defaults
Simple tasks use the same large model as the difficult ones.
03
Failed attempts
Retries and tool loops add cost without adding a useful result.
02 / How Caveman helps
Less waste.
At every step.
There is no single trick. Reuse what repeats, shrink what travels, choose a suitable model, and fix steps that waste money.
Hosted · in development
Prompt caching
Your app sends the same instructions and background information across many calls.
Repeated instructions→
Reuse eligible context→
Cached-input pricing
What Caveman does
Caveman’s hosted optimization layer is designed to reuse eligible prompt content through provider caching, and account for cache state before changing models.
What changes for you
Repeated input can use the provider’s lower cached-input price.
Cache writes, misses, and rebuilding all cost money. The full cost matters, not the hit rate alone.
Explore the Platform03 / The impact
Lower cost.
Per successful task.
Count everything spent on a batch, then divide by the tasks it solved. A lower bill can still mean worse value if too many tasks fail.
Illustrative example · 100 tasks
Original setup
$0.133per successful task$12 total ÷ 90 successful tasks
Proposed change
$0.100per successful task$9 total ÷ 90 successful tasks
Worth testing further
25% lower cost per successful task. The same 90 tasks succeeded.
Example numbers, not measured Caveman performance. Totals include every attempt and retry; changing this control only changes the assumed success count.
What we can show today
The local engine has a published input-token benchmark. The hosted Platform and Router remain in private development. There is no universal production-savings percentage to promise. Your own workload, quality tests, and provider costs decide the result.
Explore the evaluation work04 / What reliable means
The work still
has to work.
Fewer tokens only help when the task gets done. Keep the original details, test the result, and compare the complete cost.
Check the answer.
Compare the same tasks before and after a change. Check correctness and response time alongside cost.
Count the whole bill.
Include prompt overhead, cache charges, recovery calls, and retries. A smaller token count is not a guaranteed bill reduction.
More useful work.
Less wasted spend.
Bring your workflow and your quality bar. Explore where Caveman could lower your costs.
Hosted optimization: in development. Agent SDK: public source available.