Heavyweight by default
Using your most capable model for every call is the most expensive habit there is.
Cost optimization
Spend less, ship the same
Most inference spend is avoidable. Use the right model for each job, stop paying for repeats, cap output length and make retries safe. Every lever is in your hands.
Right-sized models
for every task
Free repeats
cache hits cost nothing
No double metering
with idempotency keys
The problem
What is slowing you down today
Paying for the same answer
Repeated prompts billed again and again.
Responses longer than needed
Unbounded output means unbounded cost.
Capabilities
Everything you need, nothing you don't
Route by capability
Send a capability instead of a model name, and the router picks the lowest-priced enabled model that can do it.
Compare before you call
Sort the model catalog by price and test candidates in the playground.
Cap max_tokens
Set max_tokens to the longest useful answer. You pay for output tokens produced.
Design for cache hits
Identical prompt, model and temperature inside the cache window is a free cache hit. Keep prompts stable.
Idempotent retries
An Idempotency-Key on every request means retries are metered once.
Watch the dashboard
Daily and month-to-date usage show where spend is growing before the invoice does.
Use cases
What teams run on it
01
Classification
Lower-priced chat models with temperature 0 and short max_tokens.
02
FAQ answers
Stable prompts that hit the cache.
03
Batch enrichment
Idempotency keys per item and paced workers.
04
Mixed agent steps
Reasoning only where reasoning is needed.
How it works
From sign-up to production
- 1
Measure
Use per-response price and the usage dashboard to find your biggest spenders.
- 2
Right-size
Move routine calls to capability routing or lower-priced models.
- 3
Trim and reuse
Cap max_tokens and stabilise prompts for cache hits.
Control
Built so you stay in charge
Keys you control
API keys are shown once at creation and stored only as a hash. Revoke any key instantly from Settings.
Spend you can see
Every request carries its own receipt: tokens in, tokens out, latency and price. Nothing is billed that you cannot trace.
Limits you set
Per-account rate limits and a free monthly allowance keep experiments safe before they become production.
Data you decide on
You choose exactly what you send. Repeat prompts can be served from your cache, and cache hits are never billed.
For developers
One endpoint. Every model.
Send a prompt and a capability, or name a model. One bearer key, one small JSON body, and a full receipt on every response.
curl https://YOUR-DOMAIN/api/ainf/inference \
-H "Authorization: Bearer $AI_INFERENCE_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: order-1042" \
-d '{"prompt":"Summarise this ticket in one line","capability":"chat"}'Questions, answered
What is the single biggest saving?
Usually sending routine work to a lower-priced model by requesting a capability rather than pinning your most capable model.
Are cached responses really free?
Yes. Cache hits report a price of zero.
Do failed attempts cost anything?
No. You pay for calls that succeed, at the price of the model that served them.
How do I know what a change saved?
Compare price_micros on responses and daily totals in the usage dashboard before and after.
Find your first saving today
Open the usage dashboard and the model catalog side by side.