The bill gets heavier every month
You default to the most famous model.
The better it does, the more it costs.
68
Everyone uses it, so you picked
GPT-5.6 Sol
Is it the best one for your product too?
That model costs twice as much. Does it earn twice as much?
Find out with ABTO, on your real users.
Same feature, real users split in half
AYour current model
gpt-5.6 sol
BThe model ABTO foundBEST
gemini 3.6 flash
Total paying users
Total paying users
Total paying users
9224 more
Total call cost
Total call cost
$860
Total call cost
$47045% less
The cheaper model still brought more checkouts
Does any of this sound familiar?
You worry a cheaper model will hurt quality
Switching might cost you revenue, so you don't.
And you stay on the expensive model.
Even after switching, you can't tell if it was the right call
You do have a cost chart.
Nothing shows you whether it led to revenue.
The most famous model
may not be the best one for your product
The performance model makers claim is a ranking scored on someone else's data.
The model that works for your users has to be found inside your own product.
The performance model makers claim
A score every company sees the same way.
Your users and your features are not in it.
Criteria
- Coding scores
- Math scores
- Quiz accuracy
The ranking your users rewrite
Ranked by what your users actually did.
A different feature gives a different ranking.
Criteria
- Checkout conversion
- Time spent
- Customer reply rate
Split your users across models,
then see which one does better
You decide which user action counts as success, feature by feature.
For examplesaw the recommendation and completed checkout, saw the draft and finished signing up, saw the answer and kept the subscription
1Users use the feature
2ABTO splits and sends them
3Cost and behavior get recorded
- gpt-5.6 solusers 33%, checkout rate 6.9%Total cost$120
- claude opus 5users 33%, checkout rate 7.1%Total cost$144
- gemini 3.6 flashusers 34%, checkout rate 7.0%Total cost$60BEST
4Same results, so traffic moves to the cheaper model
Checkout rates came out about the same, but gemini 3.6 flash cost half as much
Keep the code you have,
change only the address and the headers
It's OpenAI SDK compatible, so there's no call code to rewrite.
chat.pyOpenAI SDK
import osfrom openai import OpenAI client = OpenAI( base_url="https://gateway.abto.app/v1", # point the address at ABTO api_key=os.environ["ABTO_API_KEY"],) client.chat.completions.create( model=model, messages=messages, extra_headers={ "x-abto-key-openai": os.environ["OPENAI_API_KEY"], # the same OpenAI key you already use "x-abto-device-id": device_id, "x-abto-node-key": "checkout.suggest", },)If you use Claude Code or Codex
The agent finds and instruments your success metrics
It reads your repo and points out the moments worth counting as success, like checkout or signup.
Then it drops the ABTO SDK calls right there, so you never instrument by hand.
Just tell it this one line
/abto add analytics events to our AI features
Just a config swap
The baseURL and a few headers are the whole change. Your call code stays put.
A separate test per feature
Every call carries its feature name in x-abto-node-key, and results are compared separately under that name.
Instrumenting is one line too
Call the skill and it finds the spots worth measuring, and even adds the SDK calls for you.
Teams that tried it first
The changes teams saw after adopting ABTO in beta.
AI resume service
Cover letter draft featureResume review featureUsers preferred different models for drafting and polishing.
They now run a different model for each feature.Model cost
28%saved
Paid conversion
12%higher
AI study-notes service
Problem explanation featureAnswer evaluation featureThey assumed more detailed explanations were better,
but concise explanations and more accurate evaluations drove more return visits.Model cost
41%saved
Return rate
18%higher
Write the next one with us
Reserve now and you can be the first to try it on launch day.
Reserve your spot