reflexer · reflexer.ai

You are paying frontier prices
for tasks that aren’t frontier work.

1 / 12

We are all paying way too much in tokens right now.

Look at your last month of traffic.

The same jobs come back again and again — greetings, confirmations, a field pulled out of a message, a decision from a short list.

Every one of them was paid at frontier prices, as if it were the first time.

Your requests by volume, most frequent first: a short head of the requests your customers make all day, then a long tail where every request is different."hi!"get_calendar()"price?"extract_address()write_to_crm()the tail: each one differentyour requests, most frequent first →height = volume
Your requests by volume, most frequent first: a short head of the requests your customers make all day, then a long tail where every request is different."hi!"get_calendar()"price?"extract_address()the tail: each one differentyour requests, most frequent first →height = volume

2 / 12

Repetition should be cheaper.

Humans get faster at what they repeat. After a hundred times it costs almost nothing.

A frontier model does not work that way. Providers discount the repeated part of the prompt, but the answer itself is made from zero on every call — the first time and the thousandth.

Intelligence used on one job as it repeats. A frontier model stays flat: the thousandth answer costs what the first did. A human falls along an S curve and by the hundredth time the job costs almost nothing.intelligence used— which is what the call costs youa frontier model: the same every timea human: it stops needing thought1st10th100th1000thtimes you have done the same job
Intelligence used on one job as it repeats. A frontier model stays flat: the thousandth answer costs what the first did. A human falls along an S curve and by the hundredth time the job costs almost nothing.intelligence used— which is what the call costs youa frontier model: the same every timea human:it stops needing thought1st10th100th1000thtimes you have done the same job

3 / 12

How repetition looks in real traffic

Your assistant says hello by calling a frontier model and having it write the sentence out, start to finish, in every new chat.

And where it applies a rule, it composes a paragraph explaining the outcome, in new words each time, for a decision the policy had already made.

Jobs that come back in the traffic, with the output each one returns: in text generation; in tool calling.in text generationcomposed again for every new chat{"output": "Hi! You’ve reached Solano Air billing. I can help with charges, refunds and receipts — what do you need?"}in tool callingthe same rule, argued from scratch each time{"name": "refund_decision", "args": {"eligible": false, "policy": "same_day_rule", "rationale": "Booked six days ago, outside the same-day window, on a non-refundable fare, so only a credit applies."}}
Jobs that come back in the traffic, with the output each one returns: in text generation; in tool calling.in text generationcomposed again for every new chat{"output": "Hi! You’ve reached Solano Air billing. I can help with charges, refunds and receipts — what do you need?"}in tool callingthe same rule, argued from scratch each time{"name": "refund_decision", "args": {"eligible": false, "policy": "same_day_rule", "rationale": "Booked six days ago, outside the same-day window, on a non-refundable fare, so only a credit applies."}}

4 / 12

reflexer answers the routine work.

reflexer sits in front of the frontier provider you already call.

The routine requests are answered by your model, which we build from your own traffic: the same answer your provider would have given, faster, at half of the price they would have charged.

When a request needs frontier intelligence, it goes to your frontier provider, on your key. We charge only for the requests we answer.

Your app calls reflexer. A short loop returns the answer from your model. A longer loop reaches your provider for what needs frontier intelligence.your appunchangedreflexeryour modelyour provideryour keyrequestansweredneeds frontier intelligenceyour provider’s answer
Your app calls reflexer. A short loop returns the answer from your model. A longer loop reaches your provider for what needs frontier intelligence.your appunchangedreflexeryour modelyour provideryour keyrequestansweredneeds frontier intelligenceyour provider’s answer
Keep your frontier intelligence, add reflexes.

5 / 12

No work lands on your team.

reflexer improves while your product runs, on your traffic, without anyone on your side doing anything about it.

Change a prompt whenever you like. Those requests go back to your frontier provider, and reflexer learns the new answers.

No data preparation, no training schedule, no review of your model on your side.

Your app calls reflexer. A short loop returns the answer from your model. A longer loop reaches your provider for what needs frontier intelligence. Your provider’s answers feed the model that answers next time.your appunchangedreflexeryour modelyour provideryour keyrequestansweredneeds frontier intelligenceyour provider’s answeryour model learns from what your provider answered
Your app calls reflexer. A short loop returns the answer from your model. A longer loop reaches your provider for what needs frontier intelligence. Your provider’s answers feed the model that answers next time.your appunchangedreflexeryour modelyour provideryour keyrequestansweredneeds frontier intelligenceyour provider’s answeryour model learns from whatyour provider answered

6 / 12

One line, and everything else stays as it is.

reflexer speaks the same API as the major providers, in the SDK and at the endpoint. You swap one line in your code.

Your key stays your key. Your rates, your credits and your terms stay as they are.

Your model is built from your traffic, used for you only, and destroyed when you leave.

Every response says which side answered it.

The only change: client = openai.OpenAI(api_key=os.environ["OPENAI_API_KEY"]) becomes client = reflexer.OpenAI(api_key=os.environ["OPENAI_API_KEY"]).- client = openai.OpenAI(api_key=os.environ["OPENAI_API_KEY"])+ client = reflexer.OpenAI(api_key=os.environ["OPENAI_API_KEY"])
The only change: client = openai.OpenAI(api_key=os.environ["OPENAI_API_KEY"]) becomes client = reflexer.OpenAI(api_key=os.environ["OPENAI_API_KEY"]).- client = openai.OpenAI(api_key=os.environ[…])+ client = reflexer.OpenAI(api_key=os.environ[…])

7 / 12

Your quality is never what is at stake.

We answer only when we know our answer matches the one your provider would have given.

Every request we send on comes back with the frontier answer, and we compare it with the answer we would have given.

When there is any risk we do not match, we abstain and the request goes to your provider.

So the thing that moves is how much of your traffic we cover — our bill, not your product.

For each kind of request, how often reflexer's answer is the same as the frontier one, against the bar it must clear. The two kinds under the bar go to the provider: one reflexer would have matched, one it would have got wrong.greeting✓ we answerconfirm an order✓ we answerextract address✓ we answerlong conversation✓ would have matchedpick next step✓ we answerquote a price✓ we answeropen question✗ would have missedthe bar we have to clearhow often our answer is the same as the frontier one →
For each kind of request, how often reflexer's answer is the same as the frontier one, against the bar it must clear. The two kinds under the bar go to the provider: one reflexer would have matched, one it would have got wrong.greeting✓ we answerconfirm an order✓ we answerextract address✓ we answerlong conversation✓ would have matchedpick next step✓ we answerquote a price✓ we answeropen question✗ would have missedthe bar we have to clearhow often our answer is thesame as the frontier one →

8 / 12

How reflexer would look in real traffic

The same request answered by reflexer and by a general-purpose model, field by field, with reflexer’s confidence below.reflexeryour frontier providertoolorder-statusorder-statusorder_ref"TR-90412""TR-90412"answer"Shipped on Tuesday, arrivingFriday.""Shipped on Tuesday, arrivingFriday."confidence✓ we answer · exact matchthe bar we have to clearThe same request answered by reflexer and by a general-purpose model, field by field, with reflexer’s confidence below.reflexeryour frontier providertoolfind-contactfind-contactgiven_name"Van""Van Der"family_name"Der Meer""Meer"order_ref"TR-88207""TR-88207"confidence✗ we abstain · would have been wrongthe bar we have to clearThe same request answered by reflexer and by a general-purpose model, field by field, with reflexer’s confidence below.reflexeryour frontier providerrequest_typeBillingBillingseats_requested1212renewal_at_riskyesyesowner_notifiedfalsefalsewhat_they_asked_forThey cannot get two of theirsites onto one invoice, andthey are holding off on theextra seats until that issorted out.The customer wants a singleinvoice covering all theirsites before adding more seats.confidence✓ we flag it · same meaning, other wordsthe bar we have to clear
The same request answered by reflexer and by a general-purpose model, field by field, with reflexer’s confidence below.reflexeryour frontier providertoolorder-statusorder-statusorder_ref"TR-90412""TR-90412"answer"Shipped on Tuesday,arriving Friday.""Shipped on Tuesday,arriving Friday."the bar we have to clearconfidence✓ we answer · exact matchThe same request answered by reflexer and by a general-purpose model, field by field, with reflexer’s confidence below.reflexeryour frontier providertoolfind-contactfind-contactgiven_name"Van""Van Der"family_name"Der Meer""Meer"order_ref"TR-88207""TR-88207"the bar we have to clearconfidence✗ we abstain · would have been wrongThe same request answered by reflexer and by a general-purpose model, field by field, with reflexer’s confidence below.reflexeryour frontier providerrequest_typeBillingBillingseats_requested1212renewal_at_riskyesyesowner_notifiedfalsefalsewhat_they_asked_forThey cannot get two oftheir sites onto oneinvoice, and they areholding off on the extraseats until that issorted out.The customer wants asingle invoice coveringall their sites beforeadding more seats.the bar we have to clearconfidence✓ we flag it · same meaning, other words

9 / 12

A circuit breaker with fallback to your frontier provider.

If anything on our side is degraded, a circuit breaker sends your traffic straight to your frontier provider. Your product keeps working and your users see nothing.

And if at any time you want out, you just stop calling us. No migration, no rewrite, nothing to unwind.

We want you to stay because the bill is lower, not because leaving is hard.

If reflexer is degraded, the circuit breaker opens and the traffic runs straight from your app to your frontier provider, as it did before.circuit breakeryour appunchangedreflexerdegradedyour provideryour keyyour traffic goes where it went before
If reflexer is degraded, the circuit breaker opens and the traffic runs straight from your app to your frontier provider, as it did before.your appunchangedreflexerdegradedyour provideryour keycircuit breakeryour traffic goes where it went before

10 / 12

Half price on what we answer. Nothing on the rest.

For every request reflexer answers: half of what your frontier provider would have charged for it.

For every request we pass to your provider: nothing.

Every response carries what it cost, so the invoice can be checked request by request.

The response as it comes back through reflexer: which side answered, what your provider would have charged, and what you pay: half of it.through reflexer{ "model": "your-frontier-model", "usage": { "prompt_tokens": 11000, "cached_tokens": 6000, "completion_tokens": 100 }, "reflexer": { "answered_by": "reflexer", "provider_price": "$0.008", "you_pay": "$0.004" }}half of what your provider would have charged
The response as it comes back through reflexer: which side answered, what your provider would have charged, and what you pay: half of it.through reflexer{ "model": "your-frontier-model", "usage": { "prompt_tokens": 11000, "cached_tokens": 6000, "completion_tokens": 100 }, "reflexer": { "answered_by": "reflexer", "provider_price": "$0.008", "you_pay": "$0.004" }}half of what your provider would have charged

11 / 12

Closed beta, starting Q4 2026.

Not live yet. We start with a small group of companies in Q4 2026.

Want in? Tell us what you run at hello@reflexer.ai.

Companies on token bill against how long users wait, split into four: we cut the wait; we want to hear from you; you do not need us; we cut your bill.how long your users waityour token billwe cut the waitwe want to hear from youyou don’t need uswe cut your bill
Companies on token bill against how long users wait, split into four: we cut the wait; we want to hear from you; you do not need us; we cut your bill.how long your users waityour token billwe cut the waitwe want to hear from youyou don’t need uswe cut your bill

12 / 12

Team: one founder.

I am Rafael Carrascosa. At Mercado Libre I was Director and then Senior Principal of machine learning, with over a hundred people in my org.

Before that I co-founded a startup, meshh, and was part of the team that sold Machinalis to Mercado Libre.

reflexer is all I do, and then some. In the beta you talk to me directly, the person who builds the models.

Know a team with a big token bill? Send them my way.
LinkedIn · hello@reflexer.ai

Rafael CarrascosaRafael Carrascosafounder @ reflexer
Rafael CarrascosaRafael Carrascosafounder @ reflexer