Most of the AI traffic we build into products is not a conversation. It is one piece of software asking another a question and needing an answer it can act on: Is there personal data in this text? Which queue does this ticket go to? Should this prompt leave the company?
Chat models are a poor fit for that job. You ask for a decision and you get a string back: a label, a paragraph, a refusal, or something the next line of code can't parse. Getting a decision out of it means parsing, and parsing can fail.
Jev, from TypeSafe, takes a different approach. In TypeSafe's own words: unstructured state in, typed probabilistic decisions out. We spent an internal session going through it and running it against scenarios from the AI firewall we are building. This post covers what we learned.
Facts and numbers below were checked on 15 October 2026 against TypeSafe's published models page and live runs on the public Jev playground. Prices and versions move quickly, so check them again before you rely on them.
Two names worth knowing
System One. The idea comes from Daniel Kahneman's Thinking, Fast and Slow. System 1 is fast, reflexive judgment. System 2 is slow, deliberate reasoning. Reasoning models are System 2 tools. Jev is built for System 1: the quick judgment call. When a task really needs a long chain of thought, a reasoning model is still the right tool.
Jevons. The model is named after William Stanley Jevons, the economist who noticed that more efficient steam engines didn't reduce coal use. They increased it, because cheaper power made new uses worthwhile. TypeSafe expects the same thing to happen here: each drop in the cost of a decision makes work viable that made no sense at chatbot prices.
What Jev is: one state, many questions, one call
A Jev request has three parts:
- State: the evidence. It can be a string, a JSON object, or an array of text. Think of it as the packet you'd hand to a panel of reviewers.
- Questions: each has an
id, atype, andinstructions. You define the allowed answers. Jev doesn't invent new kinds of answer. - Answers: one typed value per question, each with a probability. Every question sees the same state, and they are all scored together in parallel.
The contract is small enough to show in full:
{
"model": "jev-latest",
"state": "Johnathan Doe, john.doe@company.com",
"questions": {
"has_contact": {
"type": "noul",
"instructions": "A real person or a contact?"
}
}
}
{ "type": "noul", "noul": 0.93 }
That 0.93 is the probability that the answer is yes. The playground labels this type Boolean; the API calls it Noul. They are the same number.
Because all the questions share one state and run in one call, asking twelve questions about one document is much cheaper than sending that document twelve times.
The three question types
| Type | Ask it when | You get back |
|---|---|---|
| Noul (Boolean) | How likely is this claim? | A probability from 0 to 1 |
| Choice | Which of these labels? | The chosen label, a probability for each option, and a confidence |
| Score | Where on this ordered scale? | A value that may fall between levels, plus a confidence |
A few rules make each type work well:
- Noul: phrase the question so a high number means yes, and ask about one claim only. A value near 0.5 means the model is split. It is P(yes), not a measure of degree. "0.95 angry" doesn't mean "very angry"; it means "almost certainly angry".
- Choice: make each option a distinct action. Overlapping labels split the probability between them. Add a
noneoption when nothing might fit. The API allows up to 255 options; the free playground allows 10. - Score: levels are ordered. A result of 1.4 means the probability mass sits between level 1 and level 2. Set a threshold on it, but don't treat 1.4 as a measured quantity.
Calibration, not eloquence
TypeSafe describes three training objectives that produce three kinds of model:
| Objective | The model learns to produce | Typical home |
|---|---|---|
| RLHF: human preference | Text people like to read | Chat |
| RLVR: verifiable rewards | A long solution you can check | Reasoning models |
| RLCD: calibrated decisions | A decision whose probability matches how often it is right | Jev |
Calibration is a property of many predictions taken together. If Jev says 0.80 on a thousand similar questions, about eight hundred of them should turn out to be yes. A single 0.80 is still a guess, and you may decide to override it.
Confidence is the shape of the distribution
For a Choice with n options and top probability p, Jev reports:
confidence = (p − 1/n) / (1 − 1/n)
If all the probability sits on one option, confidence is 1. An even split gives 0. Only the top probability counts.
This gives you a simple policy pattern with three bands:
- High: the code acts.
- Middle: confirm, flag, or gather more evidence.
- Low: don't guess; send it to a person or a stricter system.
Where you put the cutoffs depends on the cost of being wrong. A wrong help-article suggestion can tolerate 0.6. Letting a card number through cannot.
Jev decides. Software acts. Jev doesn't delete a row, send an email, or rewrite a prompt. That is a deliberate contrast with LLM agents that can edit your code or call services on their own.
Jobs it is already used for
TypeSafe's cookbooks and demos all follow the same pattern: Jev returns a label or a probability, and ordinary code performs the action.
| Job | Jev decides | Code then |
|---|---|---|
| Support desk | Queue, frustration, asked for a human? | Route, raise priority, or hand off |
| Refunds | Asked, policy allows it, duplicate charge? | Issue, refuse, or escalate |
| Chat guardrails | Jailbreak, harm, dosage, self-harm, severity | Pass, review, block, or crisis path |
| Citations and RAG | Does this source support the claim? | Drop the quote or keep the passage |
| Checked extraction | Per field: missing, invented, or mis-copied? | Keep the cheap parse or re-run it |
| Pick a value | Which regex hit is the receipt address? | Copy that span; never retype it |
| Agents | Which skill, which tool, which arguments? | Call, ask the user, or do nothing |
| Large catalogs | Which industry, patent branch, same product? | Take the fine label only if confidence is high |
| Real-time toys | Doom: which move? Wiki race: which link? | The game takes the step, about 10 per second |
Whatever the input (a ticket, a paragraph, a tool call, a game frame), the product around Jev comes down to a few if statements, with a different threshold for each risk.
Where it stops
These limits matter most in practice:
- The schema is closed. A Choice returns one of your labels. It can't invent a sixth team name, but it can still pick the wrong one of the five. "No hallucinated answers" doesn't mean "no wrong answers".
- It doesn't write the value. It won't redact a paragraph or retype an email. A regex over-finds candidate spans, Jev picks one, and code copies that span unchanged.
- It is a poor calculator. Counting, date ordering, and exact arithmetic belong in code. A plain regex is also better than Jev at spotting card numbers. Ask Jev which span is the mobile number, not what the number is.
- It answers the words you wrote. Ask one claim per question. Compound questions ("angry and wants a refund?") give muddy results, and lots of irrelevant detail in the state hurts.
- Order can matter. On 1.13 at least, reordering Choice options has been observed to shift the probabilities slightly when the top answer wasn't dominant.
- The state is untrusted. Text inside the state can argue for its own label. Treat user prompts as hostile until you've tested that case.
- It is hosted. There is no weight file to download and no fine-tuning. Every account uses the same weights, and you specialize the model through the request. English is its strongest language. TypeSafe says customer inputs aren't used for training, and zero-data-retention is available on request.
Testing it in an AI firewall
We are building an AI firewall: a plugin that sits on the network edge, intercepts prompts headed to third-party AI services, and masks sensitive data before it leaves the company. Today it combines regex with an open-source named-entity recognizer (spaCy). NER is useful but noisy: it flags things that look like names or identifiers and isn't always right. We tested Jev as the judge that checks the detector's output.
Example A: real contact details plus a codename
Primary contact Johnathan Doe can be reached at john.doe@company.com. For urgent matters call +1 (415) 555-0192. Tax identifiers on file include SSN 123-45-6789. Helpdesk extension 4512 routes to the generic queue, not a person. Project codename NIGHTINGALE-42 launches Q3 without customer data export. Server hostname prod-app-03.internal.example.corp is infrastructure naming.
| Question | Result |
|---|---|
| Does this text include a real person's name or a direct way to contact them? | 93% yes |
| Is NIGHTINGALE-42 personal data that should be removed? | 12% yes |
| What should a privacy firewall do? | Redact and forward 89% · Block 10% · Human review 1% · Allow 0% |
525 input tokens, 455 ms, confidence 0.85: four options with a top probability of 0.89 gives (0.89 − 0.25) / 0.75 = 0.85. The call returned a label, not a redacted paragraph. Code does the redaction.
Example B: same questions, personal data removed
With the name, email, phone, and SSN taken out, Jev correctly scores "person or contact" at 13% and the codename at 16%. The action question is a close call this time: Redact and forward 52%, Allow unchanged 42%, confidence 0.36.
This is the useful result. The low confidence tells the firewall that the evidence doesn't support an automated override. Below about 0.6, this Choice shouldn't override the span detector, and in this case the paragraph should go through unchanged.
Example C: four addresses, one gets the receipt
From: Dana Whit <dana.whit@acme-corp.com>
To: billing@acme-corp.com
Cc: orders@acme-corp.com
Reply-To: dana.personal@gmail.com
Hi team - please do not use the billing alias for this one.
Send my receipt to my personal address instead. Thanks, Dana.
A regex finds all four addresses, but only the prose says which one to use. Jev puts 100% on dana.personal@gmail.com (confidence 1.00, 664 ms). Code copies that span and lowercases it. The model never retypes the address.
The muddy question, on purpose
"Is this text free of personal data, safe to forward, and is NIGHTINGALE-42 a codename?" That is three claims in one question, and the first is inverted. Whatever number comes back is hard to act on. Split it back into Example A's atomic questions and the number becomes usable as policy again.
Our takeaway: Jev on its own isn't the firewall, but alongside NER it makes a strong second opinion. Our next steps are combining it with spaCy output, then setting probability and confidence thresholds that separate automatic redaction from human review.
What a query costs
At the time of writing, both Jev 1.13 (jev-latest, stable) and Jev 1.14 (jev-preview, shipped 14 October 2026) are priced at $0.042 per million input tokens, with $0 for output. The context limit is 64k per request and 32k per question. Text is the main input; image input is experimental.
dollars = input tokens / 1,000,000 × 0.042
| Query we ran | Tokens | One call | 1,000 calls | 1,000,000 calls |
|---|---|---|---|---|
| A: mixed PII | 525 | $0.000022 | $0.022 | $22.05 |
| B: no personal data | 490 | $0.000021 | $0.021 | $20.58 |
| C: pick the receipt address | 503 | $0.000021 | $0.021 | $21.13 |
A thousand checks cost about two cents. A company of 500 people sending 20 prompts a day through a check like A costs about $0.22 a day, or roughly $5 a month. A full 64k-token request costs $0.0027.
For comparison, here is the same judgment from a model that has to write a ~300-token JSON verdict (cookbook list prices, not a quote):
| Judge for Example A | One call | A million calls |
|---|---|---|
| Jev (measured tokens) | $0.000022 | $22 |
| Mini model, writing the verdict | $0.0017 | $1,740 |
| Reasoning model, same guess | $0.012 | $11,600 |
On this small call, Jev is roughly 80× cheaper than a mini model, mostly because the mini model has to pay for the output tokens it writes. Plan your budget around the input line. Free output is TypeSafe's current policy and could change.
The short version
Jev is a function from evidence to probabilities. It is cheap enough that the question stops being "can we afford to ask?" and becomes "which question is atomic enough to threshold?"
Good questions: is this span personal data, and which allowed action applies? Wrong job: rewriting the prompt, or typing the email address out itself.
If you're working out where a decision model like this fits into your AI security or automation stack, get in touch.
Najveći dio AI saobraćaja koji ugrađujemo u proizvode nije razgovor. To je softver koji drugom softveru postavlja pitanje i treba mu odgovor na osnovu kojeg može odmah da djeluje: Ima li u ovom tekstu ličnih podataka? U koji red ide ovaj tiket? Smije li ovaj prompt da izađe iz firme?
Chat modeli tu nisu dobar alat. Tražite odluku, a dobijete string: labelu, pasus, odbijanje ili nešto što sljedeća linija koda ne može da parsira. Da biste iz toga izvukli odluku, morate parsirati, a parsiranje može da padne.
Jev, model kompanije TypeSafe, ide drugim putem. Kako to TypeSafe kaže: nestrukturirano stanje unutra, tipizirane probabilističke odluke napolje. Prošli smo kroz njega na internoj sesiji i pustili ga kroz scenarije iz AI firewalla koji razvijamo. U nastavku je ono što smo naučili.
Činjenice i brojevi u nastavku provjereni su 15. oktobra 2026. na TypeSafe stranici sa modelima i kroz pokretanja na javnom Jev playgroundu. Cijene i verzije se brzo mijenjaju, pa ih provjerite prije nego što se na njih oslonite.
Dva imena koja vrijedi znati
System One. Ideja dolazi iz knjige Daniela Kahnemana Thinking, Fast and Slow. Sistem 1 je brzo, refleksno rasuđivanje. Sistem 2 je sporo, promišljeno razmišljanje. Reasoning modeli su alati Sistema 2. Jev je napravljen za Sistem 1, za brzu procjenu. Kada zadatak zaista traži dug lanac razmišljanja, reasoning model ostaje pravi izbor.
Jevons. Model je dobio ime po Williamu Stanleyju Jevonsu, ekonomisti koji je primijetio da efikasnije parne mašine nisu smanjile potrošnju uglja. Povećale su je, jer je jeftinija energija učinila isplativim nove primjene. TypeSafe očekuje isti obrazac i ovdje: svaki pad cijene jedne odluke otvara posao koji po cijeni chatbota nije imao smisla.
Šta je Jev: jedno stanje, mnogo pitanja, jedan poziv
Jev zahtjev ima tri dijela:
- Stanje (state): dokazi. Može biti string, JSON objekat ili niz tekstova. Zamislite ga kao paket koji biste predali komisiji recenzenata.
- Pitanja (questions): svako ima
id,typeiinstructions. Vi definišete dozvoljene odgovore. Jev ne izmišlja nove vrste odgovora. - Odgovori: jedna tipizirana vrijednost po pitanju, svaka sa vjerovatnoćom. Sva pitanja vide isto stanje i ocjenjuju se zajedno, paralelno.
Cijeli format zahtjeva i odgovora je dovoljno mali da stane ovdje:
{
"model": "jev-latest",
"state": "Johnathan Doe, john.doe@company.com",
"questions": {
"has_contact": {
"type": "noul",
"instructions": "A real person or a contact?"
}
}
}
{ "type": "noul", "noul": 0.93 }
Vrijednost 0.93 je vjerovatnoća da je odgovor da. Playground ovaj tip zove Boolean, a API Noul. U pitanju je isti broj.
Pošto sva pitanja dijele jedno stanje i idu u jednom pozivu, dvanaest pitanja o jednom dokumentu je mnogo jeftinije nego dvanaest puta poslati isti dokument.
Tri tipa pitanja
| Tip | Kada ga koristiti | Šta dobijate |
|---|---|---|
| Noul (Boolean) | Koliko je vjerovatna ova tvrdnja? | Vjerovatnoću od 0 do 1 |
| Choice | Koja od ovih labela? | Izabranu labelu, vjerovatnoću za svaku opciju i confidence |
| Score | Gdje na ovoj uređenoj skali? | Vrijednost koja može biti između nivoa, plus confidence |
Nekoliko pravila pomaže da svaki tip radi dobro:
- Noul: formulišite pitanje tako da visok broj znači da i pitajte samo za jednu tvrdnju. Vrijednost blizu 0,5 znači da je model podijeljen. To je P(da), a ne mjera intenziteta. „0,95 ljut“ ne znači „veoma ljut“, nego „skoro sigurno ljut“.
- Choice: neka svaka opcija bude zasebna akcija. Labele koje se preklapaju dijele vjerovatnoću između sebe. Dodajte opciju
nonekada možda ništa ne odgovara. API dozvoljava do 255 opcija, a besplatni playground 10. - Score: nivoi su uređeni. Rezultat 1,4 znači da je vjerovatnoća između nivoa 1 i 2. Postavite prag na njega, ali 1,4 nemojte tretirati kao izmjerenu veličinu.
Kalibracija, a ne elokvencija
TypeSafe opisuje tri cilja treniranja koja daju tri vrste modela:
| Cilj | Model uči da proizvede | Tipična primjena |
|---|---|---|
| RLHF: ljudske preferencije | Tekst koji ljudi rado čitaju | Chat |
| RLVR: provjerljive nagrade | Dugo rješenje koje se može provjeriti | Reasoning modeli |
| RLCD: kalibrisane odluke | Odluku čija vjerovatnoća odgovara tome koliko je često tačna | Jev |
Kalibracija je svojstvo mnogo predviđanja zajedno. Ako Jev kaže 0,80 na hiljadu sličnih pitanja, oko osamsto bi trebalo da bude da. Jedno pojedinačno 0,80 je i dalje procjena koju možete i nadjačati.
Confidence je oblik raspodjele
Za Choice pitanje sa n opcija i najvećom vjerovatnoćom p, Jev vraća:
confidence = (p − 1/n) / (1 − 1/n)
Ako je sva vjerovatnoća na jednoj opciji, confidence je 1. Ravnomjerna raspodjela daje 0. Računa se samo najveća vjerovatnoća.
To daje jednostavan obrazac politike u tri pojasa:
- Visok: kod djeluje.
- Srednji: potvrdi, označi ili prikupi još dokaza.
- Nizak: ne pogađaj; pošalji čovjeku ili strožijem sistemu.
Gdje ćete postaviti granice zavisi od cijene greške. Pogrešan prijedlog članka iz pomoći može da prođe sa 0,6. Propuštanje broja kartice ne može.
Jev odlučuje. Softver djeluje. Jev ne briše red u bazi, ne šalje mejl i ne prepisuje prompt. To je namjerna razlika u odnosu na LLM agente koji mogu sami da mijenjaju vaš kod ili pozivaju servise.
Za šta se već koristi
TypeSafe cookbook primjeri i demo aplikacije prate isti obrazac: Jev vraća labelu ili vjerovatnoću, a običan kod izvršava akciju.
| Posao | Jev odlučuje | Kod zatim |
|---|---|---|
| Support desk | Red, frustracija, traži li čovjeka? | Rutira, podiže prioritet ili predaje dalje |
| Povrat novca | Traži li, dozvoljava li politika, dupla naplata? | Odobrava, odbija ili eskalira |
| Zaštita chata | Jailbreak, šteta, doziranje, samopovređivanje, ozbiljnost | Propušta, šalje na pregled, blokira ili krizni tok |
| Citati i RAG | Podržava li izvor tvrdnju? | Izbacuje citat ili zadržava pasus |
| Provjerena ekstrakcija | Po polju: nedostaje, izmišljeno ili pogrešno prepisano? | Zadržava jeftino parsiranje ili ga ponavlja |
| Izbor vrijednosti | Koji regex pogodak je adresa za račun? | Kopira taj dio teksta; nikad ga ne prekucava |
| Agenti | Koji skill, koji alat, koji argumenti? | Poziva, pita korisnika ili ne radi ništa |
| Veliki katalozi | Koja industrija, grana patenta, isti proizvod? | Uzima finu labelu samo uz visok confidence |
| Igre u realnom vremenu | Doom: koji potez? Wiki trka: koji link? | Igra pravi korak, oko 10 puta u sekundi |
Šta god da je na ulazu (tiket, pasus, poziv alata ili frejm igre), proizvod oko Jeva se svodi na nekoliko if naredbi, sa drugačijim pragom za svaki rizik.
Gdje mu je granica
Ova ograničenja su najvažnija u praksi:
- Šema je zatvorena. Choice vraća jednu od vaših labela. Ne može da izmisli šesto ime tima, ali i dalje može da izabere pogrešno od ponuđenih pet. „Nema haluciniranih odgovora“ ne znači „nema pogrešnih odgovora“.
- Ne piše vrijednost. Neće sam maskirati pasus niti prekucati mejl. Regex pronađe više kandidata nego što treba, Jev izabere jedan, a kod taj dio teksta kopira nepromijenjen.
- Loš je kalkulator. Brojanje, redoslijed datuma i tačna aritmetika pripadaju kodu. Običan regex je i bolji od Jeva u prepoznavanju brojeva kartica. Pitajte Jeva koji dio teksta je broj mobilnog, a ne koji je to broj.
- Odgovara na riječi koje ste napisali. Jedna tvrdnja po pitanju. Složena pitanja („ljut i traži povrat?“) daju mutne rezultate, a mnogo nebitnih detalja u stanju škodi.
- Redoslijed može da utiče. Bar na verziji 1.13 primijećeno je da promjena redoslijeda Choice opcija može blago da pomjeri vjerovatnoće kada najvjerovatniji odgovor nije bio izrazito dominantan.
- Stanju se ne vjeruje. Tekst unutar stanja može da „lobira“ za sopstvenu labelu. Korisnički prompt tretirajte kao neprijateljski dok taj slučaj ne istestirate.
- Hostovan je. Nema fajla sa težinama i nema fine-tuninga. Svi nalozi koriste iste težine, a model specijalizujete kroz sam zahtjev. Najjači je na engleskom. TypeSafe navodi da se podaci klijenata ne koriste za treniranje, a zero-data-retention je dostupan na zahtjev.
Test u AI firewallu
Razvijamo AI firewall: plugin na ivici mreže koji presreće promptove upućene AI servisima trećih strana i maskira osjetljive podatke prije nego što napuste firmu. Trenutno kombinuje regex i open-source prepoznavanje imenovanih entiteta (spaCy). NER je koristan, ali bučan: označava stvari koje liče na imena ili identifikatore i ne pogađa uvijek. Jev smo testirali kao sudiju koji provjerava ono što detektor pronađe.
Primjer A: pravi kontakt podaci plus kodno ime
Primary contact Johnathan Doe can be reached at john.doe@company.com. For urgent matters call +1 (415) 555-0192. Tax identifiers on file include SSN 123-45-6789. Helpdesk extension 4512 routes to the generic queue, not a person. Project codename NIGHTINGALE-42 launches Q3 without customer data export. Server hostname prod-app-03.internal.example.corp is infrastructure naming.
| Pitanje | Rezultat |
|---|---|
| Sadrži li tekst ime stvarne osobe ili direktan način da je kontaktirate? | 93% da |
| Da li je NIGHTINGALE-42 lični podatak koji treba ukloniti? | 12% da |
| Šta privacy firewall treba da uradi? | Maskirati i proslijediti 89% · Blokirati 10% · Ljudski pregled 1% · Propustiti 0% |
525 ulaznih tokena, 455 ms, confidence 0,85: četiri opcije i najveća vjerovatnoća 0,89 daju (0,89 − 0,25) / 0,75 = 0,85. Poziv je vratio labelu, a ne maskiran pasus. Maskiranje radi kod.
Primjer B: ista pitanja, bez ličnih podataka
Kada se uklone ime, mejl, telefon i SSN, Jev ispravno daje „osoba ili kontakt“ 13%, a kodnom imenu 16%. Pitanje o akciji je ovaj put tijesno: Maskirati i proslijediti 52%, Propustiti nepromijenjeno 42%, confidence 0,36.
To je koristan rezultat. Nizak confidence govori firewallu da dokazi ne podržavaju automatsku odluku. Ispod otprilike 0,6, ovaj Choice ne bi trebalo da nadjača detektor, a u ovom slučaju pasus treba da prođe nepromijenjen.
Primjer C: četiri adrese, samo jedna dobija račun
From: Dana Whit <dana.whit@acme-corp.com>
To: billing@acme-corp.com
Cc: orders@acme-corp.com
Reply-To: dana.personal@gmail.com
Hi team - please do not use the billing alias for this one.
Send my receipt to my personal address instead. Thanks, Dana.
Regex pronalazi sve četiri adrese, ali samo tekst poruke kaže koju treba koristiti. Jev daje 100% adresi dana.personal@gmail.com (confidence 1,00, 664 ms). Kod kopira taj dio teksta i pretvara ga u mala slova. Model adresu nikada ne prekucava.
Namjerno mutno pitanje
„Da li je ovaj tekst bez ličnih podataka, siguran za slanje, i da li je NIGHTINGALE-42 kodno ime?“ To su tri tvrdnje u jednom pitanju, a prva je još i obrnuta. Šta god da se vrati, teško je postupiti po tome. Razbijte ga nazad na atomska pitanja iz Primjera A i broj ponovo postaje upotrebljiv kao politika.
Naš zaključak: Jev sam po sebi nije firewall, ali uz NER je jako drugo mišljenje. Sljedeći koraci su nam kombinovanje sa spaCy izlazom, a zatim postavljanje pragova vjerovatnoće i confidence-a koji razdvajaju automatsko maskiranje od ljudskog pregleda.
Koliko košta jedan upit
U trenutku pisanja, i Jev 1.13 (jev-latest, stabilna verzija) i Jev 1.14 (jev-preview, objavljen 14. oktobra 2026) koštaju $0,042 po milion ulaznih tokena, a izlaz je $0. Kontekst je 64k po zahtjevu i 32k po pitanju. Tekst je glavni ulaz, a slike su eksperimentalne.
dolari = ulazni tokeni / 1.000.000 × 0,042
| Upit koji smo pokrenuli | Tokeni | Jedan poziv | 1.000 poziva | 1.000.000 poziva |
|---|---|---|---|---|
| A: mješoviti PII | 525 | $0,000022 | $0,022 | $22,05 |
| B: bez ličnih podataka | 490 | $0,000021 | $0,021 | $20,58 |
| C: izbor adrese za račun | 503 | $0,000021 | $0,021 | $21,13 |
Hiljadu provjera košta oko dva centa. Firma od 500 ljudi koja dnevno pošalje po 20 promptova kroz provjeru poput A plaća oko $0,22 dnevno, odnosno otprilike $5 mjesečno. Pun zahtjev od 64k tokena košta $0,0027.
Poređenja radi, evo iste procjene od modela koji mora da napiše JSON presudu od ~300 tokena (cijene iz cookbooka, ne ponuda):
| Sudija za Primjer A | Jedan poziv | Milion poziva |
|---|---|---|
| Jev (izmjereni tokeni) | $0,000022 | $22 |
| Mini model koji piše presudu | $0,0017 | $1.740 |
| Reasoning model, ista procjena | $0,012 | $11.600 |
Na ovako malom pozivu Jev je otprilike 80× jeftiniji od mini modela, najviše zato što mini model mora da plati izlazne tokene koje piše. Budžet planirajte prema ulaznim tokenima. Besplatan izlaz je trenutna TypeSafe politika i može da se promijeni.
Ukratko
Jev je funkcija koja od dokaza pravi vjerovatnoće. Dovoljno je jeftin da pitanje više nije „možemo li sebi priuštiti da pitamo?“, nego „koje pitanje je dovoljno atomsko da mu postavimo prag?“
Dobra pitanja: da li je ovaj dio teksta lični podatak i koja dozvoljena akcija važi? Pogrešan posao: prepisivanje prompta ili samostalno prekucavanje mejl adrese.
Ako razmišljate gdje bi ovakav model za odlučivanje mogao da se uklopi u vaš AI security ili automatizaciju, javite nam se.