Jev for Meta Ads: 8 jobs a decision model can take over
Jev is TypeSafe's new decision model: pick-one, score and yes/no answers for about $0.0004 each, no text. The most shared Jev demo so far is a Meta one: Matthew Berman reported breaking down 724 live ads from 37 brands in 40 seconds for 9 cents. Here are 8 Meta Ads jobs that fit, and the ones that do not. Ryze AI publishes this blog and does not run Jev in production yet.
Built by our community of 2,000 marketers
Free skills and prompts for paid ads and SEO
Templates for Claude, ChatGPT and Perplexity.
Clients we work withClients we
work with
The 8 jobs at a glance
Meta Ads runs on volume: dozens of new creatives a week, thousands of competitor ads in the Ad Library, comments under every ad. Reading all of that with a frontier model is slow and expensive, which is why most teams sample. Tagging is a bounded question, so a decision model can read everything.
| # | Job | Answer type | Question Jev answers | What happens next |
|---|---|---|---|---|
| 1 | Creative tagging | Choice ×4 | Which hook, format, offer and awareness stage does this ad use? | Every ad gets four tags, and performance is grouped by tag |
| 2 | Ad Library scanning | Noul | Is this competitor ad new in concept or a variation of one we have tagged? | Only new concepts reach the creative team |
| 3 | Ad to landing page match | Score | Does the landing page deliver what this ad promises? | Low scores are flagged before spend, not after |
| 4 | Fatigue triage | Choice | Given frequency up and CTR down for 3 days, replace, refresh or leave? | Replace goes to the creative queue, leave is logged |
| 5 | Comment moderation | Choice | Is this comment a question, a complaint, spam or praise? | Questions and complaints are routed, spam is hidden |
| 6 | Policy pre-check | Score | Is this copy likely to trip a personal-attributes or health-claims policy? | High risk is rewritten before upload |
| 7 | Audience and creative fit | Choice | Is this creative written for a cold, warm or hot audience? | Mismatched ad sets are listed |
| 8 | UGC brief compliance | Noul ×5 | Does this creator video cover the 5 points in the brief? | Misses go back to the creator with the checklist |
Every row has a known answer set, which is the test for whether a decision model fits. If the answer has to be written, it belongs to an LLM.
Quick answer: Jev for meta ads means using TypeSafe's decision model for the high-volume judgment steps in meta ads. 1) It answers pick-one, score and yes/no questions for about $0.0004 each. 2) It cannot write, explain or plan. 3) The top fit here is creative tagging, then ad library scanning. 4) Run it behind a confidence threshold, with an approval queue for the middle band. 5) Keep a frontier model for anything a person reads and a human on anything hard to undo. 6) Access is by waitlist as of September 2026.
What Jev is, in one table
Jev is a model from TypeSafe AI, the company founded by Diogo Almeida after his work on ChatGPT at OpenAI. It opened in early access on September 15, 2026. TypeSafe calls it a System One model: you send it a block of text (the state) and a set of typed questions, and it returns typed answers with a probability attached. It does not write sentences.
| Property | What TypeSafe reports |
|---|---|
| Question types | Choice (pick one option), Score (a number in a range), Noul (a yes/no probability) |
| Price | $0.042 per 1M input tokens. Output is free. Roughly $0.0004 per decision |
| Latency | 70 to 500 ms per call, all answers returned in one parallel pass |
| Structured-output errors | 0%, because the output is schema-constrained |
| Accuracy on TypeSafe's 4-workflow benchmark | Jev 67.8%, Claude Opus 5 73.1%, GPT-5.6 Sol 74.1% |
| Access | Waitlist. Early users on X report approval the same day or the next |
| What it cannot do | Write text, code or a rationale. It gives a number and no explanation |
Two caveats belong next to those numbers. The benchmarks are vendor-reported and measure agreement with frontier models, and no large independent reproduction exists yet. And TypeSafe says itself that it cannot prove the launch price is unsubsidized. Treat the price as today's price.
The 8 jobs, one by one
1. Creative tagging
Question: "Which hook, format, offer and awareness stage does this ad use?" Answer type: Choice ×4. Next: Every ad gets four tags, and performance is grouped by tag.
This is the job in Berman's demo. One state, four Choice questions, one call. Your own account and the Ad Library scrape go through the same tagger, so your winners and your competitors' are comparable.
2. Ad Library scanning
Question: "Is this competitor ad new in concept or a variation of one we have tagged?" Answer type: Noul. Next: Only new concepts reach the creative team.
A brand running 1,000 ads usually has 30 ideas. The yes/no collapses variations so the weekly scan is a short list.
3. Ad to landing page match
Question: "Does the landing page deliver what this ad promises?" Answer type: Score. Next: Low scores are flagged before spend, not after.
Berman's demo lists landing page mismatch as one of the outputs. It is also the cheapest CVR fix on most accounts.
4. Fatigue triage
Question: "Given frequency up and CTR down for 3 days, replace, refresh or leave?" Answer type: Choice. Next: Replace goes to the creative queue, leave is logged.
The numbers come from the Insights API. Jev only makes the call from a fixed set, and the 50-conversion rule stays in code so nothing is judged early.
5. Comment moderation
Question: "Is this comment a question, a complaint, spam or praise?" Answer type: Choice. Next: Questions and complaints are routed, spam is hidden.
Comments under ads affect relevance and nobody reads them at scale. Four labels and a threshold cover it.
6. Policy pre-check
Question: "Is this copy likely to trip a personal-attributes or health-claims policy?" Answer type: Score. Next: High risk is rewritten before upload.
Not a replacement for Meta's review. It catches the obvious rejections before they cost a day.
7. Audience and creative fit
Question: "Is this creative written for a cold, warm or hot audience?" Answer type: Choice. Next: Mismatched ad sets are listed.
Retargeting copy in a prospecting ad set is a common, silent waste. One label per ad exposes it.
8. UGC brief compliance
Question: "Does this creator video cover the 5 points in the brief?" Answer type: Noul ×5. Next: Misses go back to the creator with the checklist.
Works on the transcript. Five yes/no answers per video replace a manual review pass.
Get a free instant audit
Get a free, instant read on your paid ads or SEO — and fix it right away.
Paid ads audit
- Catch wasted spend & broad-match leaks
- Find account structure gaps
- Rank your quickest wins
- Spot PMax & brand-search overlap
- Check conversion-tracking health
- Benchmark CPC vs your industry
- Catch wasted spend & broad-match leaks
- Find account structure gaps
- Rank your quickest wins
- Spot PMax & brand-search overlap
- Check conversion-tracking health
- Benchmark CPC vs your industry
Free · no credit card · instant
SEO audit
- Find keyword & ranking gaps
- Catch technical SEO issues
- Rank your fastest wins
- Surface thin & duplicate pages
- Check indexing & crawl coverage
- Compare backlinks vs competitors
- Find keyword & ranking gaps
- Catch technical SEO issues
- Rank your fastest wins
- Surface thin & duplicate pages
- Check indexing & crawl coverage
- Compare backlinks vs competitors
Free · no credit card · instant
Ryze AI — Autonomous Marketing
A decision is only useful if something acts on it
- ✓Ryze AI reads the live account and makes the change
- ✓Approval queue for budgets, bids and money pages
- ✓Free to connect, flat fee for autopilot
2,000+
Marketers
$500M+
Ad spend
23
Countries
The pattern behind every use case
Every use case on this page follows one architecture. Code finds the candidates, Jev judges them, a threshold decides what happens, and a frontier model writes only what passed.
The decide-then-act loop
- Collect the rows with code or an API: search terms, ads, pages, AI answers. No model needed.
- Shortlist with rules or embeddings, so Jev sees 10 to 15 candidates per item and not the whole account.
- Ask Jev one bounded question per row: a Choice, a Score or a Noul. Every answer comes back with a probability.
- Apply a threshold. High confidence is applied automatically. The middle band goes to an approval queue. Low confidence is dropped.
- Write with an LLM only where text is needed: the new ad, the new title, the one-line reason a client will read.
- Verify the LLM's output with Jev again before it ships.
This is the same split Ryze AI already uses between changes it makes on its own and changes that wait for approval. A cheap decision model widens the first group without touching the rule for the second.
What it costs at the list price
These are arithmetic at TypeSafe's published price of $0.042 per 1M input tokens with free output. They are not measured runs, and token counts depend on how much context you put in each state.
| Job | Rows | Approx. input tokens | Cost at $0.042 per 1M |
|---|---|---|---|
| Creative tagging, own account | 500 ads | ~0.3M | ~$0.01 |
| Ad Library scan, 10 competitors | 5,000 ads | ~3M | ~$0.13 |
| Comment moderation | 20,000 comments | ~1.2M | ~$0.05 |
The point of the table is the order of magnitude. Jobs that cost tens of dollars per run on a frontier model, and therefore run monthly, cost cents and can run nightly.
What Jev cannot do
Jev is narrow by design, and the narrow part matters as much as the price.
- No rationale. You get 0.91, never a sentence explaining it. Anything a client or a finance team will read still needs a frontier model to write the reason.
- Bounded answers only. If the set of possible answers is not known up front, it is the wrong tool. Strategy, copy and diagnosis stay with an LLM.
- Accuracy is triage-grade. 67.8% agreement on the vendor's own benchmark is fine behind a confidence threshold and wrong for auto-applying a budget change.
- "Cannot hallucinate" means it cannot break the schema. It can still pick the wrong option. Measure it against human labels on your own data before trusting a threshold.
- Waitlist and pricing risk. There is no general availability date, and the price may move.
Keep the frontier model for writing hooks and scripts, for explaining why a creative won, and for the brief. Keep a human on turning ads on, budgets over 20% and anything near the pixel.
Where to go deeper on each job
A decision model only sorts. The guides below cover the work around it: getting the data, writing the rules, and making the changes.
Alex M.
VP Growth
Multi-Brand Ecommerce Group
We had five channels and five people each defending their own budget. Handing allocation to Ryze AI cut blended CAC 27% in a quarter — not because the bidding got smarter, but because money finally moved to whichever channel was winning that week.”
27%
Blended CAC reduction
5
Channels automated
12 weeks
To full autonomy
1,000+ marketers use Ryze
Automating hundreds of agencies
How to start before you have access
Jev is in early access. There is no public self-serve signup yet.
1. Join the waitlist
Request access at typesafe.ai. Early users report approval the same day or within a day, but TypeSafe publishes no guarantee.
2. Label a sample first
Before you have a key, hand-label 500 rows of the job you care about. That set is how you will pick thresholds, and it works for any model.
3. Build the loop on a model you can call today
The decide-then-act loop runs on Claude Haiku 4.5 or Gemini Flash-Lite with structured outputs. It costs more per call, and the code does not change when you swap Jev in.
4. Connect the account data
The rows have to come from somewhere. The Ryze AI Meta Ads MCP gives Claude, ChatGPT, Cursor and Grok live access to the account, so the collect step is one prompt.
Sources and how this page was written
What is verified and what is not
- Verified from TypeSafe's launch post: the price, early access and the System One definition.
- Vendor-reported, not reproduced: latency, error rates and the benchmark scores.
- Third-party claims from X, not reproduced by us: the 724-ads, 586-pages and 700-leads demos.
- Our own arithmetic: every cost in the cost table, at the list price.
- Disclosure: Ryze AI publishes this blog and does not run Jev in production as of 2026-09-19.
Sources
- Introducing System One Models & Jev — TypeSafe AI, Sep 15, 2026Launch post: pricing ($0.042 per 1M input tokens, output free), early access and the System One definition.
- Jev: TypeSafe's System One Model — DataCampAPI shape, the three question types, vendor benchmark table and the stated weaknesses.
- AINews: Jev, a System One model that only decides, classifies, routes and scores — Latent SpaceIndependent summary of the launch and the open questions about the benchmarks.
- awesome-jev-by-typesafe — GitHubCommunity patterns (retrieve candidates, then judge) and the operational limits reported by early users.
Frequently asked questions
What is Jev for meta ads?
Jev is a decision model from TypeSafe AI, released in early access on September 15, 2026. Used for meta ads, it answers bounded questions about rows of account or site data: pick one option, give a score, or return a yes/no probability. It costs about $0.0004 per decision at list price and cannot write text, so it handles the sorting and an LLM handles the writing.
How much does Jev cost in 2026?
TypeSafe lists Jev at $0.042 per 1 million input tokens with output tokens free, which it describes as roughly $0.0004 per decision. That is the early access launch price from September 2026. TypeSafe itself says it cannot prove the price is unsubsidized, so budget on it as today's price and re-check before you scale a job.
How accurate is Jev compared with Claude and GPT?
On TypeSafe's own four-workflow benchmark Jev scores 67.8%, against 73.1% for Claude Opus 5 and 74.1% for GPT-5.6 Sol. The benchmark measures agreement with frontier models and nobody has independently reproduced it yet. Treat Jev as triage-grade: run it behind a confidence threshold and test it against human labels on your own data first.
Is the 724 ads in 40 seconds claim verified?
It is a first-hand claim from Matthew Berman's post on X in Jev's launch week in September 2026: 724 live ads from 37 brands broken down in 40 seconds for about 9 cents of tokens. Nobody has independently reproduced it, and Ryze AI has not run it. Treat it as a demo of the pattern, not a benchmark.
Can Jev read images and video?
TypeSafe describes Jev's input as unstructured text state, not images or video. To tag a Meta Ads creative you first turn it into text: the ad copy, a transcript of the audio and a description of the visual written by a vision model. Jev then answers the tagging questions on that text for a fraction of a cent.
Is Jev an LLM?
TypeSafe says no. It calls Jev a System One model: a new architecture with a parallel sampler that returns every typed answer in a single pass, trained with a method named Reinforcement Learning for Calibrated Decisions. In practice the difference is simple. An LLM produces text token by token, and Jev returns a choice, a number or a probability and nothing else.
What are Choice, Score and Noul in Jev?
They are Jev's three question types. Choice picks one option from a list you provide, for example buyer, researcher or junk. Score returns a number inside a range you set, such as 0 to 100. Noul returns a yes/no probability. Every answer arrives with a calibrated confidence, which is what lets you auto-apply the confident ones and queue the rest.
How do I get access to Jev?
Jev is in early access behind a waitlist at typesafe.ai, with no general availability date as of September 2026. Early users on X report being approved the same day or within a day. You do not need to wait to start: label a sample of your data and build the same loop on Claude Haiku 4.5 or Gemini Flash-Lite, then swap the model.
Can Jev hallucinate?
Jev cannot produce malformed output, because its answers are constrained to the schema you define, and TypeSafe reports a 0% structured-output error rate. It can still be wrong. It may pick the wrong option or give a confident yes where the truth is no. That is why every workflow here uses a confidence threshold and a human approval queue for the middle band.
What can Jev not do?
Jev cannot write text, code, ad copy, page content or an explanation of its own answer. It cannot browse, call tools or read an account by itself, and it only works when the possible answers are known in advance. Strategy, diagnosis, copywriting and anything a client will read still belong to a frontier model or a person.
Does Ryze AI use Jev?
Not in production as of September 2026, because Jev is waitlist-only. Ryze AI runs the same decide-then-act loop on models that are available today: code collects the rows, a model judges them, small reversible changes are applied and larger ones wait for approval. The design lets the decision step be swapped to Jev when access opens.
What is the best way to start using Jev for meta ads?
Pick one high-volume job from this page, hand-label 500 rows of it, and build the loop on a model you can call today with structured outputs. Set the auto-apply threshold from your labels, not from a vendor benchmark. Connect live account data first, since a decision model is only useful when something can act on its answers.
Should a decision model replace my frontier model?
No. The two do different jobs. A decision model such as Jev takes the thousands of small, bounded calls in meta ads, which is where most of the token bill goes. A frontier model keeps the work that needs language: writing, rewriting, diagnosing and explaining. Most of the saving comes from routing each task to the cheapest model that can do it.
Related guides
Ryze AI — Autonomous Marketing
Put the decisions to work on a live account
- ✓Connect Google Ads, Meta, GA4 and Search Console
- ✓Let the agent apply the safe changes
- ✓Review the rest in one queue
2,000+
Marketers
$500M+
Ad spend
23
Countries
