Your AI Bill Isn't Mostly Thinking. It's Mostly Deciding.
A support ticket lands in the queue. Billing or bug? Is this the customer with the four-hour response guarantee? Angry enough to jump the line? And which of your nine queues does it go to?
Four decisions, and not one sentence written yet. If a frontier model is doing that step, you are paying an essay writer to fill in multiple-choice bubbles, and waiting a second or two while it spells out each answer a token at a time.
On 15 September 2026, TypeSafe AI launched a model that does only the multiple-choice part.
What Jev actually is
Jev generates no text. You hand it a state, either plain text or JSON, along with a set of typed questions, and it returns one typed answer per question: a yes or no with a probability, a pick from a list of options you defined (up to 255 of them), or a position on a scale you defined. Each answer comes back with a confidence score, and they all come back in a single parallel pass rather than token by token. The API is one endpoint, POST /v1/systemone, with client SDKs on top.
The interesting structural consequence is that because the output space is restricted to types you defined up front, Jev cannot hallucinate a category that does not exist and cannot hand back a type error. That is not a claim about it being right. It is a claim about the shape of what comes back, which is a different and smaller promise, and a genuinely useful one if you have ever had a routing step return a queue name your system has never heard of.
TypeSafe calls this a “System One model,” borrowing Kahneman’s split between fast reflexive judgment and slow deliberate reasoning. The company raised $40 million, and co-founder and CEO Diogo Almeida is a former OpenAI researcher and a co-inventor of RLHF. Jev was trained with what the company calls reinforcement learning for calibrated decisions. It is currently in limited closed beta, which matters and we will come back to it.
The job list is narrow and clear: classify, route, score, detect, rank, extract, verify, gate. High-volume repeated decisions over a shared state where the set of possible answers is known before you start. Named uses include support ticket routing, invoice handling, security alerts, resume screening, lead and vendor and claim scoring, sentiment and policy checks across large tables, and screening the output of AI agents for jailbreaks. LangChain has already added it to the agent control loop.
The money, with the assumptions on the table
List price is $0.042 per million input tokens. Output tokens are free. Call it four hundredths of a cent per decision. Most calls come back in about 100 milliseconds, and the quoted range tops out at 500.
The models you would otherwise point at the same job, per million tokens in and out:
- Claude Haiku 4.5: $1.00 / $5.00
- GPT-5 mini: $0.25 / $2.00
- Claude Sonnet 5: $2.00 / $10.00
Three worked examples. The assumptions are stated so you can swap in your own numbers.
Triaging support tickets. 20,000 tickets a month. Each one runs about 700 tokens going in and 40 coming back, so 14M and 0.8M:
- Jev: $0.59 a month
- GPT-5 mini: $5.10
- Claude Haiku 4.5: $18.00
- Claude Sonnet 5: $36.00
Per ticket, that is roughly 100 milliseconds against one to three seconds.
Scoring a backlog. Say you have 500,000 product reviews and you want every one of them checked for sentiment and for policy violations. Figure 400 tokens in per review and 30 back. That is 200M and 15M:
- Jev: $8.40
- GPT-5 nano: $16.00
- GPT-5 mini: $80.00
- Claude Haiku 4.5: $275.00
- Claude Sonnet 5: $550.00
Worth saying plainly: against the cheapest text models the gap closes to about 2x, not the 400x TypeSafe advertises. The headline multiple compares Jev against frontier models that nobody should be pointing at a sentiment pass anyway. A 2x saving on a backlog job is real but it is not a reason to rebuild anything.
Screening agent output. Now put a check on every agent response before it reaches a customer. 100,000 turns a month at 1,200 tokens in and 20 out, so 120M and 2M:
- Jev: $5.04 a month
- GPT-5 mini: $34.00
- Claude Haiku 4.5: $130.00
- Claude Sonnet 5: $260.00
Cost is the less interesting half here. A guardrail taxes every response you send, which makes latency the thing that decides whether the check survives. At 100 milliseconds nobody notices it. At a second and a half, someone eventually says the assistant feels sluggish, and the check is what comes out.
Nobody decides to stop screening output. They decide to make the thing faster, and screening is what goes.
The evidence is thinner than the launch
TypeSafe’s own numbers are up to 193.6x faster and 444.6x cheaper than frontier models, with 76.0% accuracy in its eval tables. Those are self-reported, and The Register noted that the eval scored Jev against two other models’ answers rather than against an independent ground truth. No large-scale independent reproduction has surfaced yet.
What independent testing exists is small and points in two directions. One third-party test put Jev at about 5x faster and 8.6x cheaper than Mistral Small 4, with accuracy at 67.8% against 74.1% for the best comparison model, which the tester summed up as overwhelmingly better on speed and cost and on par or slightly worse on accuracy. A UK events company called Near Here ran 50 real listing-moderation decisions and got 96% for Jev against 84% for Mistral Small 4. A developer reported a flight search running in 7 seconds for $0.0039. Fifty decisions is a promising afternoon, not a benchmark. There are useful walkthroughs at DataCamp and Classmethod if you want to see the shape of the thing in code.
So: a well-funded launch, a credible founder, real architectural logic, self-reported benchmarks, mixed small-sample outside results, and a closed beta you probably cannot get into this quarter. Treat it as a signal about where inference is heading rather than something on your shortlist.
The part you can act on this week
The useful idea here does not depend on Jev shipping. It is that most real AI workflows are a handful of genuine reasoning steps wrapped in hundreds of small decisions, and it is the small decisions, not the reasoning, that make those workflows slow and expensive once volume shows up. A chat model produces an answer for a person to read. A decision model produces a value your system branches on. Nearly every team we talk to is buying the first thing and using it to do the second.
Take one workflow you already run through a model and go call by call. For each one, ask a single question: is the output a sentence a human reads, or is it a value something downstream branches on? Tally the two piles. If the branching pile is most of the calls and most of the tokens, you have found the part of your AI spend that is not buying you any thinking, and that will be true whichever model you eventually point at it. Confidence scores are the other half of the design: the reason to separate the piles is so the uncertain cases can escalate to a person or to a reasoning model, and the routine 90% can run cheap and fast without one.
None of that requires access to a closed beta, a migration, or a new vendor. It requires knowing what your workflow is actually asking for, which is the same problem underneath most AI spend that disappoints. If you want a structured way to look at that across the business rather than one workflow at a time, the AI Readiness Guide is free and self-administered, and it will tell you honestly whether this is your bottleneck or whether something earlier is.