OpenAI's Decisions API cut Speak's wait from 1.4 seconds to 0.13, but Jev missed one bad answer in six

There's a new kind of model showing up in your stack, and it doesn't write anything. It reads a ticket, picks from a short list, and hands back a number with a confidence score. Fast, cheap, no text to parse. The big labs are racing to own it, and your engineers are already trying it on your classification jobs. Let me catch you up.

The deep cut

• A decision model picks, it does not write. Jev, Clef, and OpenAI's Decisions API return a typed answer with a confidence score, not prose.
• Speed is the whole pitch, so prove it on your data. Speak cut time between actions from 1.4 seconds to 0.13 using OpenAI's Decisions API.
• Cheap verdicts are expensive when they are wrong. Jev accepted 16% of wrong answers as a judge while Claude Sonnet caught nearly all of them.

What a decision model actually does

A normal language model writes text one word at a time. A decision model does not write at all. You hand it a state, like a support ticket, and a set of options, and it scores each one in a single pass. Cloudflare's Clef returns a number like 0.9917 for "is this urgent," with zero output tokens. Your code branches on the number. Nothing to parse.

This started with TypeSafe's Jev, which hit 35% of paid teams on Vercel inside two weeks, the fastest adoption Vercel had ever logged. Vercel even swapped out GPT-6 Luna for Jev to classify harmful commands. The category is now crowded with Clef, Amazon's entry, and OpenAI's own version.

The speed is real and it opens new jobs

The pitch is latency. Nikunj Handa, who leads OpenAI's APIs, told TechCrunch the Decisions API shrinks response times "from one to three seconds to 100, 120, 180 milliseconds." Speak, which builds language tutors, cut the gap between actions from 1.4 seconds to 0.13.

At that speed, work you never bothered to automate becomes cheap. Handa is already asking whether 60 milliseconds would let you run a decision on every frame of a video. That is the real upside here, not a better answer, but answers fast enough to drop into places a full LLM was too slow and too costly to sit.

One honest note from the vendor himself: TypeSafe's CEO Diogo Almeida distrusts benchmarks and tells people to trial the models on their own work. The public indexes disagree with each other. Measure from your own region, on your own tickets.

Where cheap verdicts get expensive

A fast wrong answer is still wrong, and the cost moves downstream. One test put Jev in a judge's seat against Claude. One-shot, Jev tied Claude Sonnet on binary correctness. But turn thinking on, and the Claude judges hit 97% and above, about 11 points over Jev. Jev's false-accept rate was 16%. Sonnet's was under 2%.

Read that gap in plain terms. Used as a gate, Jev let one bad answer in six slip through. For routing a ticket, fine. For deciding which model version ships or which response passes a safety guard, that error rate is a problem you inherit every time. The catch is the task, not the model. Jev's confidence scores were well calibrated, so you can route the shaky cases to a stronger judge instead of trusting every verdict.

Pick the right small tool for the job

Decision models are not your only cheap option, and the lines matter. Claude Haiku 5.5 is a small LLM built for the same high-volume work, classification, extraction, summaries, at $0.10 per million input tokens, with a sub-agent pattern where Haiku workers feed a Sonnet lead. It writes text, so it fits jobs that need a few words out, not just a label.

Clef splits the difference on price and openness. Clef-flash now runs at $0.038 per million input tokens, under Jev's $0.042, with weights public under Apache 2.0. But the knowledge benchmarks tell you the boundary: Jev beat Clef 78 to 48 on GPQA Diamond. A decision model is not a reasoner. If the question needs hard thinking, these are the wrong tool.

Three questions for your team

• Which of our current LLM calls are really just classification or routing, and what would they cost on a decision model instead of per-token prose?
• Where have we put a cheap model in a gate that blocks a release or a safety check, and what is its false-accept rate on our own data?
• If we standardize on one of these, how locked in are we, given Replit dropped Anthropic as default over cost and performance and models keep getting retired out from under teams?