THE SHORT ANSWER
● Text only, cost-sensitive → Jev. $0.042 per million input against $0.10. Both charge nothing for output.
● Images in the decision → OpenAI. Jev is text only. This is the one difference that cannot be priced around.
● Nobody has published an accuracy comparison. For a product whose output is a calibrated probability, that is the number that matters most — and it does not exist.
● Test on your own labelled data. It is a day of work and it is the only evidence either way.
Side by side
| OpenAI Decisions API | TypeSafe Jev |
|---|
| Input per 1M tokens | $0.10 | $0.042 |
| Output | Free | Free |
| Cache read / write | Free | Not published |
| Inputs accepted | Text and images | Text only |
| Answer types | predicate, choice, score | Boolean, choice, score |
| Latency | "~10x faster" (vendor, unmeasured) | 70–500ms end to end (vendor) |
| Model | gpt-6-luna only | Jev 1.13 |
| Endpoint | POST /v1/decisions | Vercel AI Gateway |
| Status | Public beta, 6 Oct 2026 | Early access |
| Compliance | ZDR, HIPAA, US/EU residency | Not published |
What the price difference is actually worth
Take a million items at 500 input tokens each — a realistic monthly classification volume for a mid-size pipeline. That is 500 million input tokens.
- OpenAI Decisions API: 500M × $0.10/M = $50
- TypeSafe Jev: 500M × $0.042/M = $21
A $29 difference. At ten times that volume it is $290 a month, and at a hundred times it is $2,900.
Which tells you the honest answer: for most teams the price gap is not the deciding factor. Both products already removed the expensive part when they stopped charging for output. Running the same work on a chat endpoint at $0.10 input and $0.50 output would cost substantially more than either. The migration from a chat endpoint is where the money is; the choice between these two is a rounding difference unless you are at serious scale.
Pick on images, latency and compliance. Pick on price only if your volume makes $29 per million items matter — and if it does, you already know it.
The one difference you cannot engineer around
Jev is text only. The Decisions API accepts images, with one condition: they must be inline base64. Hosted image URLs and file IDs are refused.
So if your decision involves looking at something — is this receipt legible, does this photo show damage, which of these four layouts is this document — OpenAI is the only option of the two, and you will be base64-encoding on the way in.
What neither of them will tell you
Neither company has published a head-to-head accuracy benchmark.
That matters more here than it would for a chat model. The entire product is a calibrated probability — a 0.7 that means 0.7, so that when you threshold at 0.8 you get the behaviour you expected. A fast, cheap, badly calibrated probability is worse than an expensive well-calibrated one, because it fails quietly and you only find out downstream.
TypeSafe claims 193.6x faster and 444.6x cheaper than conventional LLMs, on its own workflow evals, and is unusually straight about what that is worth — its funding post says: "We can't prove it isn't subsidized." OpenAI claims roughly 10x the speed of its own Responses API, also unmeasured externally.
Treat both sets of numbers as marketing until somebody independent runs them.
How to decide in a day
- Pull 500 real items from your own traffic and label them by hand. This is the part people skip and it is the only part that produces evidence.
- Run both and compare accuracy at your actual decision threshold, not overall accuracy.
- Check calibration, not just correctness. Bucket the predictions by confidence and see whether the 0.7 bucket really is right about 70% of the time. That is the whole proposition.
- Measure latency from your own region, since both vendor figures are theirs rather than yours.
- Then compare cost — last, because by this point you will usually find it is not the binding constraint.
FAQ
Do both really charge nothing for output?
Yes. The Decisions API has no output, cache-read or cache-write charge; Jev's output tokens are free. Regional processing premiums and long-context multipliers still apply on the OpenAI side.
Can either generate structured JSON objects?
No, and this is the most common misunderstanding. Both produce answers to questions, not populated schemas. For object generation you want Structured Outputs on the Responses API, which is a different product with different pricing.
Which model does the Decisions API run on?
gpt-6-luna only, with no fallback. It has been in public beta since 6 October 2026 and OpenAI expects general availability "in the coming weeks", so field names and limits may still change.
Is TypeSafe a safe bet as a vendor?
It raised $870 million at a $7.5 billion valuation on 9 October 2026, led by Andreessen Horowitz with Sequoia and DCVC participating, and says a third of the Fortune 500 already use Jev. Funding is not durability, but it is runway.
What about open-weight options?
Liquid AI released d1-3B on 8 October 2026 — open weights, typed answers in a single forward pass, zero output tokens, on Hugging Face. If you can self-host, that removes the per-token question entirely.
Sources
Full launch detail: OpenAI Decisions API — $0.10 in, nothing out.
UPDATE — 11 OCTOBER 2026: A THIRD ENTRANT
Microsoft shipped Decision-1 on 9 October at $0.042 per million input tokens with free output — matching Jev to the cent and undercutting OpenAI by 58%.
It also published the accuracy comparison this page said did not exist. See below, with the caveat it deserves.
All three, side by side
| OpenAI Decisions API | TypeSafe Jev | Microsoft Decision-1 |
|---|
| Input per 1M | $0.10 | $0.042 | $0.042 |
| Output | Free | Free | Free |
| Inputs | Text and images | Text only | Text only |
| Context window | Large | Not published | 32,768 tokens |
| Base model | gpt-6-luna | Jev 1.13 | Qwen3.5-9B, post-trained |
| Output format | answers array | Typed values | JSON, one probability per option |
| Latency (vendor) | "~10x faster" | 70–500ms | p50 85ms, p95 125ms |
| Available on | OpenAI API | Vercel AI Gateway | Microsoft Foundry, OpenRouter |
| Published accuracy | No | No | Yes |
The accuracy table, and how much to trust it
Microsoft ran 36 benchmarks across 147,137 questions and nine systems:
| Model | Accuracy | Calibration |
|---|
| Microsoft-Decision-1 | 83.5% | 92.2 |
| Quyet-1.0-Large | 81.9% | 93.1 |
| GPT-6 Luna Decisions | 79.4% | 89.9 |
| H2O-Lightning-4B | 77.2% | 91.8 |
This is a vendor benchmark on which the vendor wins, which is what vendor benchmarks are for. Jev is not in it at all.
The detail that gives it some credibility is the column Microsoft lost: Quyet-1.0-Large beats Decision-1 on calibration. For products whose whole proposition is that a 0.7 means 0.7, calibration is arguably the more important number — and Microsoft published a table where it comes second in it.
Robustness: Microsoft reports the decision flipped on 1.3% of 8-way perturbations on average, and never when options were paraphrased, reversed or shuffled.
The speed claim is contested
Microsoft says Decision-1 is 4.5x faster than Quyet-1.0-Large and 35x faster than GPT-6 Sol at 3.01 seconds. Competitor figures use JevBench's adjusted median, and H2O.ai argues that method inflates measured times. Decision-1 is not on the JevBench board.
Take the p50 of 85ms as a reasonable indication it is fast. Discount the multiples.
What changes in the recommendation
- Text-only and cost-sensitive → Decision-1 or Jev, identical at $0.042. Decision-1 is generally available on Foundry and OpenRouter; Jev is still early access. Availability probably decides it.
- Images → OpenAI, unchanged. Both competitors are text only.
- Long inputs → OpenAI. 32,768 tokens is a real ceiling, and it is the clearest reason to pay the extra 58%.
- Regulated decisions → read Microsoft's model card first. It explicitly excludes sole automated decisions on credit, employment, housing, healthcare and legal rights.
- Want an explanation with the answer → none of the three. Decision-1's card says so outright.
Full launch detail: Microsoft Decision-1 — $0.042, built on Alibaba's Qwen.
Added sources