Jev vs OpenAI Decisions API
OpenAI shipped a Decisions API at DevDay 2026: constrained picks on GPT-6 Luna from questions with finite answers, aimed at classify / route / agent next action. TypeSafe's Jev launched into the same category two weeks earlier. Jev is generally available with a documented API, probabilities and confidence; Decisions is in public beta with vision support. Since Jev 1.14 on 14 October 2026, Jev reads images too, as a preview feature. This is category validation, not a funeral.
Updated · by Jon Kraayenbrink
In short
- Jev: GA with documented Choice / Score / Noul, per-option probabilities, confidence, $0.042/M input tokens (output free).
- Decisions: public beta (Oct 6, 2026), powered by GPT-6 Luna, supports text or images, available on OpenRouter. No public schema or Decisions-specific pricing yet.
- Vision is no longer only OpenAI's: Jev reads PNG and JPEG in preview, and adds a Box answer. Jev's clear edges: published contract, multi-question types, confidence scores.
- Use both when it helps. Jev for calibrated gates today; test both on your own images before either one acts on a picture.
What OpenAI announced
From the DevDay 2026 recap:
Decisions API enables real-time decision-making by focusing Luna's intelligence on a specific set of user-defined questions with finite pre-defined answers. Developers supply context using text or images, and get back answers they can use to classify content, route requests, or choose an agent's next action. Entered public beta on October 6, 2026, with general availability expected in the coming weeks.
Primary launch posts:
- @OpenAIDevs — Luna, finite answers, classify / route / agent next action, public beta
- @thsottiaux — constrained decisions, visual inputs, less than a few hundreds of milliseconds end to end
Secondary explainers (not OpenAI docs): Firecrawl comparison, Vercel explainer
What Jev is
TypeSafe's Jev is a decision model (System One), not a chat model that happens to pick a label. Launched September 15, 2026.
- Question types: Choice, Score, Noul (mixable in one call)
- Returns: per-option probabilities and confidence
- Input: text (string, JSON object, or array of strings), plus PNG and JPEG images in preview on 1.13 and 1.14 with the
TypeSafe-Danger-Zone-Imagesheader. No audio or video - Box: a fourth answer type in Jev 1.14 that returns one bounding box on an image, plus confidence (preview)
- API: documented
POST /v1/systemonewith Python and JavaScript SDKs - Pricing: $0.042 per million input tokens, output free. Images are billed as input tokens, with no fee per image
- Limits: 64k tokens per request and 32k per question by default; 262k for both with the opt-in long-context header (preview)
- Versions: jev-1.13.0 and jev-1.14.0. jev-preview is 1.14, and jev-latest moves from 1.13 to 1.14 around 21 October 2026
Sources: TypeSafe models, image docs, Box docs, Jev 1.14 post
Side by side
Facts only. When OpenAI has not published a detail, that is what the table says.
| OpenAI Decisions API | TypeSafe Jev | |
|---|---|---|
| Announced | Sep 29, 2026 (DevDay) | Sep 15, 2026 |
| Availability | Public beta | Generally available |
| Engine | Constrained GPT-6 Luna | Purpose-built System One |
| Public contract | Not published (as of Sep 30) | Documented POST /v1/systemone + SDKs |
| Question types | Finite pre-defined answers (details unpublished) | Choice, Score, Noul |
| Probabilities / confidence | Not in official docs (some press claims) | Per-option probs + confidence |
| Input modalities | Text or images | Text; PNG and JPEG images in preview (header). No audio or video |
| Bounding box answer | Not published | Box type, Jev 1.14 only (preview) |
| Context length | Not published | 64k per request, 32k per question; 262k opt-in (preview) |
| Named jobs | Classify, route, agent next action | Same jobs + score / noul workflows |
| Latency | “Few hundreds of ms” (OpenAI statement) | 70–500 ms (TypeSafe docs) |
| Price | Decisions-specific: unknown | $0.042 / M input; output free |
Latency caveat: Do not treat a vendor slide (“~150 ms” cited in secondary coverage of DevDay) and third-party OpenRouter telemetry (P50 ~0.21 s for Jev, per Firecrawl) as the same benchmark. Label methods when comparing numbers.
Price caveat: Jev's documented price vs Decisions with a big asterisk until OpenAI publishes a Decisions-specific line item. Luna list prices ($0.10/$0.50 per million tokens in secondary coverage) are not the same as Decisions billing.
What builders should do
- If you need vision (screenshots, UI state, computer-use frames): both now take images, and both are in preview for it. Try Jev's image header on your own labelled images; for production, preprocess to text (OCR / DOM / captions) until one of them calls vision production-ready. More on Jev and images.
- If you need calibrated gates (confidence thresholds, Score, multi-question fan-out): Jev's published surface is the control plane you can write policy against today.
- If you need GA with no waitlist: Jev. Decisions FOMO is real on X; shipping is better marketing than waiting.
- If you're building demos: put the answer space on the card. OpenAI's whole pitch is finite options. Make the option list visible on madewithjev builds so visitors instantly get the category.
- Name collision: OpenRouter already branded a “Decisions API” path that can serve Jev. Say OpenAI Decisions vs OpenRouter Decisions vs TypeSafe System One / Jev so readers don't fuse three products.
Builds that show the job
Routing, classification, and agent next action — with the cost and latency each builder published.
Duncan
@ephraimduncan
Built a model router with Jev by @typesafeai. Jev decides what model fits your request best and the request is sent to that model.
XRouting and model choice
A model router on Jev
Gregor Zunic
@gregpr07
Breaking: Browser Use + Jev = Ultrafast ⚡ Findings flights took 7s and cost only $0.0039 🤯 > new action space every step > DOM state space > small LLM fallback to type (this video is at 1x speed btw) Built a tiny open source browser agent. try it below ↓
tamara
@tamarajtran
found the perfect use case for @typesafeai Jev: instant compaction in 2026, why is compaction still a summarization prompt? Jev can make it instant by scoring every tool call and dropping what’s irrelevant
XContext and memory
PickInstant compaction for Claude
Hassan
@nutlope
I used Jev to classify 1,018 AI research papers. The result: $0.08 total cost and 256ms median end-to-end latency per paper. The pipeline was: 1. Summarize each paper with DeepSeek V4 Flash 2. Send the title + summary + 24 possible topics to Jev 3. Use Jev to classify each paper 4. Visualize everything on http://1kpapers.com The summaries cost $3.99 on @togethercompute. The classifications cost $0.08 on @typesafeai. So for just over $4 of inference, I ended up with a pretty useful way to explore the top AI research papers from the past year. I think this is where things are heading: different models for different parts of the workflow, instead of using one model for everything. I’m running evals on the Jev classifications before replacing the current ones, but the site is already live: http://1kpapers.com
More on Jev use cases.
Try it today
Next steps
- OpenAI Decisions alternatives — who should use Jev instead, and when Decisions still wins
- How to use Jev — your first call, SDKs, and gateways
- Jev pricing — what a decision costs across real builds
- Jev vs an LLM — same job both ways, with published numbers
- Agent prompts — wire Jev into Claude Code, Cursor, Codex
Common questions
- What is OpenAI's Decisions API?
- A constrained decision-making API on GPT-6 Luna. You define questions and finite answer options; it picks one using text or image context. Announced at DevDay 2026 for classify, route, and agent next action jobs. Entered public beta on October 6, 2026, and is available on OpenRouter.
- How is Jev different from OpenAI Decisions?
- Jev is generally available with a documented API (POST /v1/systemone), three question types (Choice, Score, Noul), per-option probabilities and confidence scores, and published pricing ($0.042/M input tokens, output free). Decisions entered public beta on October 6, 2026, but has no public schema or published pricing yet.
- Is OpenAI Decisions generally available?
- Not yet. As of October 2026, it is in public beta. OpenAI expects GA in the coming weeks.
- Does OpenAI Decisions support images?
- Yes. OpenAI states developers can supply context using text or images. Jev now does too, in preview: since 14 October 2026, Jev 1.13 and 1.14 accept still PNG and JPEG images with the header TypeSafe-Danger-Zone-Images: true. TypeSafe says the image feature is not ready for production yet, and neither model takes audio or video.
- Can Jev read images now?
- Yes, in preview. Turn it on per request with the TypeSafe-Danger-Zone-Images: true header on jev-1.13.0, jev-1.14.0 or jev-preview. Images are billed as input tokens at $0.042 per million, with no fee per image. Jev 1.14 also adds a Box answer type that returns one bounding box on the image.
- Does Jev return probabilities and confidence?
- Yes. Every Choice answer includes a probability for each option and a confidence score. Your code can act on confident answers automatically and escalate the rest. OpenAI has not published whether Decisions returns probabilities; some press claims exist but are not in official docs as of September 30, 2026.
- Which should I use for agent routing today?
- If you need calibrated confidence thresholds, multi-question batching, or GA access today, use Jev. If you need vision in production, test both: Jev's image input is a preview feature, and so is Decisions as a whole. For text-based routing with GA availability, Jev is documented and ready.
Made with Jev is independent and not affiliated with TypeSafe AI. Every figure on this page is the one its author published, linked to where it can be checked.