OpenAI makes its Jev competitor widely available as decision model competition heats up
- ID
- 38405
- Status
- summarized
- Published
- 23 Oct 2026, 3:00 AM
- Fetched
- 23 Oct 2026, 3:29 AM
- Provider
- TechCrunch
- Category
- technology
- Original URL
- https://techcrunch.com/2026/10/22/openai-makes-its-jev-competitor-widely-available-as-decision-model-competition-heats-up/
- Source URL
- https://techcrunch.com/feed/
Summary
- Score
- 7.0
- Created
- 23 Oct 2026, 3:29 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_userssaas_foundersvibe_coders
What happened
OpenAI has made its Decisions API generally available to all customers, a fast, cheap 'super classifier' model that picks among options and returns a confidence rating rather than generating text. It is positioned against TypeSafe's Jev model, which created the category and helped make TypeSafe a $7.6 billion company. OpenAI's Nikunj Handa says the team reconfigured ChatGPT Luna rather than building a purpose-calibrated decision model, cutting inference from 1-3 seconds to 100-180 milliseconds; the API was announced at OpenAI's September 29 developer day, gained voice and visual modalities in the public version, and inference is now geographically targeted for lower latency, with Box, Datadog, Ramp and Speak as early users.
Why it matters
If any of your LLM calls are really just routing, triage, moderation or 'pick one of these' choices, this is a concrete cost and latency swap: 100-180ms versus 1-3 seconds, on plain Luna rather than a specially trained classifier. Speak's CTO Andrew Hsu reports cutting time between actions from 1.4 seconds, so the honest test is to take your slowest classification prompt, run it against the Decisions API, and compare p95 latency and per-call cost against your current model. The article says inference is geographically targeted but does not name which regions, so measure latency from your own deployment location before committing to a migration.
Discussion angle
Which of your current LLM calls are actually decisions rather than generation — and what breaks when you trade a generative model's flexibility for a 100ms classifier with a confidence score?