OpenAI Launches Decisions API to Compete with Jev — the AI Model That Sparked a Developer Gold Rush

TypeSafe AI, a three-month-old startup founded by Diogo Almeida, a former OpenAI researcher, has a US$ 7.6 billion valuation with its Jev model. OpenAI just responded by making its own Decisions API publicly available -- a reconfiguration of ChatGPT Luna focused on speed and cost, not text generation. What seemed like a niche has become, in weeks, the most contested front in the race for cheap intelligence.

What "decision models" are

"Decision models" -- called "System One" by TypeSafe -- do one thing: receive predefined options and return a choice, with a calibrated probability for each alternative. They don't generate text. They are, in practice, deep classifiers with quantitative confidence.

Three primitives make up the vocabulary:

- Choice (select): "Which queue does this ticket go to?" -- picks one label from many. - Noul (proposition): "Does this need human escalation?" -- yes/no with probability. - Score (rating): "How urgent is this message?" -- ordinal score.

Diogo Almeida described Jev as System 1 -- fast, intuitive reasoning, the same concept as Kahneman's TypeSafe blog -- in contrast to general-purpose LLMs, System 2, which "think" before answering and produce long reasoning chains.

Why developers are switching

The discovery was almost accidental: by placing Jev in routing and classification pipelines where LLMs used to live, developers noticed dramatic drops in latency and cost. Speak, an AI language tutor platform, reduced action time from 1.4 seconds to 0.13 seconds by swapping GPT for Jev TechCrunch. Datadog and Ramp are already testing internally.

Torsten Volk, an analyst at Omdia, sums up the problem Jev solves: "Lack of deterministic outcomes and token cost" -- two chronic problems for AI applications in enterprise TechTarget. An LLM generating chain-of-thought reasoning can consume hundreds of tokens just to decide which module to send a customer to; a decision model delivers the choice in milliseconds, for a fraction of the price.

OpenAI moved fast. Nikunj Handa, APIs product lead, told TechCrunch the team saw Jev and thought: "We have like a million use cases for this." Instead of training a model from scratch, they took ChatGPT Luna and reconfigured it to accelerate inference and reduce costs TechCrunch. The result: 100-180 ms latency vs. 1-3 seconds for conventional LLMs.

What the comparison shows

The "decision model" field now has four serious competitors, each with distinct advantages:

| Model | Provider | Best for | |---|---|---| | Jev | TypeSafe | Long text, low cost | | OpenAI Decisions | OpenAI | Multimodality, existing ecosystem | | Clef / Clef-flash | Cloudflare | Raw speed, lowest cost | | Microsoft-Decision-1 | Microsoft | Azure workflows |

DataCamp published a detailed independent comparison DataCamp: in real latency, OpenAI Decisions and Jev were practically tied (268 ms vs. 269 ms). On price, Clef-flash won by a margin ($0.038 per million input tokens vs. $0.10 for OpenAI and $0.042 for Jev). But Jev remains leader in context window (64K tokens) while Clef-flash caps at 24K.

For image classification, OpenAI has the edge: "Luna is extremely powerful and has shown better calibration" for the task, Handa said TechCrunch. Jev has supported images since version 1.14. Clef-omni adds audio and video.

Benchmarks and the metrics confusion

Here's the trap: public benchmarks favor different models depending on the index. On the AIM System One Index and the Artificial Analysis Decision Index, D1 Ultra (from Liquid AI) and Elephant-1 (from Base Lab) lead. Jev 1.13 ranked seventh and third respectively; OpenAI's Decisions Model fell eighth and fourteenth TechCrunch. But on JevBench -- the community's most popular benchmark -- Jev is the reference itself (score 77.1 vs. 73.5 for OpenAI). Diogo Almeida has no patience for benchmarks: "Test them in your own work," he says TechCrunch. The Decisions API, an independent platform, shows that teams testing internally often reach conclusions different from public tables Decisions API.

Academia is already tracking the field. An arXiv paper compared JEV with GPT-6 Luna across seven political science replications and found: JEV is faster (0.09-0.27 seconds vs. 0.75-1.04), but not necessarily cheaper -- Luna at Batch rates costs less in six of seven scenarios arXiv. The speedup is real; the discount is relative.

What to watch

OpenAI already uses its own API to improve Dots, its voice assistant, making it more responsive TechCrunch. TypeSafe launched image support and "smart-if" statements in Jev 1.14. Cloudflare's Clef-flash cut pricing to US$ 0.038/million tokens. The field is moving at hurricane speed.

What matters is not which model wins the benchmark today, but the cascade effect: by making the decision itself cheaper -- routing logic, classification, escalation -- each heavy LLM call gets avoided at dozens of points per application. The expansion of demand for cheap intelligence is exactly what the Jev name itself honors: the Jevons effect, where efficiency in a resource makes it more abundant, not more replaced.

If decision models keep getting cheaper and faster, the next six months will reveal that many of the "use cases" Handa predicted are waiting for a millisecond response, not a text.

---

Article based on independent sources. All source links are verifiable and registered in the source manifesto.

Sources: TechCrunch, TechTarget, DataCamp

✓ Independent sources cross-checked and verified before publishing