OpenAI's Reasoning Model Goes Mainstream: The New Battle for AI Supremacy

Open-AI has made its Decisions API widely available, stepping into a new corner of the AI market fast, cheap models that make choices ...

Open-AI has made its Decisions API widely available, stepping into a new corner of the AI market fast, cheap models that make choices and report how sure they are. The category was opened by Type-safe, a startup founded by a former Open-AI researcher, whose Jev model has since pushed the company to a valuation of roughly $7.6 billion. Now the lab that once employed its founder is competing for the same developers. The result is a quick, crowded contest over speed, cost, and accuracy.

What Decision Models Are

Type-safe's Jev created a category of deep learning models that act as super classifiers. You give one a set of options, and it quickly picks among them and returns a confidence rating for its choice. These are often called decision models or System One models, after the fast, intuitive mode of thinking in psychology.

Developers noticed quickly that such models can replace large language models in many workflows. Tasks like routing a request, labeling content, or choosing an agent's next action do not need paragraphs of generated text. They need a fast, bounded answer. A decision model delivers that at lower cost and with far less delay.

Why Open AI Built One

Nikunj Handa, the product lead for Open-AI APIs, says his colleagues saw Jev and reacted with something like "we have a million use cases for this." His team sprinted to ship a version of its own.

OpenAI took a practical route. Instead of training a model built specifically for calibrated decisions, the team took Luna, one of its existing models, and reconfigured it to run faster and cost less. Developer documentation and other coverage refer to the model as GPT-6 Luna. Handa is candid about the approach. "Right now, it's just plain old Luna that gets the job done," he said, while adding that Open-AI may add training aimed at these tasks in the future.

The big gain, he says, is speed. Calls that took one to three seconds now come back in roughly 100 to 180 milliseconds.

Early Customers and Real Results

In initial trials, engineers at companies including Box, Data dog, Ramp, and Speak used the API in place of Open-AI's regular models to get higher throughput. Speak builds AI language tutors, and its CTO, Andrew Hsu, said the Decisions API cut the time between actions from 1.4 seconds to 0.13 seconds. For a conversational product, that is the difference between a pause and a natural back-and-forth. Open AI also used the product to improve the responsiveness of its own Dots voice assistant.

A Crowded Field

Open-AI is not the only challenger. Several other companies now offer decision models:

Liquid released D1 Ultra last week.

Base Lab has a model called Elephant-1.

Perplexity offers its own decider model.

Open and local options can run on your own hardware. A tooling update for the local model runner Ollama now serves decision models behind the same style of API.

Developer tooling is racing to keep up. One AI gateway now serves decision models through a single decisions endpoint from eight providers: TypeSafe, OpenAI, OpenRouter, Perplexity, Cloudflare, Azure, Databricks, and Ollama. Requests can fall back from one provider to another if the first fails. Another popular developer toolkit has added an experimental decision function that works with models from TypeSafe, OpenAI, and others. For models that are not natively built for decisions, it adapts a language model's structured output to fit.

This tooling matters because it lowers the cost of switching. Developers can compare models behind one interface and swap them without rewriting their code.

What Makes a Good Decision Model?

Here the picture gets murky. As the report puts it, it is not clear what makes a good decision model beyond "it works for you." The balance among intelligence, cost, calibration, and speed seems to vary by workload. A routing task might favor raw speed. A risk check might favor calibration, meaning the stated confidence matches how often the model is actually right.

Type-safe's CEO, Diogo Almeida, distrusts benchmarks and urges people to try models on their own work. Even so, Type-safe released its own rankings last week alongside the latest Jev model, which adds support for images and "smart-if" statements. Other coverage also mentions a new "box" question type in that release.

What the Benchmarks Say

Public rankings give a mixed picture, and they should be read with care.

Leaders on Two Indices

On the AIM System One Index and the Artificial Analysis Decision Index, D1 Ultra and Elephant-1 tend to rank near the top.

Where Jev and Open-AI Land

The latest Jev model is competitive, placing seventh on one index and third on the other. Open-AI's decision model places eighth and fourteenth.

Real-World Usage

Even so, TypeSafe's models are the most widely queried on Open Router for now, which suggests developers are voting with their traffic.

A Playful Test

Jev tops a benchmark called JevMan, where decision models play Pac-Man. Felix Wunderlich of Opper AI, which built it, says the classic game works as a test because of its nonstop stream of quick, high-stakes decisions: at every junction the model must decide where to go in a few hundred milliseconds.

Treat the rankings as a starting point. They measure specific tasks, and your own task may behave differently.

Where Open-AI May Have an Edge

Open-AI's argument rests on a broader and arguably more capable underlying model, especially for images. For image classification, Handa says Luna is extremely powerful and has been showing better calibration. Customers also want more flexibility to use images as outputs, not just inputs, which he says OpenAI is working toward.

Pricing may be where TypeSafe pushes back. One report notes that TypeSafe lists Jev at $0.042 per million input tokens, which it describes as under half of OpenAI's rate. These are list prices from vendors and can change, so compare current pricing before you decide.

Faster, Cheaper, and Everywhere

Handa sees further speed gains opening new uses. He wonders whether the response time could fall to about 60 milliseconds, at which point a developer could imagine making a decision call for every single frame of a video. That is the kind of use case that is impossible at one second per call and plausible at 60 milliseconds.

The category's name points to an economic idea. Jev is named after the economist William Stanley Jevons, who explained how using a resource more efficiently tends to make it more widely used, not less. Applied here, cheaper and faster decisions should mean software makes far more of them. Handa expects an expanding "crazy set of use cases" that were hiding in plain sight until decision models revealed them.

What Developers Should Do

Test on your own data: Run candidate models on a sample of your real inputs and measure accuracy, latency, cost, and calibration together.
Check calibration, not just accuracy: A confidence score is only useful if it matches reality. Compare stated confidence with actual hit rates.
Set thresholds and review the low-confidence cases: A common pattern is to act automatically when confidence is high and send uncertain cases to a person.
Avoid lock-in: Use a gateway or a common interface so you can switch providers as the leader board shifts.
Pin model versions: Confidence thresholds can change between releases, so lock a specific version in production.
Match the model to the task: Text classification, image classification, and voice use cases may favor different models.
Treat vendor numbers as claims: Speed and price figures come from the companies selling the models.

Open Questions

Several things remain uncertain. Independent testing is still thin, and rankings differ by index. Open-AI's Luna-based approach and Jev's purpose-built approach may behave differently on hard cases, and nobody has shown yet which holds up better in production. The market is also moving fast, with new models arriving weekly, so today's leaderboard may not last. And as decisions get cheap, teams will embed more of them in software, which raises the question of who reviews the results.

The Bottom Line

Open-AI's move confirms that decision models are a real category and not a passing curiosity. Type-safe got there first and still leads in usage, Open-AI brings scale and strong image skills, and several others are climbing the rankings. For developers, the practical outcome is more choice, lower prices, and a strong reason to test models on their own workloads. As these tools get faster and cheaper, expect them to show up in places nobody planned for.

Frequently Asked Questions (FAQs)

What is Open-AI Decisions API?

It is a dedicated API that lets a developer define questions and possible answers, then returns a choice with a confidence rating instead of free-form text. It is built on a reconfigured version of Open-AI's Luna model for speed and lower cost.

How is it different from a normal chat API?

A chat model writes text. The Decisions API picks from options you define, such as classifying content, routing a request, or choosing an agent's next action, and it does so in a fraction of the time.

What is a decision or System One model?

It is a fast model that acts as a super classifier. Given a set of options, it chooses one and reports its confidence. Type-safe's Jev created the category.

How fast is the Decisions API?

Open-AI's product lead says calls shrank from one to three seconds to roughly 100 to 180 milliseconds. One customer, Speak, reported cutting the time between actions from 1.4 seconds to 0.13 seconds.

How does it compare with Jev?

Rankings are mixed. On two public indices, the latest Jev model placed seventh and third, while Open-AI's model placed eighth and fourteenth. Jev is the most widely queried on Open Router for now. Open AI may have an advantage in image tasks.

Which companies tried it first?

Box, Data dog, Ramp, and Speak were named as early customers. Open AI also used it to improve its Dots voice assistant.

Which other decision models exist?

Liquid's D1 Ultra, Base Lab's Elephant-1, Perplexity's decider model, and local options through Ollama are among them.

Should I switch to a decision model?


Loaded All Posts Not found any posts VIEW ALL Readmore Reply Cancel reply Delete By Home PAGES POSTS View All RECOMMENDED FOR YOU LABEL ARCHIVE SEARCH ALL POSTS Not found any post match with your request Back Home Sunday Monday Tuesday Wednesday Thursday Friday Saturday Sun Mon Tue Wed Thu Fri Sat January February March April May June July August September October November December Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec just now 1 minute ago $$1$$ minutes ago 1 hour ago $$1$$ hours ago Yesterday $$1$$ days ago $$1$$ weeks ago more than 5 weeks ago Followers Follow THIS PREMIUM CONTENT IS LOCKED STEP 1: Share to a social network STEP 2: Click the link on your social network Copy All Code Select All Code All codes were copied to your clipboard Can not copy the codes / texts, please press [CTRL]+[C] (or CMD+C with Mac) to copy Table of Content