OpenAI launches Decisions API to compete with TypeSafe's Jev model

OpenAI has officially released its Decisions API to all customers, marking a significant entry into the rapidly growing market for high-speed, low-cost decision models. This move directly challenges TypeSafe, a startup valued at $7.6 billion, which pioneered the category with its Jev model. The release follows an initial announcement at OpenAI’s developer day on September 29 and aims to provide developers with a faster alternative to traditional large language models for specific classification and choice-making tasks.

What happened

The Decisions API is designed to handle workflows where speed and cost are critical constraints. Unlike standard LLMs that generate text through complex reasoning processes, these so-called System One models act as super classifiers. They evaluate a set of options and return a chosen path along with a confidence rating. Nikunj Handa, the product lead for OpenAI’s APIs, noted that internal teams recognized the potential immediately after seeing TypeSafe’s Jev model, identifying millions of potential use cases within their own infrastructure.

Rather than training a new model from scratch, OpenAI’s engineering team reconfigured ChatGPT Luna to optimize for inference speed and reduced cost. This approach allowed them to deploy the service quickly while leveraging existing capabilities. The public version includes support for voice and visual modalities, and OpenAI has distributed its inference resources geographically to lower latency for users around the world. Early trials with companies such as Box, Datadog, Ramp, and Speak demonstrated significant performance improvements, with some engineers replacing slower LLM calls entirely.

How it works

The core mechanism behind the Decisions API is the optimization of an existing foundation model for rapid classification rather than generative output. By stripping away the overhead associated with long-form text generation, the system can process inputs and return decisions in milliseconds. Handa explained that the primary benefit is the reduction in response time, shrinking delays from one to three seconds down to ranges between 100 and 180 milliseconds. Currently, the model relies on the underlying power of Luna without specialized training for decision-specific logic, though future iterations may include targeted fine-tuning.

This architecture supports multi-modal inputs, allowing the model to process images and voice data alongside text. For image classification tasks, Handa stated that Luna demonstrates strong calibration capabilities. The system is also being tested for even lower latencies, with goals of reaching 60 milliseconds. At that speed, the API could potentially handle decision-making for every frame in a video stream, opening up real-time interactive applications that were previously impractical due to computational lag.

Key details

• OpenAI’s Decisions API reduces inference time from seconds to approximately 100–180 milliseconds.
• The model is built by reconfiguring ChatGPT Luna rather than training a new specialized architecture.
• Early adopters like Speak reported reducing action times from 1.4 seconds to 0.13 seconds.
• The API supports voice and visual modalities, with geographic resource distribution to minimize latency.
• TypeSafe’s Jev model remains the most queried decision model on OpenRouter despite new competition.
• Public benchmarks show OpenAI’s model ranking eighth on the AIM System One Index and fourteenth on the Artificial Analysis Decision Index.

Why it matters

For software engineers and technical founders, the emergence of decision models represents a shift in how AI is integrated into production systems. Traditional LLMs are often overkill for tasks that require simple choices, such as routing customer support tickets or classifying user intent. By offering a dedicated API for these tasks, OpenAI enables developers to reduce costs and improve user experience through faster response times. This efficiency allows AI to be embedded in more frequent, high-volume interactions without prohibitive expense or latency.

The competition in this space is intensifying, with multiple players releasing specialized models. Liquid recently launched D1 Ultra, and Base Lab released Elephant-1, both of which perform well on public indices. TypeSafe continues to innovate with features like smart-if statements and image support. However, the practical value of these models often depends on specific workload requirements rather than benchmark scores alone. As Diogo Almeida, CEO of TypeSafe, suggests, the best model is the one that works for your specific use case, emphasizing the need for direct trial and evaluation.

What you can do

• Audit your current LLM workflows to identify tasks that involve simple classification or choice selection.
• Test the Decisions API against your existing LLM calls to measure improvements in latency and cost.
• Evaluate multi-modal capabilities if your application processes images or voice data alongside text.
• Compare performance metrics with competitors like TypeSafe’s Jev, Liquid’s D1 Ultra, and Base Lab’s Elephant-1.
• Consider real-time applications, such as video frame analysis, that may become feasible with sub-100 millisecond latency.
• Monitor benchmark results but prioritize internal testing to determine which model best fits your specific operational needs.