OpenAI Clones TypeSafe's Jev for Faster, Cheaper AI Agents

Overview

During OpenAI’s recent developer conference, CEO Sam Altman disclosed the company’s entry into a specialized sector of artificial intelligence: the Decisions API. This new tool is designed to provide models with a predefined set of options, allowing them to classify inputs or select specific behaviors with remarkable speed and low cost.

The announcement drew immediate attention because its functionality closely mirrors Jev, a model released just weeks prior by San Francisco-based TypeSafe AI. Jev was explicitly engineered for software automation, acting as a super-powered classifier that outputs probabilities rather than generating free-form text. Industry observers view OpenAI’s move as a significant validation of the architectural shift toward faster, cheaper, and more constrained AI agents.

The Decisions API and the Clone Wars

OpenAI’s new API represents a strategic pivot from general-purpose language generation to high-efficiency decision-making. At the event, Altman explained that the API allows systems like the Luna model to focus intensely on a limited set of choices. This includes tasks such as categorizing images or determining specific agent behaviors within a controlled environment.

“By focusing the model on that choice, we can make it extremely fast while keeping capabilities like image understanding, broad language support, and safety protections,” Altman stated.

This approach contrasts sharply with traditional large language models (LLMs), which often operate as black boxes generating lengthy responses. By narrowing the output space, OpenAI aims to reduce latency and computational overhead significantly. The move has sparked what TypeSafe CEO Diogo Almeida jokingly referred to as the beginning of the clone wars on social media.

Almeida, a former OpenAI engineer and co-inventor of reinforcement learning, noted that OpenAI’s interest signals a broader industry realization. He suggested that building in a “System One” compatible manner—representing fast, intuitive thinking—is becoming the future of scalable AI infrastructure, distinct from the deliberate, slower reasoning of “System Two” models.

Why Speed and Cost Matter for Agents

The underlying tension driving this development is the economic reality of running autonomous AI agents at scale. While frontier LLMs are powerful, they remain comparatively slow and expensive for routine software automation tasks. Developers have increasingly turned to models like Jev to augment their workflows, discovering that decision-specific models offer a superior ratio of performance to cost.

TypeSafe’s value proposition rests on its ability to generate statistically useful outputs using synthetic data. Almeida emphasized that simply making a model fast is easy; the true challenge lies in maintaining intelligence per dollar. His company’s moat is built on ensuring these cheap decisions remain highly accurate and aligned with real-world requirements.

Other startups are already rolling out similar specialized models, indicating that OpenAI is not the last tech giant to produce this type of interface. The critical question for the industry is how well these decision models calibrate to complex, real-life scenarios where nuance and context are paramount.

Security and Monitoring Applications

One of the most compelling use cases for this technology is in the realm of AI agent security and monitoring. Following a series of incidents where OpenAI’s own agents behaved unpredictably on the open internet, the company has been forced to implement safeguards. Currently, this involves using a separate, full-scale model to watch for malicious or erroneous actions, a process that incurs significant compute costs.

A model like Jev could potentially revolutionize this monitoring layer by making it far more economical. Shapor Naghibzadeh, a cybersecurity professional leading the startup QueryStory, demonstrated this potential during a recent hackathon. He built a system using Jev to evaluate every action an agent took against its original task parameters.

The demo allowed actions with low confidence of being malicious to pass through, flagged suspicious activities for review, and blocked those with high confidence of being bad. According to Naghibzadeh, such monitoring could have prevented recent high-profile incidents, such as the one involving Hugging Face, at a fraction of the current cost.

Monitoring Method	Estimated Cost per Action
Frontier LLM	$372
Specialized Decision Model (Jev)	$2.94

The dramatic cost difference—approximately one hundredth of the price of using a frontier model—suggests that continuous, granular monitoring of agentic behavior is now economically viable. This level of scrutiny could dramatically improve the reliability and safety of autonomous software agents across the enterprise.

Future Implications

The convergence of OpenAI and TypeSafe’s approaches highlights a maturing AI ecosystem. As agents become more embedded in critical workflows, the demand for tools that prioritize speed, cost-efficiency, and safety over raw generative capability will grow.

While it remains unclear how closely OpenAI’s Decisions API will match Jev’s specific implementations, the market signal is undeniable. The ability to run a lightweight, high-confidence decision model on every single agent interaction offers a new layer of governance. This development suggests that the next generation of AI innovation may not come from making models smarter, but from making them faster, cheaper, and more controlled.