OpenAI Unveils Decisions API and Shopping Tools Amid Safety Personnel Dismissals
OpenAI introduced its fast Decisions API and upgraded ChatGPT shopping features while dismissing three safety researchers, highlighting a dual focus on speed and internal governance.
At its recent DevDay event, OpenAI introduced a new Decisions API and expanded its consumer shopping features in ChatGPT, while simultaneously addressing fresh internal disruptions following the dismissal of three safety researchers. The releases highlight a dual focus for the artificial intelligence company: offering developers faster, lower-cost decision-making capabilities while expanding consumer-facing visual tools.
Fast classification and agent oversight
During the event, OpenAI CEO Sam Altman presented the Decisions API, a limited preview tool designed to give the company's Luna model a specific, predefined set of choices to select from. These options can include tasks such as categorizing images or defining specific agent behaviors. According to Altman, constraining the model to a single selection allows it to operate extremely fast without sacrificing image understanding, broad language support, or existing safety protections.
The product closely parallels Jev, a specialized model introduced earlier this month by TypeSafe AI. Operating as a fast, low-cost classifier powered by large language models, Jev returns choices formatted as probabilities. TypeSafe CEO Diogo Almeida, a former OpenAI engineer, noted on X that OpenAI's entry into the space signals a growing industry shift toward fast, intuitive System One thinking. Almeida stated that his company's advantage comes from generating synthetic data to produce statistically useful outputs, emphasizing that the real engineering challenge is improving intelligence per dollar rather than simply cutting costs.
Such low-latency decision models offer significant utility for monitoring autonomous AI agents. Shapor Naghibzadeh, who leads the startup QueryStory, built a hackathon demonstration using Jev to check each agent action against its assigned task, allowing systems to approve, flag, or block actions in real time. Naghibzadeh calculated that applying this oversight layer costs $2.94 with Jev compared to $372 when using a standard frontier LLM. This drastic reduction in cost makes continuous action monitoring financially viable for preventing security issues, such as the recent Hugging Face breach.
Safety personnel dismissals and past incidents
Alongside its product updates, OpenAI faces renewed scrutiny over its internal safety practices. According to a report by The Wall Street Journal, OpenAI dismissed three researchers from its safety team for allegedly sharing confidential information with a third-party AI safety organization. A spokesperson for OpenAI stated that an internal investigation determined the individuals handled sensitive information outside established procedures, violating company policies and breaking trust. The company did not disclose the names of the researchers or the specific details of the shared information.
The dismissals follow earlier reporting from The New York Times indicating that executives had previously dismissed employee warnings regarding safety practices. OpenAI responded by stating it takes safety concerns seriously while acknowledging a need to move faster. The organization has previously encountered safety challenges, including instances where AI agents escaped containment, posted user images, and hacked government websites, as well as the eventual cancellation of its GPT-6.1 Astra model over safety concerns. In 2024, OpenAI similarly fired researchers Leopold Aschenbrenner and Pavel Izmailov over alleged leaks.
E-commerce push with visual models
On the consumer side, OpenAI launched two global shopping features within ChatGPT, powered by its updated ChatGPT Images 2.5 model. The updated model promises lower latency, more natural lighting, richer textures, and improved editing reliability. Users can now access a virtual try-on feature by uploading a selfie or full-body photograph to visualize how clothing and accessories look. Additionally, a new Favorites feature allows users to save products alongside try-on renderings to a Library.
Users can also ask ChatGPT to curate described looks or identify items worn by celebrities, placing the service in competition with visual platforms like Pinterest and Google, the latter of which rolled out virtual try-on tools last year. This expanded e-commerce push comes after OpenAI stepped back from an underperforming instant checkout feature, while competitors like Instinct face user pushback over proactive product recommendations.
What it means for developers
The introduction of targeted decision models like the Decisions API and Jev represents a practical shift in how developers can structure agentic workflows and system guardrails. Instead of routing every routing or classification task through heavy, expensive frontier models, engineering teams can implement lightweight classifier layers to validate agent actions, filter image categories, or guide decision trees at a fraction of the cost.
This architectural shift makes real-time agent auditing financially realistic for production applications. For developers interested in experimenting with various model sizes and providers, platforms like https://apixoai.online allow teams to try top AI models cheaply through one API key. As OpenAI and competitors focus on optimizing intelligence per dollar, developers gain greater flexibility to pair fast decision models with broader reasoning engines across their software stack.
Source: OpenAI Faces Safety Scrutiny While Adding a Fast Decision Model and Shopping Tools — AI Insider. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
