A narrow product with a wide blast radius
OpenAI first previewed the Decisions API at its September 29 developer day, and it is now generally available with added voice and visual inputs and markedly lower latency, achieved through geographically targeted inference. According to Nikunj Handa, OpenAI's API product lead, the team got there by reconfiguring ChatGPT Luna for faster, cheaper inference rather than training a dedicated model from scratch, a telling choice about where OpenAI sees the margin in this category. The product targets a specific job: routing a high volume of narrow, structured decisions, approve or deny, escalate or resolve, which model or workflow to call next, cheaply enough and fast enough that a business can run it on every transaction rather than a sampled subset.
The numbers are the story. Response times dropped from a range of one to three seconds down to 100 to 180 milliseconds. Speak's CTO Andrew Hsu said the API cut the time between conversational actions from 1.4 seconds to 0.13 seconds in his product, a difference that moves voice AI and real-time transaction routing from 'noticeably laggy' to 'usable in production.' Box, Datadog, and Ramp have all trialed the product, which tells you this is being evaluated as infrastructure, not a novelty feature.
The competitor OpenAI can't out-brand
What makes this GA launch more than a routine product update is who OpenAI is competing against. TypeSafe, a startup founded by a former OpenAI researcher and now valued at $7.6 billion, built its business specifically around fast, low-cost 'super classifier' decision models, choosing between options and attaching confidence ratings, through its Jev model. On two leading public benchmark indices, Jev currently ranks seventh and third, while OpenAI's own decision model ranks eighth and fourteenth. Liquid's D1 Ultra and Base Lab's Elephant-1 lead the rankings outright.
TypeSafe CEO Diogo Almeida's advice to buyers, test models on your own workloads rather than trust the leaderboard, is sound and also self-serving: benchmark rankings on narrow classification tasks often don't predict real-world performance, and a smaller specialized vendor benefits every time a buyer discovers that the giant's general-purpose model underperforms a focused one on a specific job. For CTOs evaluating this category, the practical lesson is to run your own bake-off on your actual decision-routing workload before committing, because the brand name leading the AI conversation is not leading this particular leaderboard.
Uber bets on breadth over depth
Separately, and more significant for anyone managing enterprise AI rollout strategy, Uber is expanding ChatGPT Work, OpenAI's enterprise agent product introduced in July, from a 1,000-employee pilot to its entire workforce of roughly 30,000 people. Uber VP of Engineering Sumanth Sukumar said about 80 percent of the pilot's early users were non-engineers, and that they used the tool to analyze operational data, localize training materials, and build dashboards, tasks that sit well outside what most enterprises initially imagine as the use case for an AI coding or agent tool.
Uber says it will track usage and outcomes to guide the next phase of its AI deployment, though it has not disclosed cost. The non-engineer adoption pattern is the detail worth sitting with: it suggests the highest-value early use case for general-purpose enterprise agents is not developer productivity, where most vendor pitches are aimed, but operational and analytical work done by staff who previously had to request a report or a dashboard from someone else and wait. That pattern should reshape how large operational employers, logistics, retail, hospitality, prioritize rollout, since the biggest early win may sit in operations and support functions rather than in the engineering org that usually gets first access to new tooling.
What the pairing tells you about OpenAI's strategy
Taken together, these two moves show OpenAI fighting on two fronts at once: a developer-infrastructure front, where it's racing narrow, well-funded specialists on latency and cost per decision, and an enterprise-deployment front, where it's using a flagship customer's full-company rollout as proof that ChatGPT Work scales past a pilot. Both are necessary. Infrastructure wins without enterprise logos don't generate revenue at the scale OpenAI needs to justify its compute commitments, and enterprise logos without competitive infrastructure economics eventually get undercut by a cheaper, faster alternative.
The risk for OpenAI is that these two fronts pull in different directions. The Decisions API GA is explicitly about being fast and cheap at narrow classification tasks, a commodity game where TypeSafe, Liquid, and Base Lab are already ahead on independent benchmarks. ChatGPT Work's pitch to Uber is about broad, flexible utility across unpredictable use cases, a premium game where the competition is Anthropic and Microsoft's enterprise agent offerings rather than narrow decision-model startups. Winning the commodity game on price while winning the premium game on capability inside the same company is a tension every major cloud and infrastructure vendor has eventually had to resolve with separate product and pricing lines, and OpenAI is now far enough into both markets that it will have to do the same.
The roadmap implication
If you're evaluating decision-routing infrastructure for a latency-sensitive product, voice, real-time fraud scoring, dynamic pricing, treat this moment as the point where sub-200-millisecond decision inference became a baseline expectation rather than a differentiator, and benchmark across OpenAI, TypeSafe, Liquid, and Base Lab on your own data before signing anything multi-year. The leaderboard changes fast enough that a twelve-month contract locked to today's ranking is already a bet on the wrong horse by Q2.
If you're planning an enterprise agent rollout, Uber's pilot-to-company-wide expansion is a usable reference point for sequencing: start with a bounded pilot, measure adoption by function rather than by seniority, and expect your biggest unplanned win to come from non-technical staff doing analysis and reporting work, not from engineering productivity gains. Build your rollout's success metrics around that pattern rather than assuming developers will be the primary beneficiaries, because the two biggest enterprise AI deployments in this story both found the opposite.
