OpenAI 全面开放 Jev 竞品 Decisions API，决策模型赛道竞争升温

OpenAI 首次向用户全面开放其 Decisions API 模型，正式进军快速、低成本智能市场。此前，由前顶尖实验室研究员创立的初创公司 TypeSafe 正是凭借该赛道迅速成为估值 76 亿美元的公司。

TypeSafe 的 Jev 模型开创了一个全新类别：充当“超级分类器”的深度学习模型。在给定一组选项的情况下，它们能迅速做出选择并给出决策置信度。软件开发者很快发现，这类决策模型（或称“系统 1（System One）模型”）可以在多种工作流程中替代传统大语言模型（LLM），且运行速度更快、成本显著降低。

OpenAI API 产品负责人 Nikunj Handa 向 TechCrunch 透露，他的同事们看到 Jev 时直呼：“哇，我们这儿有成千上万个这类用例。”Handa 的团队随即全力冲刺推出自己的版本：他们没有专门重新训练一套校准的决策模型，而是将 ChatGPT Luna 进行了重新配置，以大幅提升推理速度并降低成本。

“Decisions API 带来的最大突破在于速度，它将耗时从原来的 1 到 3 秒大幅压缩至 100、120 或 180 毫秒，”Handa 表示，“未来我们可能会进行针对性训练，让模型在这些方面表现得更好。但目前来说，原版 Luna 就足以胜任这项工作。”

Decisions API 最早于 9 月 29 日的 OpenAI 开发者日上宣布，并于周四向所有客户正式开放。OpenAI 为公开发布版本增加了全新的语音和视觉模态，并通过按地理区域部署推理资源，确保全球各地区都能实现低延迟。

在与 Box、Datadog、Ramp 和 Speak 等客户的初期测试中，工程师利用该 API 替代了吞吐量要求更高的 OpenAI 大语言模型。AI 语言家教应用 Speak 的首席技术官 Andrew Hsu 表示，Decisions API 使公司动作间的延迟从 1.4 秒大幅缩短至 0.13 秒。OpenAI 还将该产品用于提升其 Dots 语音助手的响应速度。

尽管如此，除了“适合具体业务”之外，目前尚无明确标准界定什么是优秀的决策模型——在智能水平、成本、置信度校准和速度之间的权衡优化似乎因不同工作负载而异。TypeSafe 首席执行官 Diogo Almeida 不信任公开基准测试，呼吁大家在实际业务中直接测试模型；但即便如此，TypeSafe 上周在发布支持图像和“智能 if 语句”（smart-if statements）等新功能的最新 Jev 模型时，也发布了自己的排名榜单。

类似 AIM System One Index 和 Artificial Analysis Decision Index 等公开基准测试往往更青睐 Liquid 上周发布的全新决策模型 D1 Ultra 以及 Base Lab 的 Elephant-1。最新的 Jev 模型在两份榜单中依然保持竞争力，分列第七和第三，而 OpenAI 的决策模型则分别位列第八和第十四。不过目前来看，TypeSafe 的产品依然是 OpenRouter 上被调用最广泛的模型。

颇具趣味性的是一个名为 JevMan 的基准测试，该测试让决策模型来玩吃豆人（PacMan）游戏；最新的 Jev 模型在该榜单中拔得头筹。构建该测试的 Opper AI 首席增长官 Felix Wunderlich 表示，这款经典游戏“是一个非常有趣的测试，因为它充满了不间断、高风险的快速决策需求。在每一个路口，模型都必须在几百毫秒内决定行进方向。”

OpenAI 这款综合能力可能更强的模型，其优势或许在于多模态——Handa 告诉 TechCrunch，在图像分类方面，“Luna 极其强大，并展现出了更好的校准表现。”他补充道，OpenAI 的客户还希望 Decisions API 在输出图像方面能够具备更高的灵活性。

Handa 还认为，速度的进一步提升将解锁全新的应用场景：“如果我们能把延迟降到 60 毫秒左右，你可以想象对视频里的每一帧都进行一次 API 调用吗？”

随着这类工具速度越来越快、成本越来越低，我们或许能见证它们兑现 Jev 得名背后的经济学内涵——经济学家杰文斯曾阐释过，更高效地利用某种资源反而会促使其变得更加普及（杰文斯悖论）。或者正如 Handa 所言，随着决策模型的出现，大量原本显而易见却未被发掘的“疯狂用例”将迎来集中爆发。

OpenAI made its Decisions API model widely available to users for the first time, entering the market for fast, cheap intelligence that recently made TypeSafe, a startup founded by a former researcher at the frontier lab, a $7.6 billion company.

TypeSafe’s Jev model created a category for deep learning models that act as super classifiers: Given a set of options, they can quickly make choices and offer a confidence rating in their decision. Software developers quickly found that these decision or System One models could replace LLMs in a variety of workflows, operating more quickly and a at dramatically lower cost.

Nikunj Handa, the product lead for OpenAI’s APIs, told TechCrunch that his colleagues saw Jev and said “wow, we we have like a million use cases for this.” Handa’s team sprinted to offer their own version: Rather than building a specifically calibrated decision model, the team took ChatGPT Luna and reconfigured it to speed up the inference and lower the cost.

“The big unlock with Decisions API has been the speed, shrinking things from one to three seconds to 100, 120, 180 milliseconds,” Handa said. “In the future, we may sort of like have specific training to make our models like even better in these ways. But right now, it’s just plain old Luna that gets the job done.”

Announced at the company’s developer day on September 29, the Decisions API became available to all customers on Thursday. OpenAI added new voice and visual modalities to the public version, and worked to ensure lower latency in regions around the world by targeting its inference resources geographically.

In initial trials with customers like Box, Datadog, Ramp, and Speak, engineers used the API to replace OpenAI’s LLMs with higher throughput. Andrew Hsu, the CTO at Speak, which builds AI language tutors, said the Decisions API allowed the company to cut the time between actions from 1.4 seconds to .13 seconds. OpenAI also used the product to improve their Dots voice assistant’s responsiveness.

Still, it’s not clear exactly what makes a good decision model, other than “it works for you” — optimization across intelligence, cost, calibration, and speed seems to differ by workload. TypeSafe CEO Diogo Almeida distrusts benchmarks and urges people to trial the models in their own work, but even TypeSafe released its own rankings last week alongside the release of its latest Jev model, which includes new features including support for images and smart-if statements.

Public benchmarks like the AIM System One Index and the Artificial Analysis Decision Index tend to favor D1 Ultra, a new decision model that Liquid released last week, and Base Lab’s Elephant-1. The latest Jev model remains competitive on both indices, in seventh and third, with OpenAI’s Decision Model falling eighth and fourteenth, respectively. However, TypeSafe’s offerings are, for now, the most widely-queried on OpenRouter.

The most fun Jev benchmark, however, is JevMan, a benchmark where decision models play PacMan; the latest Jev model tops that ranking. Felix Wunderlich, the chief growth officer at Opper AI, which built the benchmark, said the classic game “is an interesting 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.”

Where OpenAI’s arguably more capable model may have an advantage is in multi-modality—for image classification, “Luna is extremely powerful and has been showing better calibrations,” Handa told TechCrunch, adding that OpenAI customers want the Decisions API to be more flexible about using images as outputs as well.

Handa also saw further increases in speed enabling new use cases: “Can we get this to something like you know 60 milliseconds, and then you could imagine like an API like call like for every single frame in like a video?”

As these offerings get faster and cheaper, we can expect them to live up to Jev’s namesake, an economist who explained how using a resource more efficiently makes it more ubiquitous. Or, as Handa sees it, there will be an expanding “crazy set of use cases” that were hiding in plain sight until the decision models revealed them.