OpenAI Releases Decisions API for General Use to Rival TypeSafe Jev

OpenAI has made its Decisions API generally available to all developers, pivoting its platform toward lightweight, low-latency evaluations that ditch conversational text in favor of structured answers. Running on the company's GPT-6 Luna foundation model, the specialized endpoint aims to handle routine tasks such as routing customer tickets, moderating text, categorizing transactions, and orchestrating autonomous agents^{[3, 4]}.

The general release, announced by the company on October 22, marks an unusually rapid push from limited developer preview to full commercial availability. The launch accelerates a sudden industry race against TypeSafe, an emerging startup whose dedicated decision model, Jev, spawned a fast-growing software category tailored for automated System One reasoning.

Engineering for Sub-Second Classification

Rather than generating multi-paragraph prose that software systems must later parse with regex or JSON filters, the Decisions API accepts unstructured evidence and scores it against predefined schemas. Developers define explicit questions that yield three distinct response formats: a predicate estimating a yes-or-no probability, a multiple-choice selection from a constrained list, or a score mapped to a rubric^[4].

According to OpenAI product lead Nikunj Handa, the engineering team opted against creating a separate classifier architecture from the ground up, choosing instead to reconfigure existing weights within GPT-6 Luna.

The biggest breakthrough of the Decisions API is speed, compressing what used to take one to three seconds down to 100, 120, 180 milliseconds.

Nikunj Handa, OpenAI API Product Lead

TradingView reports that the interface can process requests up to ten times faster than calls routed through standard completion APIs. In testing leading up to general availability, language-learning platform Speak reported slashing the round-trip latency between interactive user actions from 1.4 seconds down to 0.13 seconds by switching to the endpoint. Early production adopters cited by OpenAI include Datadog, financial platform Ramp, and cloud workspace provider Box^[2].

Feature Expansion and Expanded Request Capacities

The general availability update introduces several capabilities that were absent during the beta period. To satisfy complex data pipelines, OpenAI doubled the evaluation volume per API call, allowing engineers to batch up to 400 questions within a single payload.

Key technical additions incorporated in the final production release include:

• Prompt Caching: Applications can reuse extensive document or image contexts across requests, incurring heavy discounts on repetitive tokens^[4].
• Multi-Modal Inputs: The engine now accepts native JSON objects, hosted HTTP and HTTPS image URLs, and image comparisons tailored to individual question prompts.
• GPT-Live Integration: Audio agents built on OpenAI's live streaming infrastructure can trigger background decision calls simultaneously as voice transcripts arrive.
• Enterprise Data Security: Infrastructure support for Private Link and regional routing across United States and European Union facilities, alongside HIPAA eligibility^{[4, 5]}.

The High-Stakes Battle with TypeSafe Jev

The sudden emergence of decision-only AI mirrors an industry-wide pivot away from open-ended chatbots toward tightly controlled autonomous loops. TypeSafe captured significant early market momentum after debuting Jev, reaching a multibillion-dollar valuation on the back of pure-play decision endpoints that serve as rapid routers for enterprise software. Competitors including Microsoft, Cloudflare, Perplexity, and open-weight model creators quickly rolled out alternatives^[6].

Pricing structures in this emerging segment reflect intense margin compression. OpenAI has set input pricing for the Decisions API at $0.10 per million tokens, falling to $0.01 per million tokens for cached data, while waiving output fees entirely because responses consist only of mathematical weights and chosen keys. In comparison, TypeSafe Jev lists its base input price at approximately $0.042 per million input tokens. While OpenAI carries an uncached price premium over Jev, its cached rate significantly undercuts the startup on workloads involving static policies, long system prompts, or shared reference documents.

Performance Benchmarks and Enterprise Friction

Despite OpenAI's aggressive rollout, third-party benchmarks highlight real tradeoffs between generalist foundation models and purpose-trained alternatives. Data aggregated across public tracking platforms shows Jev maintaining an edge in pure calibration and confidence consistency. On composite tracking indices such as AIMultiple and Artificial Analysis, Jev ranks significantly higher in reliability scores than the OpenAI Decision model.

Platform monitoring logs from developer proxy OpenRouter indicate that TypeSafe models continue to retain a commanding share of high-throughput classifier traffic. Some engineers on technical forums noted that early iterations of the Decisions API struggled with tail latencies under heavy system strain, while others pointed out that OpenAI's custom schema is not a drop-in substitute for applications built against TypeSafe's open specifications^[13].

Nevertheless, OpenAI executives are projecting substantial performance gains as hardware allocation matures. Nikunj Handa noted that the internal engineering roadmap aims to drive response latency down to roughly 60 milliseconds per request, a threshold low enough to permit real-time decision calls on individual video frames or continuous conversational voice streams. For developers balancing the overhead of large models, the arrival of production-grade decision layers suggests that the future of enterprise automation will depend on narrow, precise judgments rather than verbose language generation.