Decision Model APIs: AWS and Ollama Back System One, Not OpenAI

Run AWS’s new open-weight decision model as a server and it answers at /v1/systemone, an endpoint a startup introduced less than three weeks earlier. That detail matters more than any benchmark in the category right now. Decision models are standardizing on one startup’s schema faster than anyone has standardized their quality.

What Happened

The New Stack’s analysis traces how TypeSafe AI’s System One API, introduced with its Jev model, became the de facto contract for the category. A decision model takes application state and a set of typed questions, then returns structured answers with a probability for each option. It never writes prose. System One reduces every decision to three question types: choice (pick one option), score (place input on an ordered rubric), and yes/no, which TypeSafe calls “noul.”

Adoption came in two groups. The first implements TypeSafe’s endpoint directly: Upstage serves Solar Decide through it, and Ollama added /v1/systemone in version 0.35. AWS’s Strands Decider 2B, released October 1 under Apache 2.0 on Qwen3.5, exposes it too, as do community models such as Jared Palmer’s Kev and Open-Jev. The second group keeps the semantics but changes the address. OpenRouter, Venice, and Perplexity each run their own decisions endpoints, and OpenRouter also offers a TypeSafe-compatible route.

OpenAI is the outlier. The New Stack describes its Decisions API, built on a specialized version of GPT-6 Luna, as a limited preview with no published schema. Stack Archive covered that API’s pricing earlier this week.

Why It Matters

The precedent is OpenAI’s own Chat Completions API, which became the default way to call an LLM because so many servers cloned it. System One is following the same path, faster, because it is much smaller: three question types and less than a month old. TypeSafe’s SDKs work against any server implementing the contract once you change the base URL, so an application can move from hosted Jev to a self-hosted Strands Decider without rewriting its decision logic.

The economics push the same way. On OpenRouter, Jev costs $0.042 per million input tokens with no output charge, and Perplexity prices its hosted model at $0.02. AWS reports a median of 115 milliseconds per question for Strands Decider on an RTX 3090. A routing or policy check becomes far cheaper than an equivalent LLM call.

What This Means for Builders

The New Stack describes decision models as the if statement of an AI application. The LLM plans and writes, and the decision model makes the small judgments between steps. Three patterns already appear in production tooling: routing requests to the right model (OpenRouter’s Jev Router), gating tool calls before they execute (an example in the Strands Decider repository), and falling back to an LLM when the decision endpoint is down (Maxim AI’s Bifrost gateway).

Portability is not the same as equivalence. Servers still differ in option limits, how they compute confidence, and answer quality. Confidence also measures how concentrated the probability distribution is, not the chance an answer is right, and TypeSafe advises setting thresholds from your own data. Robustness is the bigger warning. A University of Southern California researcher found an optimizer could redirect 61.4% of 508 initially correct Jev decisions by adding short, natural-looking context, and an open Qwen-based model flipped at 72.6%. That is an adversarial flip rate, not an everyday failure rate, but it points to a weakness in the category rather than in one vendor.

Source: The New Stack

Content created with AI assistance and reviewed for accuracy.