Decision Models Spread to Six Providers in Three Weeks, Converge on One API

Within three weeks, decision models went from a single startup's experiment to a full category, according to an October 2026 analysis published by The New Stack. AWS, OpenAI, Upstage, Perplexity, and Cloudflare now offer hosted decision models, while AWS and independent developers have released open weights on Hugging Face. What's converging fastest isn't the endpoint address — it's the structure of requests and responses, which most vendors have cloned from startup TypeSafe AI's System One schema.

The report documents how AWS's Strands Decider 2B, released October 1, answers at the /v1/systemone endpoint — the same path TypeSafe introduced with Jev just over two weeks earlier. Upstage serves Solar Decide through the System One schema, and Ollama added a /v1/systemone endpoint in version 0.35, explicitly based on TypeSafe's API. Strands Decider, Jared Palmer's Kev, Zefan Cai's community-built Open-Jev, and local runtimes including Ollaya and SGLang expose the identical interface. OpenRouter's native Decisions API uses its own address but also provides a TypeSafe-compatible route, letting existing SDK clients switch providers by changing the base URL. Venice reportedly serves Jev through a beta decisions endpoint of its own, while Perplexity's Decisions API supports the same choice, yes-or-no, and score question types but uses a different path. OpenAI remains the outlier, answering at DevDay with a Decisions API built on a specialized version of GPT-6 Luna that's still in limited preview with no published schema or pricing. On pricing, Jev costs $0.042 per million input tokens on OpenRouter with no output charge, while Perplexity prices its hosted v1.1 model at $0.02 per million input tokens. AWS reports a median of 115 milliseconds per question for the 1.9 billion-parameter Strands Decider on an RTX 3090. Florian Standhartinger and contributors maintain JevBench under the MIT license, and its v1.5.4 leaderboard tracked 112 decision systems, placing Jev third behind blockbrain's Cygnet and the open-weight Winnow-12B.

The report finds that decision models return structured answers with a probability attached to each option and never write prose, code, or explanations of their reasoning. According to the analysis, AWS reports that answers at 0.9 confidence or higher were correct about 95% of the time on short classification tasks the model hadn't seen. Three patterns emerged within days of Jev's launch: OpenRouter launched Jev Router on September 25 to choose the model and reasoning effort for each incoming request, the Strands Decider repository ships an example that gates a weather-tool call on two yes-or-no decisions, and Maxim AI's Bifrost gateway routes decisions to an LLM when Jev is unreachable. A University of Southern California researcher found that an optimizer could redirect 61.4% of 508 initially correct Jev decisions by adding short, natural-looking context that pushed the model toward a chosen wrong answer, with successful additions having a median length of 31 words. The report notes this is an adversarial targeted-flip rate, not the probability that ordinary context will cause Jev to fail, but an open Qwen-based model in the same study flipped at 72.6%, suggesting the weakness lies with the category rather than with a single vendor.

The report draws a parallel to OpenAI's Chat Completions API, which became the default interface for LLMs because so many providers cloned it. OpenAI later launched Open Responses in January 2026, an open specification based on its Responses API, with OpenRouter, Hugging Face, LM Studio, vLLM, Ollama, and Vercel as launch partners. The precedent carries two lessons: the semantic shape spreads before any formal governance, as System One is experiencing now, and the company that defines the shape eventually has to open it, or the ecosystem drifts into compatible variants. The practical result is code-level portability — TypeSafe's SDKs work against any server implementing the System One API once you change the base URL, so an application written against the contract can move from hosted Jev to a local Strands Decider without rewriting its decision logic. For enterprises, the analysis notes, a shared contract adds a second benefit: governance and procurement teams can approve one integration pattern, then swap the backend as accuracy, price, and data-residency requirements change, giving enterprise customers a stronger negotiating position than a market of incompatible endpoints would. The same contract commoditizes the interface for vendors — TypeSafe now has to win on model quality and calibration, because a developer can replace Jev with Strands Decider or Kev by editing a URL. Once published, OpenAI's schema will decide whether the category consolidates around one contract or splits into two. The report concludes that whether developers trust decision models to gate real actions will depend less on API convergence than on calibration and robustness, noting that the safeguards teams build around these probabilities will matter just as much. The standardization race mirrors the broader shift toward modular AI architectures, where specialized components replace monolithic models. If TypeSafe fails to formalize its advantage through an open specification, it risks watching competitors fragment the very interface it created.