Decision Models: The Fast Layer for AI Agents

A model that takes a block of context and a set of predefined possible answers, then returns one answer from that set plus a confidence score for each option. It never generates text; the output is always one of the choices you supplied, with a probability attached.

It removes the vocabulary head and replaces it with a pointer or decision head that scores only the specific options supplied. There is no vocabulary and no autoregressive sampling, so the model produces a score for each option in a single forward pass.

70-500ms end-to-end latency and $0.042 per million input tokens; a pipeline routing 10 million decisions per month through Jev spends $420. It also notes OpenAI's Decisions API demonstrated roughly 150ms versus 1,600ms for standard calls, about 10x faster.

Reinforcement Learning for Calibrated Decisions: rather than cross-entropy loss, models train to maximize strictly proper scoring rules such as the Brier score and logarithmic scoring rules, so a 0.6 confidence genuinely means about 60% confidence. Laya documents a GRPO-style implementation with logarithmic, spherical, and ranked probability scoring.

Jev 1.13, AWS Strands Decider 2B (Qwen3.5-2B plus a pointer head, Apache 2.0), Convai's Laya (ModernBERT-large plus a decision head, 421M, Apache 2.0), Kev-4B (Qwen3.5-4B plus LoRA, open weights), OpenAI's Decisions API, and the Ollaya runtime that bundles several open models.

Because thresholds are only as good as the confidence they gate on: if a model returns 0.92 where it is actually right 72% of the time, a 0.85 threshold gives false security. Calibrated confidence is what makes auto-act versus escalate decisions tunable against the cost of a wrong decision.

DeepInfra's blog, published October 13, 2026 by Niklas. It is a vendor making the case for its own inference category, and prices, model versions, and availability in this space are moving quickly, so verify claims against the linked providers before relying on them.