1 October 2026 was not a GPT or Claude day: it was decision-model day. Perplexity launched pplx-decider-v1-27b and a Decisions API; AWS Strands Labs released Strands Decider 2B; Cloudflare, the same day, shipped Clef. Capital & Compute counted ten new models in four days, all clustered on that date, none from the classic frontier labs.
pplx-decider-v1-27b is described as an Apache 2.0 fine-tune on a Qwen-class 27B base that answers with typed probabilities to yes/no, multiple-choice and rubric questions. It supports up to 128 questions and very large input windows (reports cite hundreds of thousands of input tokens). The cited price — about $0.04 per million input tokens with free output — places it among the cheapest hosted deciders.
Perplexity claims a slight edge over Typesafe’s Jev on an internal benchmark (published figures around 85.7% versus 84.5%). As always with vendor leaderboards, reproduce on your own data. The structural point is the product: a decisions API for agents that do not want to parse prose.
AWS, via Strands Labs, released Strands Decider 2B, a small open-source model built for routing and tool choice in fractions of a second on-device or at the edge. TechCrunch and VentureBeat framed it as a free, fast “Jev killer”: it does not compete on long essays; it competes on the micro-decision an agent takes a thousand times per task.
The convergence is not accidental. After months of agents hallucinating format, drifting off scope or spending dollars of chain-of-thought to say “yes/no,” the market is discovering that a calibrated decision head can be the cheapest guardrail. Billing input only aligns incentives: the provider does not earn on verbosity.
For architects, the emerging pattern is hybrid: a frontier LLM for planning and generation; a decider for policy, triage and “should I call this tool?” Cloudflare Clef, Perplexity and Strands fight to be that second head. Whoever wins the schema standard (Jev-compatible or otherwise) wins fine-tuning data lock-in.
Risks: misleading calibration, rubric bias, and the false comfort of a pretty probability. A decider at 0.97 may be miscalibrated out of distribution. Governance still needs continuous evaluation, human queues and immutable decision logs.
Stack geopolitics: Qwen bases inside Western deciders show how central the Chinese open ecosystem is in 2026, even as OpenAI cuts rival APIs over distillation. The irony — using open Asian weights to decide inside U.S. agents — will not escape export regulators.
Enterprises on Bedrock fleets have a natural path to Strands; search-and-synthesis shops to Perplexity; teams already on Cloudflare’s network to Clef. The battle is distribution as much as F1.
Neural Pulse recommends A/B tests of cost and tool-error rate before replacing a giant LLM on the hot path. In many flows, latency savings matter more than a 1% accuracy bump on a synthetic benchmark.
As of 4 October, with three decider families in the air, the category is no longer a Typesafe experiment: it is infrastructure. Frontier labs must decide whether to ship their own, buy, or keep pretending one model is enough.
We close the day’s slate with this reading: the news is not just another model, but a shape shift — from free text to typed action — that may define how safe agents are built in 2027.
Observability vendors (Datadog, OpenTelemetry) should add standard attributes for “decision_model,” “threshold” and “rubric_id.” Without common telemetry, comparing deciders in production will be impossible.
Free output on decision APIs may change. Budget for scenarios where output is also billed to avoid surprises when agent traffic explodes.
We close: typing the action is the logical next step after typing the tools. 2026 will be remembered as the year agents began deciding with models built to decide.
