Amazon Web Services has entered the growing field of AI decision models with the release of an open source offering aimed at developers building automated workflows. The release comes at a time when the AI industry is increasingly looking beyond large language models for tools better suited to structured, task-specific automation.
The new model, called Strands Decider 2B, is designed to rapidly select between a predefined set of options and return a confidence score alongside its choice. It is lightweight enough to run on local hardware, fully open source, and positioned as a lower-cost alternative to deploying a full-scale LLM for every step in an agentic pipeline. Its release coincided with a similar announcement from OpenAI, signaling a broader industry shift toward this category of specialized models.
From a Personal Project to an AWS Release
The model traces its origins to a personal experiment by Marc Brooker, a distinguished engineer at Amazon. Brooker came across TypeSafe's Jev model and was inspired to build his own version from scratch. The side project performed well enough to briefly claim the top position on the Jevbench leaderboard for models in its size class. That result caught the attention of colleagues, and Amazon's internal team known as Strands Labs - a group focused on developing tools and protocols for AI agent deployment - refined and packaged it into a formal release.
Strands Labs has been building out a broader ecosystem for agentic AI, and the Decider model fits into that mission as a practical utility for workflow orchestration. Rather than relying on a general-purpose LLM to reason through every decision in a pipeline, developers can use the Decider to handle discrete routing tasks more efficiently.
A Tool Built Around Real Customer Needs
Brooker has said the motivation behind the project was grounded in direct feedback from AWS customers. Many of those customers operate agentic workflows where not every step requires the full reasoning power - or the associated cost - of a frontier model. For routine decision points, a smaller, faster, and cheaper model is often sufficient.
"What originally piqued my interest in this class of models was that they make a perfect decider for a workflow step - 'what is the next thing for me to do here, based on where I am?'"
Brooker described the advantage of the model as offering a more reliable workflow component, thanks to its confidence scoring system and its operation within a closed domain of possible answers. He noted that the approach can deliver lower latency and potentially lower costs compared to routing every decision through a full LLM.
How the Model Works
Like others in this emerging model category, Strands Decider is built on the backbone of an existing language model - specifically Qwen3.5-2B. However, rather than generating open-ended text responses, it produces calibrated selections from a predefined list of options. This architectural approach allows it to operate with greater speed and predictability than a generative model in the same decision-making role.
The category itself was pioneered by TypeSafe, whose Jev model gave the benchmark its name. TypeSafe named Jev after the 19th-century economist William Stanley Jevons, invoking his well-known paradox: that as the cost of a resource falls, overall consumption of it tends to rise. The implication for AI is that cheaper, more efficient models could actually drive greater adoption and usage rather than simply replacing more expensive ones.
A Crowded Field With Open Questions
Since TypeSafe introduced the concept, dozens of similar models have been developed and released by research teams around the world. The volume of activity reflects genuine interest in the approach, but it also raises practical questions about differentiation and long-term value.
Brooker acknowledged that the key technical challenge is maintaining a careful balance between speed and intelligence. Optimizing a model too aggressively for fast, narrow decision-making risks degrading the broader language understanding that makes it generally useful across different contexts and domains.
"There is a very careful balance to be found where you want to push its performance on accuracy and calibration on these kinds of tasks, without degrading its performance on understanding different languages, on having the kind of knowledge it has, which is what makes it general purpose and interesting and useful."
Despite the growing number of entrants into this space, Brooker does not expect the largest AI labs to dominate it. He pointed out that the cost to build a compelling model in this category can be measured in the hundreds to thousands of dollars - a relatively low barrier that keeps the field accessible to smaller teams and independent researchers.
TypeSafe's Perspective on the Competition
For its part, TypeSafe appears unfazed by the surge of new entrants. The company says it is focused on continuing to improve its own models rather than reacting to competitive pressure. Diogo Almeida, the company's CEO and founder, suggested that many of the new releases in this space may be underestimating what it actually takes to build genuinely intelligent decision models.
"I get that people think it's a gold rush, but they might be underestimating the difficulty of making the models actually smart."
Almeida said he has not yet seen what he would consider real competition emerge for TypeSafe, characterizing the current wave of releases as driven more by technical curiosity than by a sustained commitment to making AI decision-making practically useful at scale.
"The current batch seems more like ML people wanting to implement a cool architecture than a team deeply dedicated to making intelligence useful," he said.
Whether Amazon's entry into this space - backed by the resources and customer base of AWS - changes that dynamic remains to be seen. What is clear is that decision models are drawing serious attention from across the industry, with developers and enterprises alike looking for more efficient ways to build and run AI-powered workflows.
