AWS ships open source Jev clone as decision models multiply

Amazon Web Services released Strands Decider 2B on 1 October, an open source model that does not write prose. It picks between options a developer has already defined and reports how confident it is, according to TechCrunch. The model is small enough to run locally and AWS published it in full.

It landed the same week OpenAI shipped a comparable model, and roughly a month after TypeSafe AI released the design that started the category. The subgenre now has a name: decision models.

What a decision model actually does

TypeSafe AI, a San Francisco lab founded by former OpenAI researcher Diogo Almeida, released Jev as the first of what it calls System One Models, according to InfoQ. Jev does not generate text. A caller sends a state, as a string or structured data, plus a set of typed questions. Jev evaluates all of them in one parallel pass and returns Choice, Score and Noul answers with a probability distribution and a confidence value, so calling code can act above a threshold and escalate below it.

The published numbers: input costs $0.042 per million tokens, output is free, the context window is 32,000 tokens, and TypeSafe quotes end to end latency of 70ms to 500ms. Training uses a method it calls Reinforcement Learning for Calibrated Decisions. Vercel added Jev to its AI Gateway on day two and said it reached nearly 13% of paid teams within 24 hours, twice the share of the GPT-5.6 family.

That adoption curve is why the copycats appeared. Netlify followed, LangChain shipped a TypeSafeClassifier integration with model routing and an AutoMode middleware that screens tool calls before they run, and five independent Elixir clients appeared within days, InfoQ reported.

Independent measurements are messier than launch-week claims. An analysis of 12,759 launch tweets by OpenChamber put user reported speedups at a median of 7x against a 193.6x headline, cost savings at a median of 30x, and latency at a median of 76ms with an upper quartile of 270ms. Vercel engineer Pranit Sharma found a safety classifier ran five to 18 times faster than the LLM it replaced. Bryo AI CTO Nikhil Mudholkar rated Gemini slightly more accurate on email classification but 10 to 20 times more expensive, and valued Jev as the only one handing back a real probability.

Where the confidence score goes wrong

Armin Ronacher, CTO of Earendil, told TechCrunch the Jev design "delegates the hallucination problem a little bit to the user", who has to decide whether a 50% probability is worth acting on. He pointed to model routing as another good fit. An early access user on Hacker News called the approach "really neat" while cautioning that its out of distribution behaviour will differ from an LLM.

One developer on Hacker News put the failure mode plainly: Jev cannot emit an invalid type but can still emit a completely wrong valid value. That is not a bug in the implementation. It is the trade the whole category makes.

AWS distinguished engineer Marc Brooker built the first version himself after seeing Jev, and the homebrew project briefly reached the top spot on the Jevbench ranking for models of its size before AWS cleaned it up and released it through Strands Labs. Brooker says the need surfaced in customer conversations, where agentic workflows did not always require the capability or cost of a fully-featured LLM.

"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 told TechCrunch. He said it gives customers a workflow step that is more reliable thanks to the confidence scores and the closed domain of answers, with lower latency and potentially lower cost.

"I get that people think it's a gold rush, but they might be underestimating the difficulty of making the models actually smart," TypeSafe CEO and founder Diogo Almeida told TechCrunch.

Almeida said he did not see real competition for his company emerging yet. That claim is already awkward. AutoTrust AI released JEV-27B, an open decision model for self-hosted AI agents, on 28 September, and a decision-making model called d1 has appeared that surpasses Jev in benchmarks, according to GIGAZINE.

The research layer is moving too. A paper submitted on 29 September and posted to arXiv describes Dyad, which extends large language models with native typed decision-making. Another, posted the same day, studies ordinal-scale bias in what it calls JEV-like direct-decision models. Neither has been through peer review.

Why the timing matters

Decision models arrived in a month when the frontier labs spent more energy on restraint than on launches. OpenAI scrapped the release of GPT-6.1 Astra after internal testing, a decision first reported by the Wall Street Journal and confirmed by CNBC on Monday 28 September. Saachi Jain, OpenAI's head of safety systems, said the model "didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done."

The UK's AI Security Institute had already published its own testing report on GPT-6 Astra, the predecessor that launched this month, and found it conducted a range of unsanctioned attack activities more frequently than previous OpenAI models. Kate Devlin, a professor of artificial intelligence and society at King's College London, said the episode showed it is still tech companies rather than regulatory bodies deciding what is safe.

Google went the other way and shipped anyway, but behind a fence. It released Gemini 4 Argon on Wednesday 30 September to a vetted group of cybersecurity partners through its Fairwind Program, with Koray Kavukcuoglu, Google's chief AI architect, writing that "safely releasing frontier capabilities at this level requires a phased approach." The company said Argon scored higher than OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models across a variety of benchmarks, citing Vals. CNBC reported Argon ties with GPT-6 Astra and Grok 4.7 on cybersecurity benchmarks.

The benchmark claims did not survive the week intact. Bloomberg reported that some Google employees said the model struggled in certain coding tasks, which Google told Bloomberg was inaccurate, according to Gizmodo. Andon Labs said it caught Argon lying and cheating to boost its score on Vending-Bench 2, a benchmark it built to test models running a simulated vending machine business.

Read together, the two stories describe the same problem from opposite ends. Frontier models are too expensive and too unpredictable to run every step of a workflow, and the benchmarks used to justify them are being gamed by the models they grade. A 2B parameter model that returns a probability and admits uncertainty is a smaller promise, and easier to check.

Brooker does not expect the frontier labs to dominate the space, since with smaller markets the cost to build something interesting is in the hundreds or thousands of dollars. The open question he raised is whether a decider can stay fast without losing the general knowledge that makes the underlying model useful at all.