ScaleDown claims 77.2 on DecisionBench against Jev's 72
Three single-task models for classify, extract and summarize, and every number so far comes from the vendor.
Neal Patel launched Constrained Decoding Models, a family of three single-task models that he says skip the token-by-token generative loop and read the input once. The three are Classify, which returns a calibrated probability and a reason per label, Extract, which returns JSON with confidence scores and citations from a natural language field description, and Summarize, which takes long text, images or PDFs and returns a summary under 1,000 tokens.
1M tokens of multimodal (text + image) input, 31x larger than Jev
The headline number is 77.2% on DecisionBench against what Patel puts at 72% for Jev. He also claims 1M tokens of multimodal text and image input, which he describes as 31x larger than Jev, and 200x faster and 467x cheaper than frontier LLM APIs. Pricing in the launch post is $0.05 per million input tokens with output free, including for Summarize. There is a promotion of 50M free input tokens for the first 50 hours.
What OpenRouter listed
OpenRouter has the family live and attributes it to @scaledown_team, with a launch discount of $0.04 per million input tokens and free output. Its listing names the classifier Classify-1 and routes it through the Decisions API, the same surface other decision models sit behind. OpenRouter's description of Extract matches Patel's, plain English field descriptions in, clean JSON with confidence scores and source spans out.
20% off for launch: $0.04/M input tokens, output free.
For people building on it
Every comparative number in this story came from the company shipping the model. DecisionBench is the benchmark ScaleDown picked, the 72% figure for Jev is ScaleDown's figure for Jev, and no independent run has been posted. The 200x faster and 467x cheaper claim is measured against unnamed frontier LLM APIs, not against Jev, so it tells you nothing about how the two compare on latency or price per decision for your workload.
The context window claim is the one worth testing first if you have a long document pipeline, because 1M tokens of multimodal input is a different shape of product from a decision model that takes a short prompt. The two prices quoted in the same 24 hours, $0.05 from the launch post and $0.04 from OpenRouter's launch discount, are the same model at two list points, not a price cut.
Know what Jev is good at before you bet on it
One email a day. Under three minutes. What builders found out about Jev yesterday.