ScaleDown claims 77.2 vs Jev's 72 on DecisionBench
Neal Patel launched a three model family and put Jev's DecisionBench score in the launch post, 77.2 against 72. The number comes from the people selling the model, and nobody outside has run it yet.
Built with it
Open SWE routing numbers land: 973 threads, merge rate unchanged
The detail behind LangChain's 64% cost cut is now out: NeilXbt's writeup says 973 threads were split half routed, half always on GPT-6 Astra, with median cost per thread $0.94 against $2.61 and merged PR rates of 29.2% versus 27.3%. A separate run that always used the fast model was killed within a day after engineers reported low output quality.
router lives in the harness, not a gateway, bc the harness knows the task
Someone measured it
ScaleDown says its Classify model scores 77.2 on DecisionBench, Jev 72
Neal Patel launched Constrained Decoding Models, three single-task models for classify, extract and summarize, and claims 77.2% on DecisionBench against Jev's 72%. He also claims 1M tokens of multimodal input, 31x larger than Jev, at $0.05 per million input tokens with output free. OpenRouter lists the family at $0.04 per million input for launch.
1M tokens of multimodal (text + image) input, 31x larger than Jev
The case against
Almeida says the Jev clones copy the interface and miss the reliability
Diogo Almeida, asked by Ash about everyone copying Jev, said the copies take the interface without understanding reliability. His worry is that people who try a weak clone will conclude the whole class of decision models is not useful. In a separate post he said he is anti public benchmark even when his side is ahead.
they copy the interface, but I don't think they understand the importance of reliability