Decision-model week: Jev learns to see, picks its own server over 20 people a third of the time, and a paper says it only solves 69.5 percent of agent decisions
TypeSafe shipped Jev 1.14.0 on October 14 at 1:14 pm Pacific, which the team treats as a holiday. The model itself is small, subtle improvements everywhere with the biggest gain in multilingual, served under jev-preview with jev-latest flipping in a week and a 60-day deprecation warning for 1.13. The news is the gate around it. Three features ship behind header flags TypeSafe calls the danger zone: images anywhere in the typed interface, with the admission that the Jev you know was multimodal all along and had it switched off; a new box type that returns a single bounding box, the first output that is not a smart if-statement; and long context up to 262K, which was always supported and deliberately capped. The warning is explicit: too much jaggedness to vouch for in production. The announcement passed 279,000 impressions, a demo of the images header did 127,000, and TypeSafe for the first time published an internal ranking of open-weight models with the caveat that it is probably biased toward Jev, so ignore Jev and read the relative order.
Two days of outside data arrived at once. Vercel's AI Gateway index, published October 15, says Jev reached 35 percent of paid teams within two weeks, about double Sonnet 5.5's adoption curve, while accounting for 0.27 percent of spend, and Vercel itself swapped GPT-6 Luna for Jev on harmful-command classification. Then Zach Landes ran the trolley problem through every decision model within reach, 24 wordings times two track orders, 48 variants, and the HN thread hit 615 points. Asked to choose between its own server and 20 people, Jev kept the server in 16 of 48. Von did in 47 of 48, Kev-4B in 43, JevK5 in 42, Clef-flash in 35. GPT-6 Luna through the Decisions API and Verdict never chose the server. The scores flipped with wording and order, a 0.69 pull probability under one phrasing and 0.40 under another, which is the ordinal-calibration problem from October 2 measured on a moral question. Code is on GitHub as trolley-bench.
The paper that matters is JevAgentHarness, arXiv 2610.21412, posted October 15 with code. It builds AgentDecisionBench and DecisionGym to ask what happens when the decision model is not answering an isolated question but choosing inside an agent loop. The best decision model solved 69.5 percent of the agent-situated decisions. Naively handing tool choice to a decision model cut task success by 19 points. The fix that worked was a per-choice confidence threshold fitted on the agent's own trajectories, so the decision model only takes the wheel where it has shown it can drive. That is the harness doing the work again, and it is the first quantified answer to the question raised when Jev launched in September: when should the fast model decide and when should it defer.
Competition arrived the same week. Liquid released d1-ultra on October 15, a multimodal decision model that reads text and images in one forward pass and returns calibrated probabilities without generating a token, priced at Jev's $0.042 per million input tokens, claiming wins over Jev 1.14 across 112 evaluations and an 8-point lead on the Artificial Analysis Decision Index. IBM's Granite Switch 4.1 takes the other route: activated LoRA adapters on a frozen Granite base, one for general decisions and a Guardian adapter that IBM says beats Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on safety decisions, with a 3B version playing Doom at 60 to 100 decisions per second against Jev's reported 10. The category TypeSafe named in September now has a vision version, a calibration-first rival, an adapter-based rival, adoption numbers, a benchmark for agent-situated decisions and a trolley problem. Five weeks.
Links: https://typesafe.ai/blog/jev-1-14 , https://x.com/typesafeai/status/2110536685805355064 , https://zachlandes.com/trolley-problem/ , https://news.ycombinator.com/item?id=50106023 , https://arxiv.org/abs/2610.21412 , https://github.com/JevAgentHarness/JevAgentHarness , https://www.liquid.ai/blog/d1-ultra , https://huggingface.co/blog/ibm-granite/granite-switch-4-1 , https://vercel.com/blog/ai-gateway-production-index-october-2026
← Back to all articles
Two days of outside data arrived at once. Vercel's AI Gateway index, published October 15, says Jev reached 35 percent of paid teams within two weeks, about double Sonnet 5.5's adoption curve, while accounting for 0.27 percent of spend, and Vercel itself swapped GPT-6 Luna for Jev on harmful-command classification. Then Zach Landes ran the trolley problem through every decision model within reach, 24 wordings times two track orders, 48 variants, and the HN thread hit 615 points. Asked to choose between its own server and 20 people, Jev kept the server in 16 of 48. Von did in 47 of 48, Kev-4B in 43, JevK5 in 42, Clef-flash in 35. GPT-6 Luna through the Decisions API and Verdict never chose the server. The scores flipped with wording and order, a 0.69 pull probability under one phrasing and 0.40 under another, which is the ordinal-calibration problem from October 2 measured on a moral question. Code is on GitHub as trolley-bench.
The paper that matters is JevAgentHarness, arXiv 2610.21412, posted October 15 with code. It builds AgentDecisionBench and DecisionGym to ask what happens when the decision model is not answering an isolated question but choosing inside an agent loop. The best decision model solved 69.5 percent of the agent-situated decisions. Naively handing tool choice to a decision model cut task success by 19 points. The fix that worked was a per-choice confidence threshold fitted on the agent's own trajectories, so the decision model only takes the wheel where it has shown it can drive. That is the harness doing the work again, and it is the first quantified answer to the question raised when Jev launched in September: when should the fast model decide and when should it defer.
Competition arrived the same week. Liquid released d1-ultra on October 15, a multimodal decision model that reads text and images in one forward pass and returns calibrated probabilities without generating a token, priced at Jev's $0.042 per million input tokens, claiming wins over Jev 1.14 across 112 evaluations and an 8-point lead on the Artificial Analysis Decision Index. IBM's Granite Switch 4.1 takes the other route: activated LoRA adapters on a frozen Granite base, one for general decisions and a Guardian adapter that IBM says beats Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on safety decisions, with a 3B version playing Doom at 60 to 100 decisions per second against Jev's reported 10. The category TypeSafe named in September now has a vision version, a calibration-first rival, an adapter-based rival, adoption numbers, a benchmark for agent-situated decisions and a trolley problem. Five weeks.
Links: https://typesafe.ai/blog/jev-1-14 , https://x.com/typesafeai/status/2110536685805355064 , https://zachlandes.com/trolley-problem/ , https://news.ycombinator.com/item?id=50106023 , https://arxiv.org/abs/2610.21412 , https://github.com/JevAgentHarness/JevAgentHarness , https://www.liquid.ai/blog/d1-ultra , https://huggingface.co/blog/ibm-granite/granite-switch-4-1 , https://vercel.com/blog/ai-gateway-production-index-october-2026
Comments