GPT Researcher drops embeddings for Jev
Assaf Elovic measured 73% versus 46% relevant context swapping embeddings for Jev, and GPT Researcher now ships that way by default. The same week, pg-jev shut down over a 50k row ceiling.
Built with it
Chase says LangChain already ships Breunig's Jev router shape
Drew Breunig's DSPy and Jev router, sketched in about 15 minutes, picks the model from the initial prompt, and Harrison Chase said LangChain ships the same thing as ModelRouterMiddleware. Chase added that Jev is cheap enough to re-pick the model after each tool result with a custom hook, rather than once per run.
jev is cheap enough that you can also re-pick after each tool result with a custom hook
Someone measured it
GPT Researcher now runs on Jev by default, embeddings removed
GPT Researcher no longer needs embeddings at all, after Assaf Elovic tested both retrieval paths on 28 research tasks from SimpleQA and open ended research. He reported 73% versus 46% relevant context, reports preferred 15 to 3 in blind comparisons, and the same cost per report.
retrieval is a decision problem, not just a similarity problem
Sydney Runkle says a coding agent router cut median cost 64%
Sydney Runkle built a model router for a coding agent and says it cut median cost per task by 64% with no measurable drop in quality. Harrison Chase laid out the recipe alongside it: understand the tasks, understand the models, build the router inside the harness, then track outcomes.
most tasks don't need top tier intelligence!
The case against
Almeida says OpenAI's Decisions API is Luna with constrained decoding
Ian Nuttall asked whether OpenAI's Decisions API, built on Luna, is a Jev killer, and Diogo Almeida replied that it looks like Luna with constrained decoding. Pressed by Justin Schroeder, he named two problems: a model shared with other objectives rather than its own dedicated one, which can cause jaggedness, and constrained decoding on string outputs instead of something machine-native. Nobody has posted numbers either way.
constrained decoding often makes models dumber for reasons I've talked about previously
Zachi ends pg-jev: query speed is the bottleneck above 50k rows
Zachi called the newest pg-jev release its last, saying it is fine under 50k rows but query speed becomes the bottleneck above that. He says he spent about 3b tokens on benchmarks and testing, and is building a replacement on Jev meant to run in Supabase, Neon or Drizzle.
its safe to say that pg-jev wont work as exptected
Also on the timeline
Cloudflare ships Clef, a Jev-API compatible decision model
Cloudflare introduced Clef and Clef-flash, open source decision models hosted on Workers AI. Its developer account says they are smarter, faster and fully Jev-API compatible.
Jev lands in Vercel's AI SDK for Python
Jev is now in the AI SDK for Python.
Frequently asked questions
What is a model router and why does it matter?
A model router picks which AI model to use based on the task, either once at the start or after each tool result. Multiple teams report it can reduce costs significantly while maintaining quality, though each measurement comes from their own internal tests.
How does GPT Researcher use Jev now?
GPT Researcher switched to running on Jev by default and removed embeddings entirely. In their internal tests on research tasks, this approach returned more relevant context and got better ratings in blind comparisons at the same cost.
What are the limits of pg-jev?
pg-jev works well with tables under 50,000 rows, but query speed becomes the bottleneck at larger sizes. The developer recommends sizing your table before adding a Jev call inside a query.
How does OpenAI's Decisions API compare to Jev?
Diogo Almeida said Decisions API appears to use constrained decoding on text strings rather than machine-native outputs, and shares a model with other tasks which can cause instability. No public benchmark has compared the two directly.
What is Cloudflare's Clef?
Cloudflare released Clef and Clef-flash, open source decision models that run on Cloudflare Workers AI and are fully compatible with the Jev API.