Quick Hits

• Jev is the fastest-adopted model in Vercel AI Gateway history. Per @vercel, it reached roughly 13% of teams on its first day, about 2x the GPT-5.6 family and 6x Fable 5.1. @rauchg credits the surge partly to the "AI is too expensive/slow" zeitgeist, and @vercel_dev has made Jev free on the Gateway until Sept 25.
• Open-source rivals arrived in days, not years. @jun_song argues an OpenAI co-founder spent three years quietly building Jev "only for someone to open source a better performing model in just 3 days," pointing to Laya via @0xCVYH, while @michael_chomsky says a free classifier he offers already outperforms it.
• NVIDIA open-sourced a self-evolving agent harness. As @omarsar0 summarizes, SoL-Pi cuts token traffic nearly in half and took about a third off API costs on the 51-task EdgeBench benchmark, an estimated $8.75 to $13.50 per hour saved versus native Codex and Claude Code harnesses.
• A 706k-parameter model fills a whole form in one 50ms pass. @trycua open-sourced CUA-S1-FORMS, which @be_arsh contrasts with an LLM agent that needed 23 turns and 39.6 seconds for the same form.
• Platform rollouts: @finkd opened Muse connector development to outside APIs, a consumer opening @VijarKohli calls a real first-mover play given Meta's assistant plus WhatsApp and Instagram DMs; @FactoryAI announced Factory Private for VPC, on-prem, and air-gapped deployment (@AlexTriplett); @claudeai's Projects beta runs parallel threads from one Claude Code conversation, and @eliasstravik already shipped a herdr-projects plugin for it; Cloudflare released the security-audit skill it used internally (@MoureDev).

Jev posts Vercel's fastest adoption curve, then goes free for a week

The day's biggest signal is a velocity number. Quoting @vercel's data, @rauchg says "the data and the anecdata on Jev's adoption are shocking," with day-one adoption roughly double the GPT-5.6 family and six times Fable 5.1. He thinks the product is good but the surge is downstream of cost and latency anxiety: "People are eager to optimize and put AI in even more places." @vercel_dev then made Jev free on AI Gateway through Sept 25, prompting @rauchg to declare his weekend plans settled.

What Jev actually is remains a set of claims, not settled fact. @sydneyrunkle amplified @matthewcanham's framing of it as "a new type of model that is optimized for making decisions, not generating text." In the launch post @OhansEmmanuel quote-tweeted, @CompleteSkeptic claims to have co-invented ChatGPT, spent two years in stealth on a training method called RLCD, and shipped Jev at "20-200x faster" and "40-400x cheaper" with output tokens free. For orientation, @AIGuide_ points to @sydneyrunkle's "Building a Harness with Jev" as required study: "if you're confused about Jev this is 100% worth your time."

Builders keep finding decision-shaped work for Jev

The most concrete Jev posts are about offloading decisions, not prose. @michael_chomsky is offering free bulk classification for personal agents (tweets, emails, Codex chats), says Jev is finally cheap enough to make that free forever, and predicts the capability gets baked into harnesses like Claude Code and Codex "kinda like web search/scraping/page summarization/compaction."

@kwindla describes "greedy inference" from his Pipecat pipelines: run checks on partial inputs and discard bad output. The blocker was always tool calls, since an LLM must emit an entire tool stanza token by token before execution. Jev changes that, running in parallel to emit a structured event that triggers the tool while the LLM is still streaming text. He builds on @BhosalePratim, who replaced the LLM's decision step with Jev so voice agents can act before a user finishes their sentence, at lower cost.

@marcklingen sees an evals story: every question is evaluated in parallel and in isolation against the same state, so adding a fourteenth question barely changes response time and cannot degrade the others. "You can ask speculatively and throw away what you do not need." @OhansEmmanuel's use case is enforcing coding-agent rules a linter can't codify, scoring every turn and telling the agent what to fix, open-sourced. @MiguelriosEN built AgentRun with @pidotdev and Jev, a harness where an agent learns a job, codes a general solution, "and then get[s] out of the way." And @amal_irgashev endorses @ctatedev's json-render plus Jev experiment, which renders your own components in milliseconds and, in his view, finally abstracts presentation logic away from the agent: "Gen UI is so back."

Free and open alternatives show up before the launch dust settles

@jun_song's blunt take: "An OpenAI co-founder spent 3 years quietly building Jev, only for someone to open source a better performing model in just 3 days," which he reads as evidence that starting an AI company right now carries too much risk. He's quoting @0xCVYH, who points to Laya, an open-source Jev that reportedly already benchmarks above it. @michael_chomsky piles on, saying his own free classifier "now outperforms jev."

Even teams without Jev access are substituting for it. @tekbog flags GitHubNext cobbling together a "poor man's Jev" on omlx in a morning, benchmarking diffusiongemma, Qwen MoE, Gemma 4 MoE, and Gemma 4 e4b/e2b, with decent evals and speeds on an M5 Max 64GB, and exposing a LocalJev server compatible with standard Jev API wrapper libraries. "This is 100% promptcoding but hey the evals look okay." For perspective, @mstockton (echoing @_chenglou) notes Google already did decision-model-style work for timeseries and tabular data with TabFM and TimesFM, and "the UX and the GTM just didn't hit." Technology is never the only thing.

Small specialists embarrass general agents

@trycua open-sourced CUA-S1-FORMS, the first in a family of "System One Models" for computer use. @be_arsh's numbers: a 706k-parameter model fills an entire form in one 50ms pass, while an LLM agent took 23 turns and 39.6 seconds. His verdict: "the specialists are going to eat the generalists from the bottom."

The same logic applies to shrinking generalists. @0xSero highlights @PrismML's Ternary Bonsai 2 27B, a ternary Qwen3.8-27B that is 9x smaller at a 5.9 GB footprint while retaining 98.2% of aggregate benchmark performance, with claimed gains in agentic coding, multimodal reasoning, and long-horizon tool use, all Apache 2.0. He calls it the smartest model runnable on 8-16GB of VRAM or 16-24GB of Mac memory. @sairahul1's advice fits here: "Learn to fine tune a model this year. Start small. Start slow. But make sure you start," with a full guide on fine-tuning a small LLM on your own data.

Supervising agents: heuristics can't certify heuristics

@DominikTornow argues "the current thinking around software factories is wrong": better workflows, skills, and instructions, or more agents checking agents, won't produce correctness, because "a heuristic cannot certify the result of a heuristic." @rjs endorses @istoica05's "Two Key Gaps in Agentic Software Engineering" as a clean articulation of the underlying problem, that requirements can't be fully pre-stated in advance.

@unclebobmartin's answer to the supervision bottleneck is a custom interrogation tool: a dynamic UML-like view of the system, color-coded with crap and mutation scores, drillable from architecture down to code, so a human can spot weak structures and direct an agent to refactor them. He suggests pointing an agent at his tool to build your own variant.

@KengGuangLong relays Matt Pocock's prompt-level fix for agents that write the whole database layer, then application, then frontend, and only discover mismatches at assembly: drop "tracer bullet" from The Pragmatic Programmer into the prompt. Pocock found the agent then uses the term in its own reasoning, a "leading word" that activates knowledge already in training data, alongside "deep modules" and "ubiquitous language." @mndoci recommends the Kiro Crew's post on shipping a lot as a small team, a follow-up to @clare_liguori's manifesto of ten principles for building with AI drawn from teams across Amazon. And @omarsar0's NVIDIA thread treats the harness itself as the search space: auto-research loops kept four mechanisms (Action Fusion, Online Context Compact, ObservationPack, Evidence-Preserving Reducer) that survived selection across many environments, cutting token traffic nearly in half while matching baseline harnesses on GPT-5.6 Sol and Opus 5, with code under NVlabs on GitHub.

Practical Takeaway

Jev is free on Vercel through Sept 25, so the cheapest experiment this week is routing one decision-shaped slice of your agent, classification, routing, per-turn rule checks, or eval questions, through it in parallel with your LLM, exactly as @kwindla and @marcklingen describe, and measuring latency and cost against your current stack. Given @jun_song's observation that open-source equivalents appeared within days, prototype against a small fine-tuned local model too (@sairahul1's guide is a starting point) and keep the interface swappable, because no model choice in this segment looks durable yet.