Jev 为何输出免费：决策模型没有解码循环

Straggler Liu | AI & Semis @StragglerLiu原推发布于 2026.10.06

在 X 上打开原推

Jev Doesn't Change Frontier Lab Spending. It Changes Something More Important. TypeSafe AI released Jev on September 15, and within 24 hours it reached nearly 13% of Vercel's paid teams — twice the adoption rate of the GPT-5.6 family and more than six times that of Fable 5.1. Every reaction I've seen treats it as either a cost optimization story or a capability story. Neither captures what's happening. Jev is a decision model. You send it unstructured state and typed questions. It returns Choice, Score, or Noul answers with calibrated probabilities in a single parallel pass. There is no autoregressive decode loop. That architectural detail is the entire story, and it has almost nothing to do with model size. Separate the two things a generative model does when you ask it a yes/no question. First, it computes the answer — one forward pass producing logits. Second, it writes the answer out token by token — the decode loop, where latency lives and output tokens accumulate. For a classification task, that second step produces zero additional information. The model knew the answer after the first forward pass. Everything after is transcription. Jev eliminates the transcription. It enumerates the legal output space before the call, runs one parallel forward pass, and returns a softmax over the options. There is no decode loop because there is nothing to decode. The output is a probability distribution, not a string. That is why Jev's pricing looks the way it does: $0.042 per million input tokens, output free. Not a marketing gimmick — a direct consequence of the architecture. Output tokens don't exist, so they can't be billed. A frontier model's output tokens typically cost 3-5x more than input, and a classification task generating JSON can emit 50-200 tokens of reasoning and formatting that carry no decision-relevant information. The cost difference is structural. Harness moved the yes/no decision step in its AppSec pipeline from a generative LLM to Jev and cut per-endpoint classification cost by roughly 150x against a single-shot LLM call, 1,200x against their original multi-step pipeline. OpenRouter's benchmark on a 40,000-ticket/month triage scenario put Jev at $0.0248 per thousand tickets versus $2.88 for Claude Opus — 1/116. An independent benchmark repository found Jev matching or exceeding two Claude models on core classification at 1/50 to 1/116 of the cost. Jev does not reduce frontier lab spending. Agentic workflows still require long-horizon reasoning — planning, multi-step code generation, error recovery, open-ended research. Jev cannot do any of that. It returns typed answers, not plans. The frontier model it displaces is the one being used as a router or classifier in the middle of an agent loop, not the one doing the reasoning at the top. The 150x reduction is not 150x more capability. It is 150x less overhead. The frontier model was never the right tool for the classification step — it was the only tool that could do classification and reasoning in the same call. Jev separates those concerns. AgentRouter's ICML 2026 paper quantified the scope. Enterprise agent systems waste 60-80% of their inference budget routing every trajectory step to a frontier model when smaller models handle the majority of steps equally well. Their per-step routing classifier achieved 72% cost reduction with under 3% task-completion degradation. Jev suggests a more radical version: for the 60-80% of steps that are classification, routing, or scoring decisions, the right architecture may be no LLM at all — a decision model with a probability read-out instead of a generation head. RLCD adds a second dimension. RLHF optimizes for preferred responses. RLVR optimizes for verifiable outcomes. RLCD optimizes for whether the stated probability matches the empirical frequency — a 0.8 decision should be correct about 80% of the time. That property is what makes a probability actionable. It is the difference between a confidence score and a control surface that lets you route the top 30% to automation and the bottom 70% to human review. Calibration is also where independent verification breaks down. A pre-registered study from NXTG AI, published October 1, tested Jev's calibration directly. Pooled test ECE was 0.156, with systematic overconfidence — mean confidence 0.66 against actual accuracy 0.53. In four of five decision classes, Jev was not calibrated as claimed. The study concluded that Jev's probabilities should not set a routing threshold without per-class measurement. If that finding generalizes, Jev's application range narrows from "automated decision layer" to "very fast, very cheap classifier." The two imply very different market sizes and valuation frameworks. The former is an architectural change; the latter is a cost optimization. The competitive response arrived within two weeks, and it was architectural. OpenAI's Dev Day on September 30 introduced the Decisions API — a specialized Luna model with predefined options returning results in roughly 150ms. AWS open-sourced Strands Decider 2B, a 2-billion-parameter decision model running locally under 150ms. Cloudflare released the Clef series at $0.24 per million tokens. Databricks shipped ai_decide. TypeSafe CEO Diogo Almeida's response: "Fast and cheap is very easy. If you want it really fast and cheap, use dice. Intelligence is the hard part, and my North Star is always pushing the intelligence-per-dollar Pareto curve." That is a calibration play, not a price play. If the calibration advantage holds, the pricing premium is defensible. If the NXTG AI finding generalizes, the moat narrows to synthetic training data — a data advantage, not an architectural one. Data moats are replicable with enough investment. Architectural moats are not. Once you separate "compute the answer" from "write the answer," every inference workload can be re-examined. How much of what you're paying for is generation, and how much of that generation actually produces decision-relevant information? For most agentic pipelines, the answer is: very little. What to watch: whether the calibration gap closes. NXTG AI's study is pre-registered and reproducible — a harder standard than vendor self-evaluation. If independent calibration curves validate Jev's probabilities, the automated decision layer thesis holds. If they don't, Jev remains a very fast, very cheap classifier. Also watch whether OpenAI's Decisions API publishes calibration figures. A frontier lab that ships a decision model without calibration data is implicitly conceding the calibration problem is unsolved.