ChatGPT Co-Inventor Launches Jev: Drops Text Generation, Claims AI Reasoning Costs Cut 400x

ChatGPT co-inventor Diogo Almeida launched the decision model Jev in September 2026, and his company TypeSafe AI officially claims that the model is 20 to 200 times faster than large language models and 40 to 400 times cheaper, priced at $0.042 per million input tokens with output tokens free. This is the most direct public challenge to the "generation paradigm" in the large-model era to date, and the challenger himself is a co-creator of that paradigm.

The Inventor Rebels Against His Own Invention

Diogo Almeida's identity needs a precise description. He is not the "inventor" of ChatGPT—ChatGPT is an engineered product with hundreds of engineers behind it. His accurate role is: co-author of the InstructGPT paper and co-contributor to RLHF (Reinforcement Learning from Human Feedback), the training method. It was InstructGPT that laid the technical foundation for ChatGPT's ability to understand user instructions, and Almeida is also listed on the GPT-4 technical report.

He now calls the RLHF era a "weird detour" in AI. This is not marketing talk, but criticism with clear technical logic: RLHF's training objective is to maximize human preference scores, which causes the model to optimize toward "making people satisfied" rather than "getting the task right." In human-supervised dialogue scenarios, this mechanism works well. But in agent automation scenarios, the same mechanism becomes a hidden risk—the model does not know when it should stop and report uncertainty, tends to "fill in answers" under ambiguity, causing downstream systems to cascade-fail after receiving incorrect classifications.

This is the direct reason Almeida bypassed RLHF when designing TypeSafe after leaving OpenAI.

Jev's Architecture: Replacing Generation with Judgment

TypeSafe's official blog disclosed that Jev uses three proprietary technologies: a completely new model architecture, a parallel sampler (which generates all possible outputs simultaneously rather than unfolding step by step as a word sequence), and a training method called RLCD (Reinforcement Learning for Calibrated Decisions).

The most fundamental difference from LLMs is the task boundary. Give Jev input data and a preset question, and what it returns is not a readable passage of text, but type-determined options, a probability distribution, and a confidence score. It cannot generate explanations, nor can it reason, but it knows how confident it is—this is precisely the attribute most lacking in existing LLMs in automation scenarios.

TypeSafe claims Jev's response time is about 150 milliseconds and that, by spec comparison, its price-performance "has nearly two orders of magnitude of Pareto frontier advantage." But this benchmark suite was designed by TypeSafe itself, using the predicted probabilities of external large models such as Astra and Fable as a reference, not from an independent third-party organization.

Independent Testing: Fast, but Not Perfect

Tech media outlet Every.to conducted hands-on testing soon after Jev launched. According to the outlet, the tester submitted 37 articles and 21 evaluation questions to Jev at the same time, and it completed them in under 0.7 seconds; in paragraph-by-paragraph comparison testing, Jev's median response time was 0.35 seconds, while the comparison model's was 8.83 seconds, about 25 times faster. Across 11 rounds of experiments, totaling 1,709 judgments, the total cost was less than one cent.

On accuracy, the tester deliberately planted 7 defects in the corpus; Jev detected 6 and missed 1. The tester commented that the results "performed well but were not perfect," and before use in production systems, they still hoped to conduct more systematic accuracy testing.

This is the only publicly available independent test that can currently be verified, and the sample size is limited, but the speed and cost figures are consistent in direction with the order of magnitude of the official claims.

The Industry Significance of Cost Restructuring

The significance only becomes apparent when the cost figures are placed into actual workflows. An enterprise-grade content moderation system needs to perform classification, prioritization, and routing decisions on millions of pieces of content every day. At the input pricing of mainstream OpenAI or Anthropic models, which ranges from $0.25 to $10 per million tokens, the decision layer alone could cost on the order of hundreds of thousands of dollars per month. If the unit cost of equivalent decision tasks is compressed to less than one percent, the economic feasibility boundary of entire AI agent applications will undergo a structural change—not in theory, but an ROI that engineering teams can recalculate today.

This also explains why some AI practitioners call Jev "the most underrated release of the year." It is not a stronger text model, but an infrastructure layer specifically built to carry the subtask of "judgment actions." Current AI application architecture is becoming layered: LLMs handle understanding and generation, while decision-intensive operations such as classification, scoring, and routing are all seeking lighter-weight alternatives in terms of cost and latency. Jev appears in the gap of this structural demand.

An Unverified Hypothesis, but the Direction Is Real

TypeSafe has so far not submitted Jev's evaluation results on any public standard benchmark set, and all existing performance figures come from its internal evaluation system. "40 to 400 times cheaper" is a range rather than a precise value, depending on the choice of comparison target and task type, and this range itself suggests a high degree of conditional dependence.

Independent judgment: Jev's product logic holds up, and this matters more than the performance claims themselves. Splitting AI workflows into "generation" and "judgment" and designing a model specifically for judgment—this classification framework has clear engineering persuasiveness regardless of what Jev's final numbers turn out to be. At OpenAI, Almeida personally experienced the success of RLHF and also observed enough agent failure cases; his choice to rebuild the training objective from first principles is not a purely commercial narrative.

But a correct technical direction does not equal execution delivered. Jev has not yet undergone sufficiently large-scale, sufficiently diverse external testing; its "zero hallucination" design promise needs to be validated in edge cases; and whether its calibrated probabilities truly reflect model capability, rather than being another form of performative confidence, still requires more evidence.