FULL STORY

TypeSafe Emerges From Stealth: Jev Decision Model Launches and Gets Benchmarked

After two years in stealth, TypeSafe launched Jev, its first System One decision model, with the new RLCD training method. The model was quickly integrated into Vercel and Cloudflare gateways, with early benchmarks showing it several times faster and far cheaper than the fastest LLMs.

2026-09-16 ~ 2026-09-18 · 9 episodes · 153 posts

Episode 1 · TypeSafe exits stealth with decision model Jev and RLCD training method (2026-09-16, 86 posts)

On September 16, TypeSafe ended roughly two years in stealth, releasing Jev, its first System One model, along with a new training method called RLCD. Founder Diogo Almeida (@CompleteSkeptic) says he worked on instruction-following research behind ChatGPT at OpenAI, and many reposts call him a "ChatGPT co-inventor," though this identity is self-described and unverified. All performance claims are vendor-side, with no independent evaluations yet, and some skepticism has already surfaced.

Confirmed

  • TypeSafe released the model Jev and the training method RLCD; per @rohanpaulai, Jev is the company's first System One model, and multiple reposts describe Almeida as an InstructGPT paper co-author.
  • Jev is not a chat model but "decision intelligence": it takes natural-language questions (including ambiguous or subjective ones) and returns probabilities for yes/no or custom categories via a pure API, with no text generation. A typical use case relayed by @hardimanjames: asking whether a support ticket customer sounds angry, returning a probability like 0.9 in about 0.7 seconds. @bindureddy characterized Jev as a classifier that makes intelligent decisions among options, arguing such near-free classification is a great complement to agent platforms; she further proposed that small models outputting option probabilities in a single forward pass could act as decision-makers, easing the serial bottleneck (tool selection, branching, retries) in current agent architectures and enabling thousands of parallel steps.
  • The division of labor: large models handle hard reasoning while massive small decisions (routing, fraud detection, moderation, verification) go to models like Jev. Technically it skips autoregressive token generation, mapping unstructured state to type-safe probabilistic decisions in parallel. @FrankFelixAI likened this to how Transformers overtook RNNs by replacing serial with parallel computation—fast and cheap, but incapable of generating text. The AI Daily Brief's video offered a similar reading.
  • Per @Profanion citing The Register, the vendor claims far lower hallucination rates than traditional LLMs and much cheaper outputs.
  • Vendor claims: 20-200x faster, 40-400x cheaper than existing models, free output tokens, sub-500ms responses, no hallucinations. @heyabusiddik relayed input pricing of $0.042 per million tokens; @TheMoonMidas relayed 155x cheaper and 20x faster than Opus; @soumitrashukla9 relayed 25x faster and 600x cheaper than Fable-level judgments—figures vary across reposts, all originating from the vendor.
  • The vendor also demoed Jev playing Doom in real time to showcase low-latency decision-making (per @alexcovoeth).

Unconfirmed

  • Speed, cost, and "zero hallucination" claims lack systematic independent evaluation or benchmark verification.
  • "ChatGPT co-inventor" is self-described; @bennash explicitly flagged it as unverified. @mohbibi's repost noted skepticism that Jev is merely a specialized diffusion model, and @karmay007 mentioned the claims were mocked online, with no official response.

Why it matters

  • If the claimed latency and cost hold, Jev could dramatically lower the barrier to embedding many judgment calls in automated workflows. @bindureddy's architectural thesis—that decision-makers can free agents from serial bottlenecks—is a hypothesis worth tracking.
  • "Structured probabilistic output instead of text generation" represents a product path divergent from mainstream chat models; third-party evaluations, founder identity verification, and technical details merit follow-up.

66 more related posts →

Episode 2 · TypeSafe AI launches Jev, a dedicated evaluation model showing major speed and cost gains in tests (2026-09-16, 8 posts)

TypeSafe AI launched Jev, a dedicated evaluation model now available on Vercel AI Gateway. Built with involvement from LangChain's Sydney Runkle and Hunt Lovell, the company claims up to 200x speed and 400x cost improvements over comparable LLMs on classification. Independent developer tests show significant speed and cost advantages with quality matching or exceeding prior flagship lightweight models, fueling discussion of replacing general LLMs with small specialized models for classification and routing.

Confirmed

  • Per @hackgoofer, Jev is a "System One" evaluation model for fast structured judgments, returning options, scores and boolean probabilities, and is available via Vercel AI Gateway.
  • Developer cramforce dropped Jev into a classifier benchmark previously based on Gemini 2.5 Flash Lite: it saturated the evaluation (100% accuracy) while running 6x faster.
  • @jonreed's Near Here team, with early access, used it for local event data validation with per-model prompt tuning: up to 5.7x faster responses, 98% lower cost, and 12 percentage points higher accuracy.
  • @hardimanjames relayed N8Programs' tests across multiple-choice benchmarks (MMLU, GPQA etc.), validating that Jev's System 1 intelligence rivals GPT-5.6 Terra.
  • @manubfr tested Jev as an alignment monitor on 4 public benchmarks for harmful-content judgment, reporting it beats mainstream options at the lowest price.
  • Per @hwchase17's repost, Jev was built by LangChain team members Sydney Runkle and Hunt Lovell, with official claims of up to 200x speed / 400x cost improvements.

Unconfirmed

  • All comparison figures come from third-party or early-access self-reports without independent replication; "Luna-level intelligence" and the 200x/400x figures are vendor claims.

Why it matters

  • Four independent production/benchmark swaps point to the same conclusion: for structured judgment tasks, dedicated evaluation models can achieve equal or better quality at far lower latency and cost than general LLMs.
  • If reproducible, Jev is a strong A/B candidate for high-throughput classification, routing, review, and content-safety monitoring workloads.

Episode 3 · Vercel fx to adopt Jev safety reviewer, up to 18x faster (2026-09-17, 3 posts)

Vercel CEO Guillermo Rauch confirmed that fx's default auto-mode safety reviewer will switch to typesafeai's Jev, which benchmarks show is roughly 5-18x faster and more accurate than the current GPT-5.6-Luna classifier.

Episode 4 · TypesafeAI Launches Zero-Shot Classifier Jev on AI Gateway (2026-09-17, 2 posts)

TypesafeAI's zero-shot classifier Jev is now live on AI Gateway for all users with zero-data retention. The cheap, fast tool promises to significantly cut text labeling costs.

Episode 5 · TypeSafe Launches Jev, a Decision-Only Model That Never Writes a Word (2026-09-17, 44 posts)

TypeSafe (founded by Diogo Almeida, one of the researchers behind ChatGPT/InstructGPT) has released its first model, Jev, positioned not as a better chatbot but as a "decision engine": given an application state plus a set of typed questions, it directly outputs options, scores, and probabilities for code to execute, without generating any text. The company claims it is 20-200x faster and 40-400x cheaper than LLMs (output tokens are free; input pricing works out to about $0.042 per million tokens), with latency of 70-500 milliseconds, and uses a new training method called RLCD. In a demo, Jev played Subway Surfers in real time, showcasing real-time decision-making.

Confirmed

  • Jev takes a "System One" approach: instead of generating text token by token, it directly selects a decision from a fixed set of candidate answers with a confidence score, in contrast to the LLM's token-by-token "System Two" generation
  • Latency of 70-500ms, free output tokens, positioned as a third path beyond hybrid classifiers and LLMs
  • Early hands-on from independent AI engineer @mostlyidempotent confirmed it returns structured decisions rather than text, with latency around 70ms
  • @mikegiannulis ran an offline replay test (real historical inputs + manually labeled ground truth, with pass/fail rules preset in advance): for manuscript material reranking, embedding similarity ranked the correct passage first 60% of the time, while Jev raised it to 90%; portal intent routing scored 90.5% vs. 83.8% for the original keyword routing; "I've already done this" detection hit 98.3% vs. 96.6% for the regex approach; 95% accuracy and Brier 0.057 on 148 faithful/fabricated pairs; the replay also caught a customer service message the old classifier had misjudged in production to the team's embarrassment (e.g., "please text, don't call" was classified as other)

Unconfirmed

  • "20-200x faster and 40-400x cheaper" are official claims; independent benchmarking remains insufficient; the founder's identity also carries传播 noise—most reposts say "ChatGPT co-inventor," while @adrianscottcom noted Diogo Almeida is an InstructGPT co-author; the exact attribution varies
  • Small-scale replays can't prove perfect calibration in production, and @mikegiannulis himself explicitly remained measured

Why it matters

  • Jev represents a "small model fills the gap" paradigm: not replacing LLMs, but replacing the roundabout practice of "have an LLM write a paragraph then dig out the answer" with probabilistic outputs for deterministic yes/no, routing, and classification tasks
  • @mikegiannulis also disclosed a failure list: their own LLM won about 2:1 on subtly-worded author emails, it couldn't help with subjective judgments like "this seems too salesy," and showed no improvement on data-sparse sales lead scoring—blind spots like these can't be found on AI leaderboards, only exposed by replaying your own data, providing a rare honest reference for the model's applicability boundaries

24 more related posts →

Episode 6 · TypeSafe AI's Evaluation Model Jev Lands on Vercel and Cloudflare Gateways (2026-09-17, 4 posts)

TypeSafe AI launched Jev, a millisecond-level evaluation model for agent decisions, routing and scoring, now available via Vercel and Cloudflare AI Gateways at $0.04 per million input tokens.

Episode 7 · Typesafe AI's New Model Jev Delivers Fast, Cheap Classification (2026-09-17, 2 posts)

Typesafe AI released Jev, a new model focused on System 1 fast classification rather than long-chain reasoning. Developer tests show it is over 5x cheaper, faster, and more accurate than smaller models previously used for cost-constrained safety pipelines.

Episode 8 · Jev Classifies Tweets 6x Faster and 40x Cheaper than Fastest LLM (2026-09-17, 2 posts)

Developer altryne swapped TypeSafe's Jev into his Chrome extension that filters the Twitter "For You" feed, finding it classifies tweets 6x faster and 40x cheaper than the fastest LLM on Cerebras.

Episode 9 · Jev Outputs Structured Decisions Directly at $0.042/M Tokens (2026-09-18, 2 posts)

TypeSafe AI's Jev skips text generation and directly outputs options, scores, and probabilities at $0.042/M tokens. An AI book-writing company validated it by replaying 3,300 real production decisions offline in an afternoon for just 20 cents.