ModelsArchitectures & capability
TypeSafe AI's Jev decision model reaches 35% of Vercel AI Gateway paid teams in two weeks
Jev, a model that returns typed decisions with probabilities instead of text, reached 35.3% of paid Vercel AI Gateway teams within two weeks. That is the gateway's fastest uptake, though Jev accounted for only about 1% of tokens and 0.15% of spend.
From the TechCrunch AI article of October 16, 2026 and cited sources: Many teams tried Jev, but each used it lightly. [2] [1]
SOURCES[2] Vercel: AI Gateway Production Index, October 2026 vercel.com[1] TechCrunch: Jev is Vercel's fastest-adopted AI model ever… techcrunch.com
AT Pulse research deskPublished · Updated · 4 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- TechCrunch AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 16, 2026 article published
- Covered
- Oct 18 edition
- Evidence
- 10 sources across 5 sites
- 1TechCrunch: Jev is Vercel's fastest-adopted AI model ever…techcrunch.com
- 2Vercel: AI Gateway Production Index, October 2026vercel.com
- 3TechCrunch: A new kind of AI model from a ChatGPT inventor…techcrunch.com
- 4InfoQ: TypeSafe AI Releases Jev (Oct 1, 2026)infoq.com
- 5Vercel: What is Jev?vercel.com
- 6TechCrunch: OpenAI's Jev clone (Sept 30, 2026)techcrunch.com
Reach: not measured yet
Many teams tried Jev, but each used it lightly. Its 35.3% reach among paid teams in two weeks is nearly double Sonnet 5.5's 18.1%, and week-two retention was 55.6%. Even so, Jev handled about 1% of monthly gateway tokens and 0.15% of spend. [2] [1]
Jev takes a state plus a set of typed questions (Boolean, Choice or Score) and answers them all in one parallel pass, returning a probability with each answer rather than generated text. It has a 32K-token context and 70–500 ms latency, and costs $0.042 per million input tokens with free output. [4] [10] [3]
Similar products followed within weeks. OpenAI launched a Decisions API on September 30, and AWS open-sourced Strands Decider 2B on October 1. Researchers have also released dozens of similar decision models, which suggests the category could quickly become a commodity. [6] [7]
Jev is unreliable at counting, arithmetic and comparing dates, and its accuracy drops on large, noisy inputs. It can return an answer that looks valid but is wrong. The claims of up to 194× faster and 445× cheaper than LLMs come from TypeSafe itself, and the field reports so far are anecdotal. [4] [8] [3]
On one gateway, many teams have tried a cheap decision layer.
Read the full assessment
The likely implication, not yet measured, is that classification, routing, moderation and agent monitoring could move off frontier LLMs, cutting revenue per call as OpenAI and AWS offer alternatives.
Executive brief
TypeSafe AI's Jev returns decisions, not text. According to Vercel's AI Gateway Production Index, 35.3% of paid teams on its gateway used Jev within two weeks of launch. That is nearly double Sonnet 5.5's 18.1% and the fastest adoption the gateway has recorded. Usage was wide but light: Jev handled about 1% of monthly tokens and 0.15% of spend. That fits its design, which picks from a fixed set of typed answers cheaply. OpenAI and Amazon released similar products within weeks. The adoption figures come from one gateway, which TechCrunch notes may undercount labs whose customers call their APIs directly.
What changed and event timeline
Jev launches
TechCrunch gives September 15 as the. Vercel dates the launch, and an earlier says September 18.
- Also
Day 2 on the gateway: 13% of paid teams in 24 hours
Vercel added Jev to AI Gateway on day two. and say it reached nearly 13% of paid teams within 24 hours.
OpenAI ships a Decisions API
OpenAI announced it at Dev Day. It does the same job: pick from preset options and return probabilities. TypeSafe's CEO called it the start of "the clone wars" ().
AWS open-sources Strands Decider 2B
It is a decision model inspired by Jev. TechCrunch says researchers have produced dozens of similar models ().
Vercel publishes the adoption data
The Production Index reports 35.3% paid-team reach in two weeks, the most requests of any model in September, and 55.6% week-two retention ().
Capabilities and access
- It is available on Vercel AI Gateway and through an
evaluatemethod in AI SDK 7 (Vercel). Forbes reports that Cloudflare also added it (Forbes). - Answer types: Boolean, Choice and Score, each returned with a probability (Firecrawl).
- Specs: 32K-token context, 70–500 ms end-to-end latency, $0.042 per million input tokens, output free (InfoQ).
Read the full section
- Model:
typesafe-ai/jev. It is available on Vercel AI Gateway and through anevaluatemethod in AI SDK 7 (Vercel). Forbes reports that Cloudflare also added it (Forbes). - Answer types: Boolean, Choice and Score, each returned with a probability (Firecrawl).
- Specs: 32K-token context, 70–500 ms end-to-end latency, $0.042 per million input tokens, output free (InfoQ).
Technical analysis for researchers and developers
- Interface: You send a state (a string or structured data) plus a set of typed questions.
- Training: TypeSafe calls its method "Reinforcement Learning for Calibrated Decisions."
- Evaluation: No standardized calibration benchmark from TypeSafe has been published. A community "Jevbench" leaderboard exists for decision models (TechCrunch).
Read the full section
- Interface: You send a state (a string or structured data) plus a set of typed questions. Jev answers all of them in one parallel pass (InfoQ). It is a transformer that outputs calibrated probabilities instead of tokens (TechCrunch).
- Training: TypeSafe calls its method "Reinforcement Learning for Calibrated Decisions." The details are proprietary, and no paper or model card appears in the reviewed sources, so the results cannot be reproduced.
- Evaluation: No standardized calibration benchmark from TypeSafe has been published. A community "Jevbench" leaderboard exists for decision models (TechCrunch).
- Implementation: Because outputs are typed, they don't need parsing, but they still need handling for low-confidence answers.
Claims and evidence
- Fastest adoption ever: This comes from Vercel's index and covers its gateway only. Vercel is also an early Jev customer (TechCrunch).
- Up to 194× faster and 445× cheaper than LLMs: This is TypeSafe's own claim, with no independent replication (Vercel).
- Field reports: A Vercel engineer reported 5–18× speedups over the classifiers Jev replaced. These are anecdotes and sentiment data, not controlled evaluations.
Read the full section
- Fastest adoption ever: This comes from Vercel's index and covers its gateway only. Vercel is also an early Jev customer (TechCrunch).
- Up to 194× faster and 445× cheaper than LLMs: This is TypeSafe's own claim, with no independent replication (Vercel).
- Field reports: A Vercel engineer reported 5–18× speedups over the classifiers Jev replaced. Bryo AI's CTO said Gemini cost 10–20× more for the same task (TechCrunch). An OpenChamber analysis of 12,759 tweets found a median 7× speedup and 30× cost savings (InfoQ). These are anecdotes and sentiment data, not controlled evaluations.
- What Vercel replaced: The sources disagree. One TechCrunch article says GPT-6 Luna (Oct 16); another says ChatGPT Luna 5.6 (Sept 18).
Context and prior work
TypeSafe was co-founded by Diogo Almeida, who co-invented RLHF and worked on ChatGPT research, along with Erik Gafni and Sasha Sheng (InfoQ). The same index shows open-weight models processing over 70% of gateway tokens and nearly 25% of spend. Frontier-lab spend still grew about tenfold since December (Vercel).
Read the full section
TypeSafe was co-founded by Diogo Almeida, who co-invented RLHF and worked on ChatGPT research, along with Erik Gafni and Sasha Sheng (InfoQ). The company markets Jev as a "System One" model: a fast, intuitive component that sits alongside slower LLM reasoning (Vercel). The same index shows open-weight models processing over 70% of gateway tokens and nearly 25% of spend. Frontier-lab spend still grew about tenfold since December (Vercel).
Limitations, safety and contested findings
- Known weaknesses: Counting, arithmetic and date comparison are unreliable, and accuracy drops on large, noisy state.
- "Cannot hallucinate": This claim (TechCrunch) only means Jev produces no free text. Wrong decisions are still possible.
- Interpreting confidence: Armin Ronacher notes that low-probability outputs need interpretation. A 50% answer may be one to throw away.
Read the full section
- Known weaknesses: Counting, arithmetic and date comparison are unreliable, and accuracy drops on large, noisy state. The model can return a valid-looking but completely wrong answer (InfoQ).
- "Cannot hallucinate": This claim (TechCrunch) only means Jev produces no free text. Wrong decisions are still possible.
- Interpreting confidence: Armin Ronacher notes that low-probability outputs need interpretation. A 50% answer may be one to throw away.
- Measurement gaps: Launch dates conflict across sources, and the adoption data covers a single gateway.
Business and practitioner implications
- Cost: Classification, routing, moderation and agent monitoring can move from frontier LLMs to decision models.
- Reach vs. revenue: 35% of teams produced only 0.15% of spend (Vercel). Decision models spread widely but earn little per call.
- Competition: With OpenAI and AWS alternatives available, commoditization looks likely. Benchmark calibration on your own data before committing.
Read the full section
- Cost: Classification, routing, moderation and agent monitoring can move from frontier LLMs to decision models. One researcher's demo put agent monitoring at $2.94 with Jev versus $372 with frontier LLMs (TechCrunch).
- Reach vs. revenue: 35% of teams produced only 0.15% of spend (Vercel). Decision models spread widely but earn little per call.
- Competition: With OpenAI and AWS alternatives available, commoditization looks likely. Benchmark calibration on your own data before committing.