Jev gives software a decision, not another paragraph

TypeSafe AI's first System One model accepts text state plus typed questions, then returns constrained values, probabilities, and confidence in parallel. The launch promises LLM-level judgment for narrow decisions at a fraction of the latency and cost. Jev is accessible through SPIRITT AI Gateway.

Use Jev in SPIRITT

See the evidence

Jev builders and evaluations across X

28 source posts: adaptive forms, real-time games, browser automation, model routing, app testing, Mac troubleshooting, safety monitors, and published evaluations. Builders describe their own setups; in SPIRITT, just ask your workspace to use Jev.

introducing JevForm, a form that dynamically branches and chooses what to ask next usinng @typesafeai’s Jev in my life i’ve made hundreds of forms with crazy if/then logic. Jev solves it. built with @vercel json-render (by @ctatedev), so theoretically it can support any generative form UI, and @DavidKPiano’s xstate for the actual state Play with it here: https://jevform.spiritt.app/
— Tamir (@TamirSPIRITT) September 18, 2026

we are officially out of stealth! join the frontier and get access to Jev on our website (link on profile)
— TypeSafe AI (@typesafeai) September 15, 2026

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x cheaper (w/ output tokens free) • Frontier composable intelligence optimized for decisions AFAICT the shortest path to AI-based economic revolution
— Diogo Almeida (@CompleteSkeptic) September 15, 2026

got @typesafeai’s new model Jev to play Super Mario Bros. fast inference + structured outputs makes it surprisingly good for real time use cases. I'm excited to see what can be done with these new models!
— Faadil Shaik (@faadilhshaik) September 16, 2026

JEV CAN DRIVE! 🤯 I hooked up Jev to the raw controls of a driving simulator, this video shows him driving in realtime. Realtime is the kicker here. Asking a model what to do every 50ms then rendering that into a "realtime" video is one thing, but in this simulation I wanted to give Jev the challenge of actually controlling a moving vehicle in a simulation that will never pause while he thinks.
— Toran Bruce Richards (@SigGravitas) September 16, 2026

Breaking: Browser Use + Jev = Ultrafast ⚡ Findings flights took 7s and cost only $0.0039 🤯 > new action space every step > DOM state space > small LLM fallback to type (this video is at 1x speed btw) Built a tiny open source browser agent. try it below ↓
— Gregor Zunic (@gregpr07) September 17, 2026

i built a codebase classifier with Jev and this might be the solution to overengineered code that agents create what should I test Jev on next?
— vogel (@ryanvogel) September 16, 2026

i made an llm from first principles with Jev 29 yes/no questions per character: should the next key be a–z, space, comma, or period? highest probability gets append to it, then fed the updated text back in & repeat an autoregressive loop made out of a classifier
— vogel (@ryanvogel) September 16, 2026

preview of how fast browser use can be powered by typesafe's jev and opencode's browser use cli
— dax (@thdxr) September 16, 2026

typesafe's jev is fun! live demo you can play with: https://typesafe-demo.val.run/
— Steve Krouse (@stevekrouse) September 16, 2026

We benchmarked fx auto mode (safety) classifier with @typesafeai's Jev. tl;dr: ~5-18x faster and more accurate than 𝚐𝚙𝚝-𝟻.𝟼-𝚕𝚞𝚗𝚊, our current top choice
— Pranit (@fazxes) September 16, 2026

alright - just got Jev deployed for a real production use case, which now performs at fable level quality but 10x faster and saves a ton of money context - a powerful capability of firstmate is that as an orchestrator it intelligently routes each task to an appropriate agent (permutation of harness, model, and reasoning effort) based on custom user preference by default, that's done by the firstmate agent and the LLM would have to do some thinking, make tool calls to read dispatch rules, quota data etc and then do the dispatch. this is slow and does cost a bit of LLM tokens i just replaced this dispatch process with Jev. it makes the same decision with no thinking or tool calls, done in ~200ms, and for the 25 tasks i evaluated this with, it gives the exact same answer fable would have given... there's still a tool call needed to invoke Jev, done by the firstmate agent which uses an LLM. but even with that counted, the saving from Jev still resulted in a -71% reduction in cost and -90% reduction in wall time of completing the whole dispatching process Jev API calls themselves are almost free.. i made 100+ calls, and my usage dashboard still shows $0.01 (which i think means "haven't reached $0.01 yet"). if we just look at the part Jev replaced and not the whole system, then the saving is on the magnitude of ~100x if you also have Jev and you use firstmate, set TYPESAFE_API_KEY in your .env file in your firstmate repo to activate this exciting times! i think this is starting to enable a whole new architectural paradigm for software. LLMs are just a small part of it. more on that soon
— Kun Chen (@kunchenguid) September 17, 2026

Tested @typesafeai's claim that their new model Jev delivered "comparable... intelligence" to GPT-5.6 Terra on "System 1" tasks. To do this, I compare both models on multiple-choice benchmarks (MMLU, GPQA, etc.). Set reasoning=none for Terra for sys 1. Result: Jev is Terra-tier.
— N8 Programs (@N8Programs) September 16, 2026

Awesome to see @EGafni & Diogo launch their first model Jev! Verification is the bottleneck for building great AI learning environments. Jev is a welcome, unique, addition to our toolkit @goodstartlabs Typed decisions, probabilities for each option, free output tokens! On our sample: • Jev agreed with Fable 5.1 nine times out of ten • ~200× lower grading cost • <0.5 seconds per grading call That means more eyes on what agents are doing and what we’re teaching them. More on how we’re using it to build game environments, experts to play them, and ensuring they teach the right things: https://goodstartlabs.com/research/verification-is-the-bottleneck
— alex duffy (@alxai_) September 15, 2026

Jev is an AI model built for fast, low-cost decisions that went viral yesterday. It plays doom on its own because of low cost/latency decision. This will be the SOTA for World of Warcraft bots.
— Chubby♨️ (@kimmonismus) September 16, 2026

Ran @typesafeai's Jev against an existing classifier eval that previously used Gemini 2.5 Flash Lite. It won both on quality (saturated the eval) and speed (6x)
— Malte Ubl (@cramforce) September 16, 2026

Jev from @typesafeai is on AI Gateway. Build agents that decide, route, score, and stop in milliseconds: 𝚊𝚠𝚊𝚒𝚝 𝚎𝚟𝚊𝚕𝚞𝚊𝚝𝚎({ 𝚖𝚘𝚍𝚎𝚕: '𝚝𝚢𝚙𝚎𝚜𝚊𝚏𝚎-𝚊𝚒/𝚓𝚎𝚟', 𝚜𝚝𝚊𝚝𝚎, 𝚚𝚞𝚎𝚜𝚝𝚒𝚘𝚗𝚜, }); https://vercel.com/changelog/typesafe-ai-jev-now-available-on-ai-gateway
— Vercel Developers (@vercel_dev) September 17, 2026

hype-free explanation of jev: jev does not replace gpt / claude jev is just a *really* smart switch statement like if 2016 ml classifiers got 2026 levels of intelligence it's a new* type of tool that will make a lot of workloads insanely fast, cheap, and accurate * = and by new, i mean rebranded ~~~ it needs a predefined set of options and it will tell you which one to take it cannot: - write code - generate natural language - reason step by step / show its work - produce any output you didn't define in advance - pick from more than ~255 options in one shot but it can: - classify, route, score, rank - give confidence - pick the right branch, tool, model, or sub-agent - judge / verify / guardrail an llm's output - label tons and tons of rows ~~~ i'd imagine a lot of workflows that look like: llm proposes options → jev decides → code executes and i see this fitting *really* well with code mode and mcp ~~~ implying this will lead to agi seems incredibly far fetched to me, but i don't want to discount the types of applications that this will make possible
— Nathan Flurry 🔩 (@NathanFlurry) September 16, 2026

Jev by @typesafeai is now on OpenRouter, in beta. Jev is a System One model. Instead of generating text, it takes your app's state plus a typed question and returns a typed decision with a probability attached. There is no JSON prompting, parsing layer, and nothing to validate against.
— OpenRouter (@OpenRouter) September 18, 2026

TypeSafe AI has introduced its first model named Jev - a small and customizable System One decision model. Not being an LLM, Jev can’t reason or write explanations; it generates a fast judgment with calibrated probabilities and a confidence score. Anything heavier gets split into separate questions and recombined in code. TypeSafe reports a response time of approximately 150ms per call, at around 100x lower cost and faster responses than LLMs, with pricing at $42 per billion input tokens and free output.
— 🚨 AI News | TestingCatalog (@testingcatalog) September 15, 2026

Built a model router with Jev by @typesafeai. Jev decides what model fits your request best and the request is sent to that model.
— Duncan (@ephraimduncan) September 17, 2026

First @typesafeai use case, live in our Mac app: setup and troubleshooting help when no model is loaded. Model downloading, load failed, API returning 503, phone won't pair: the user asks, Jev reads the question with the whole built-in manual as state and decides, with probabilities, what it is and which article answers it, or that nothing does. The app then shows the real documentation and live status. Jev decides, the app answers from its own docs. No model loaded, nothing invented. 42/42 on a held-out set: paraphrases, typos, French, German, Spanish, features that don't exist, follow-ups. Median 0.93 s. Great breakthrough by the TypeSafe team. Thank you.
— Malek Ould-Oulhadj (@malekoo) September 17, 2026

Testing apps with @typesafeai jev & @opencode will be soooo fast small demo:
— Filip (@Neriousy) September 16, 2026

Ask Jev anything. Give it a try at https://askjev.ai It won't answer. It will judge. Let's see if we can get to 1 million questions. @typesafeai 🤝 @convex work great together. @hmartenjoyer @CompleteSkeptic @justKDeng @mikeysee
— Wayne Sutton (@waynesutton) September 17, 2026

Tested TypeSafe’s Jev (no-text, probability-only model) as an AI agent safety monitor. Checking each action first worked well caught most attacks with almost no false blocks, and much faster than Gemini.
— 😎Nick 常胜 (@isNickMa) September 17, 2026

jev + kernel browser use demo pretty freaking cool! so fast! try it yourself: https://jev-browser-use.val.run/
— Steve Krouse (@stevekrouse) September 16, 2026

Jev's performance on knowledge benchmarks like MMLU and GPQA is highly impressive - especially for a non-COT model. It exceeds Terra at other linguistic reasoning tasks like WinoGrande or HellaSwag as well. It only loses substantially on math reasoning.
— N8 Programs (@N8Programs) September 16, 2026

Not to mention you also get probabilities from Jev, which are quite well-calibrated (expected calibration error is 1.74 percentage points averaged across benches - its 0.26pp at best and 6.96pp at worst)
— N8 Programs (@N8Programs) September 16, 2026

A model built to sit inside the if statement

Jev is deliberately narrower than an LLM. Its bet is that automation needs constrained judgment more often than it needs generated prose.

What TypeSafe launched, and where Jev fits

Jev is TypeSafe AI's first System One model. Developers send unstructured text state plus Choice, Score, or Noul questions whose output spaces are declared in advance. Jev returns typed answers, probability distributions, and confidence values rather than free-form text.

TypeSafe reports 193.6x faster and 444.6x cheaper results in a four-workflow evaluation, separate from its short recorded demo. Jev averages 67.8% agreement with Astra/Fable-generated reference answers there, while GPT-5.6 Sol scores 74.1%. The vendor calls the speed and cost ratios the high end of real-world gains, not universal guarantees.

Current documentation lists Jev 1.13 at a TypeSafe direct list price of $0.042 per million input tokens with free output. Context is 64K tokens across state and all questions, with a separate 32K limit for state plus the longest question. Input is text only and Choice supports up to 255 options. Typed output still can be wrong or manipulated by adversarial input; keep exact arithmetic and date comparisons in code.

Jev is accessible through SPIRITT AI Gateway. Open a SPIRITT Workspace and ask the agent to use Jev; the gateway handles model access while the workspace supplies the code, files, tools, tests, and surrounding workflow.

Typed decisionsParallel questions64K total / 32K per-question contextText onlyTypeSafe list: $0.042 / 1M inputAvailable in SPIRITT

Mean workflow agreement

GPT-5.6 Sol

74.1%

Claude Opus 5

73.1%

GPT-5.6 Terra

67.9%

Jev

67.8%

Claude Sonnet 5

67.8%

GPT-5.6 Luna

66.8%

DeepSeek V4 Pro

65.5%

DeepSeek V4 Flash

64.4%

Claude Haiku 4.5

53.6%

TypeSafe's four published workflows, equally weighted. Reference answers are GPT-6 Astra plus Fable 5.1 at high thinking, not verified human labels. Jev's reported mean is 67.8%, about $0.0004/case and 0.4 seconds/case (rounded).

Published Jev benchmarks, demos, and real-world evaluations

The full four-workflow vendor comparison, all nine N8 accuracy conditions, fx and classifier latency/accuracy, calibration, grading costs, and production routing. Each source is scoped separately; none is an overall leaderboard.

TypeSafe vendor demonstration

Recorded demo latency (ms)

Jev

114

GPT-5.6 Terra

8566

TypeSafe's short, deliberately simplified launch demo; Terra used default reasoning. About 75.1x faster here, distinct from the 193.6x four-workflow headline.

TypeSafe vendor demonstration

Cost per 1,000 demo runs

Jev

$0.08

GPT-5.6 Terra

$13.88

Scaled from $0.000081 and $0.013880 per recorded run, about 171.4x lower cost. These are direct-provider reference costs, not a SPIRITT billing quote.

TypeSafe vendor workflow evaluation

Security incidents agreement

Jev

61.7%

GPT-5.6 Terra

51.2%

GPT-5.6 Sol

62.5%

Claude Opus 5

66.2%

evals.typesafe.ai: all models run the same coded workflow at default reasoning. Reference labels are the average of Astra and Fable 5.1 at high thinking, not human-verified ground truth.

TypeSafe vendor workflow evaluation

Agent trace observability agreement

Jev

71.6%

GPT-5.6 Terra

73%

GPT-5.6 Sol

76.6%

Claude Opus 5

75.2%

evals.typesafe.ai: all models run the same coded workflow at default reasoning. Reference labels are the average of Astra and Fable 5.1 at high thinking, not human-verified ground truth.

TypeSafe vendor workflow evaluation

Invoice processing agreement

Jev

61.8%

GPT-5.6 Terra

74.7%

GPT-5.6 Sol

79.1%

Claude Opus 5

78.4%

evals.typesafe.ai: all models run the same coded workflow at default reasoning. Reference labels are the average of Astra and Fable 5.1 at high thinking, not human-verified ground truth.

TypeSafe vendor workflow evaluation

Customer service agreement

Jev

76%

GPT-5.6 Terra

72.7%

GPT-5.6 Sol

78.3%

Claude Opus 5

72.4%

evals.typesafe.ai: all models run the same coded workflow at default reasoning. Reference labels are the average of Astra and Fable 5.1 at high thinking, not human-verified ground truth.

TypeSafe vendor headline

Workflow speed headline (relative)

Jev

193.6

Compared workflow

The vendor attributes 193.6x to its separate four-workflow evaluation and calls these gains the higher end of real-world results. Not the 114 ms demo or a universal guarantee.

TypeSafe vendor headline

Workflow cost efficiency (relative)

Jev

444.6

Compared workflow

TypeSafe's separate 444.6x workflow-cost claim. Results depend on workflow and comparator; the short demo instead implies about 171.4x.

TypeSafe published reference price

Input list price per billion tokens

Jev

$42.00

Launch LLM floor

$200.00

TypeSafe lists $0.042 per million input tokens and free output. The comparator is only the $0.20/M lower bound in TypeSafe's generic launch table, not an equal-quality benchmark or SPIRITT price.

Independent: N8 Programs

MMLU

Jev

90.91%

Terra, no reasoning

87.89%

N8 Programs, September 16 numeric result table. Terra reasoning=none and letter-only output; explicit a/b/c/d instruction for chess. This is an external experiment, not an official TypeSafe benchmark or a full-reasoning comparison.

Independent: N8 Programs

GPQA Diamond

Jev

68.18%

Terra, no reasoning

56.06%

Independent: N8 Programs

ARC-Easy

Jev

99.33%

Terra, no reasoning

98.82%

Independent: N8 Programs

ARC-Challenge

Jev

97.61%

Terra, no reasoning

96.76%

Independent: N8 Programs

WinoGrande

Jev

91.63%

Terra, no reasoning

78.69%

Independent: N8 Programs

HellaSwag

Jev

94.86%

Terra, no reasoning

95.32%

Independent: N8 Programs

GSM8K: four choices

Jev

79.68%

Terra, no reasoning

87.72%

Independent: N8 Programs

GSM8K: ten choices

Jev

54.59%

Terra, no reasoning

69.83%

Independent: N8 Programs

Chess: four legal moves

Jev

49.45%

Terra, no reasoning

50.25%

Independent: N8 Programs

Expected calibration error (pp)

Lowest benchmark

0.3

Reported average

1.7

Highest benchmark

7.0

N8 reports 1.74 percentage-point expected calibration error across the benchmark suite, with a 0.26 to 6.96 range. The reliability chart covers 33,735 predictions and uses ten equal-width bins. Lower is better; individual answers may still be wrong.

Independent: Vercel fx

fx safety accuracy

Jev

98.6%

GPT-5.6 Luna

96.7%

Vercel fx safety evaluation: 70 labeled cases, three runs each (210 decisions per classifier, not 210 independent cases). Original Pranit chart, September 16. Correct decisions: 207/210 versus 203/210. Jev had zero fallbacks and identical answers on all three runs of each case. Not a general safety guarantee.

Independent: Vercel fx

fx median latency (ms)

Jev

312

GPT-5.6 Luna

1458

Vercel fx safety evaluation: 70 labeled cases, three runs each (210 decisions per classifier, not 210 independent cases). Original Pranit chart, September 16. About 4.7x faster at the median.

Independent: Vercel fx

fx p95 latency (ms)

Jev

374

GPT-5.6 Luna

6583

Vercel fx safety evaluation: 70 labeled cases, three runs each (210 decisions per classifier, not 210 independent cases). Original Pranit chart, September 16. About 17.6x faster at p95, not a 5 to 18x p95 range.

Independent: Malte Ubl

Classifier accuracy

Jev

100%

Gemini 2.5 Flash Lite

98.2%

Malte Ubl classifier evaluation: 167 prompts run three times (501 decisions each). Original screenshot says not deployed yet. 501/501 versus 492/501 correct; saturation of this evaluation does not prove universal accuracy.

Independent: Malte Ubl

Classifier median latency (ms)

Jev

111

Gemini 2.5 Flash Lite

703

Malte Ubl classifier evaluation: 167 prompts run three times (501 decisions each). Original screenshot says not deployed yet. About 6.3x faster at the median.

Independent: Malte Ubl

Classifier p95 latency (ms)

Jev

217

Gemini 2.5 Flash Lite

1074

Malte Ubl classifier evaluation: 167 prompts run three times (501 decisions each). Original screenshot says not deployed yet. About 4.9x faster at p95.

Independent: Good Start Labs

Estimated grading cost / million answers

Jev

$160.00

Fable 5.1

$33000.00

Good Start's published study: 6,003 rubric checks across 1,203 financial-research answers, 91.5% agreement with Fable 5.1. Estimated costs use direct-provider rates. Agreement is not accuracy.

Builder report: 25 evaluated tasks

Firstmate routing cost, indexed

With Jev

Previous workflow

100

Kun Chen reports 71% lower end-to-end routing cost. The same 25 decisions matched Fable, which is agreement, not independently proven correctness.

Builder report: 25 evaluated tasks

Firstmate routing time, indexed

With Jev

Previous workflow

100

Kun Chen reports 90% lower end-to-end wall time and roughly 200 ms for the Jev decision. Indexed to 100 for the prior system; includes the surrounding agent invocation.

Independent: Good Start Labs

Jev agreement with other graders

With Fable 5.1

91.5%

With Astra

90.6%

With Gemini 3.8

90.6%

With DeepSeek V4.1

91.5%

With GPT-5.6 Luna

86.5%

6,003 rubric checks across 1,203 financial-research answers. Share of matching verdicts, not accuracy. Separate initial study: 96.0% agreement with Sonnet 4.5 on 741 checks/112 answers. A separate 10,500-call game-task run had zero unusable gradings, not zero wrong verdicts.

Official sources: TypeSafe AI's launch, homepage, workflow evaluation notes, System One documentation, and current model reference. Independent signals include the Vercel fx safety evaluation, Malte Ubl's classifier evaluation, Firstmate's 25-task production-routing report, Good Start Labs' grading sample, and N8 Programs' MMLU/GPQA comparison. These use different harnesses and are not one comparable leaderboard. Reproduce calibration, latency, and cost on your own decisions before autonomous production use.

Jev gives software a decision, not another paragraph

Jev builders and evaluations across X

A model built to sit inside the if statement

What TypeSafe launched, and where Jev fits

Mean workflow agreement

Published Jev benchmarks, demos, and real-world evaluations

Recorded demo latency (ms)

Cost per 1,000 demo runs

Security incidents agreement

Agent trace observability agreement

Invoice processing agreement

Customer service agreement

Workflow speed headline (relative)

Workflow cost efficiency (relative)

Input list price per billion tokens

MMLU

GPQA Diamond

ARC-Easy

ARC-Challenge

WinoGrande

HellaSwag

GSM8K: four choices

GSM8K: ten choices

Chess: four legal moves

Expected calibration error (pp)

fx safety accuracy

fx median latency (ms)

fx p95 latency (ms)

Classifier accuracy

Classifier median latency (ms)

Classifier p95 latency (ms)

Estimated grading cost / million answers

Firstmate routing cost, indexed

Firstmate routing time, indexed

Jev agreement with other graders

How It Works

Open a workspace

Just Ask to use Jev

Build or automate

Build the workflow around the decision.

Jev FAQ