Jev Is 5 to 200 Times Faster, Depending on the Denominator

TypeSafe AI came out of stealth on September 15 with Jev, a model that writes no text. It takes some state and a list of typed questions (which of these options, does this condition hold, where on this scale) and returns a probability for each, in one parallel pass. The company’s launch post says Jev reaches “similar levels of intelligence on System One tasks compared to existing LLMs” while being “two orders of magnitude faster and more efficient,” and its home page puts numbers on that: 193.6 times faster, 444.6 times cheaper.

The launch post is unusually careful about those numbers. It says they come from one eval, and that “we expect that these are on the higher end of real world gains.” That sentence is the best guide to the claim. Every “N times faster” has a second model in it, and how large N is depends on which one.

Where the 193.6 comes from

The evidence behind the headline is TypeSafe’s own workflow evals: four workflows (security incidents, agent-trace observability, invoice processing, customer service), each broken into narrow typed questions with ordinary code making the final call. Accuracy is agreement with a consensus, the average of GPT-6 Astra and Claude Fable 5.1 at high thinking answering every question. The other models run at their providers’ default reasoning settings, behind a wrapper TypeSafe wrote to force structured answers, and the launch post says the workflows were built by its own capabilities team, so “some bias could exist.” It also says the consensus “biases answers towards OpenAI and Anthropic’s models,” which it says likely understates Jev and DeepSeek’s models.

TypeSafe's workflow eval: accuracy against cost per case for Jev and eight comparison models A scatter plot on a logarithmic cost axis from 0.0001 to 1 dollar per case, with accuracy against the consensus labels on the vertical axis. Positions are read off TypeSafe's own chart and are approximate, about half a point of accuracy and five percent of cost. Jev sits at about 68 percent accuracy and about 0.0004 dollars per case. Models of about the same accuracy cost more: OpenAI's luna at about 67 percent and about 8 times Jev's cost, terra at about 68 percent and about 75 times, and Anthropic's Sonnet 5 at about 68 percent and about 300 times. More accurate models cost far more: OpenAI's sol at about 74 percent and about 210 times, and Anthropic's Opus 5 at about 73 percent and about 440 times. DeepSeek V4 flash is at about 64 percent and 15 times, DeepSeek V4 pro at about 65 percent and 100 times, and Haiku 4.5 at about 54 percent and 50 times. TYPESAFE'S WORKFLOW EVAL · ACCURACY VS. COST PER CASE · WORKFLOW CONFIGURATIONS 50% 60% 70% $0.0001 $0.001 $0.01 $0.10 $1 Jev's accuracy OpenAI luna: about 67% at about $0.0033 per case DeepSeek V4 flash: about 64% at about $0.006 per case Anthropic Haiku 4.5: about 54% at about $0.02 per case OpenAI terra: about 68% at about $0.03 per case DeepSeek V4 pro: about 65% at about $0.04 per case OpenAI sol: about 74% at about $0.08 per case Anthropic Sonnet 5: about 68% at about $0.12 per case Anthropic Opus 5: about 73% at about $0.18 per case Jev: about 68% at about $0.0004 per case Jev luna ~8x DS v4 flash ~15x haiku 4.5 ~50x terra ~75x DS v4 pro ~100x sol ~210x sonnet 5 ~300x opus 5 ~440x Positions and multiples read off TypeSafe's chart by the writer, roughly 1 point of accuracy and 5% of cost. Accuracy is agreement with the average of GPT-6 Astra and Claude Fable 5.1. "~Nx" is a model's cost per case relative to Jev's.
At Jev's accuracy the price gap runs from about 8 times (luna) to about 300 times (Sonnet 5). By my reading, the 444.6 times matches the ratio against Opus 5, which scores about five points higher.

Read Jev’s position on that chart and the multiple stops being a single number. It lands at about 68 percent accuracy and roughly four hundredths of a cent a case. The cheapest model that scores about the same is OpenAI’s luna, at around eight times the cost. Terra is around 75 times, and Sonnet 5, which also scores about 68 percent, around 300 times. The two models that beat Jev by five or six points, sol and Opus 5, cost around 210 and 440 times as much. The page doesn’t say which models the headlines divide by, but the marker positions do. By my measurement, the 444.6 is the cost ratio against Opus 5, the priciest model on the chart, and the 193.6 is the response-time ratio against Sonnet 5 (roughly 78 seconds a case against Jev’s 0.4) on the companion latency view, where DeepSeek V4 pro is slower still. So the two headlines divide by different rivals. The launch post is explicit about the effect of the comparison in its Wikipedia-racing demo, where “our speedups here tend to be a lot less than in previous demos” because the rivals ran in their non-reasoning modes.

What outsiders measured

Within three days, several developers ran their own comparisons, and Every published one on launch day. Near Here, a UK events site, tested Jev on September 16 against Mistral Small 4 and Gemini 3.5 Flash-Lite, both at high reasoning, on rejecting unsuitable listings. On its 21-listing run, its report puts Jev at 0.59 seconds and $0.043 per 1,000 decisions, against 2.90 seconds and $0.370 for Mistral and 3.40 seconds and $2.496 for Gemini. That works out to 4.9 and 5.8 times faster, and 8.6 and 58 times cheaper, by my division. Emil Lindfors ran 24 Norwegian consultation letters through Jev and through DeepSeek V4.1 Flash. With DeepSeek’s reasoning off it took 2.7 seconds and $1.31 per 1,000 documents against Jev’s 0.32 seconds and $0.22. With reasoning on, DeepSeek needed 26 seconds and $3.08. Every’s write-up ran four writing checks across 12 synthetic passages, six clean and six with planted problems. Jev’s median was 0.35 seconds a passage against 8.83 for Claude Fable 5.1 at high effort, about 25 times faster, at an estimated cost about 580 times lower. Jev caught six of the seven planted defects and Fable all seven. TechCrunch relays two customer tests without data: Vercel found Jev 5 to 18 times faster and more accurate than OpenAI’s luna on safety classification, and Bryo AI found Gemini slightly more accurate and 10 to 20 times costlier.

How many times faster and cheaper Jev is, by which rival and which source A dot plot on a logarithmic axis from 1 to 1,000 times. For each rival, a circle marks how many times faster Jev is and a diamond marks how many times cheaper. Mistral Small 4 at high reasoning, from Near Here: 4.9 times faster, 8.6 times cheaper. Gemini 3.5 Flash-Lite at high reasoning, from Near Here: 5.8 faster, 58 cheaper. DeepSeek V4.1 Flash with reasoning off, from Lindfors: 8.4 faster, 6.0 cheaper. The same model with reasoning on: 81 faster, 14 cheaper. Claude Fable 5.1 on four writing checks over 12 synthetic passages, from Every: 25 faster, 580 cheaper. TypeSafe's own headlines: 193.6 faster and 444.6 cheaper, which by the writer's reading of the chart are the ratios against Sonnet 5 on time and Opus 5 on cost. Ratios for the first four rows are the writer's division of the reported figures. TIMES FASTER (CIRCLE) AND TIMES CHEAPER (DIAMOND) THAN EACH RIVAL · LOG SCALE 1x 10x 100x 1,000x Mistral Small 4 Gemini 3.5 Flash-Lite DeepSeek V4.1 Flash, reasoning off DeepSeek V4.1 Flash, reasoning on Claude Fable 5.1, four writing checks TypeSafe headline: Sonnet 5 time, Opus 5 cost Near Here, 16 Sep Near Here, 16 Sep Lindfors, 18 Sep Lindfors, 18 Sep Every, 15 Sep typesafe.ai home page Mistral Small 4: 4.9 times faster (2.90 s vs 0.59 s) Mistral Small 4: 8.6 times cheaper ($0.370 vs $0.043 per 1,000 decisions) Gemini 3.5 Flash-Lite: 5.8 times faster (3.40 s vs 0.59 s) Gemini 3.5 Flash-Lite: 58 times cheaper ($2.496 vs $0.043 per 1,000 decisions) DeepSeek V4.1 Flash, reasoning off: 8.4 times faster (2.7 s vs 0.32 s) DeepSeek V4.1 Flash, reasoning off: 6.0 times cheaper ($1.31 vs $0.22 per 1,000 documents) DeepSeek V4.1 Flash, reasoning on: 81 times faster (26 s vs 0.32 s) DeepSeek V4.1 Flash, reasoning on: 14 times cheaper ($3.08 vs $0.22 per 1,000 documents) Claude Fable 5.1: about 25 times faster (medians, 8.83 s vs 0.35 s), from Every Claude Fable 5.1: about 580 times cheaper (estimated), from Every TypeSafe headline: 193.6 times faster, matching Sonnet 5 on time by the writer's reading TypeSafe headline: 444.6 times cheaper, matching Opus 5 on cost by the writer's reading 4.9x, 8.6x 5.8x, 58x 8.4x, 6.0x 81x, 14x 25x, 580x 193.6x, 444.6x Each ratio is the rival's response time or price divided by Jev's. Right column: times faster, times cheaper. The first four rows are the writer's division of each source's printed figures. Fable row: Every's medians and estimate. Near Here ran both chat models at high reasoning; costs are list prices. Last row: rivals read off chart geometry.
The same model, DeepSeek V4.1 Flash, is 8 times slower than Jev with reasoning off and 81 times slower with it on, on a mix of providers by Lindfors' account. The multiple moves at least as much with that setting as with the rival's name.

Line the outside numbers up and the pattern is plain. Against small, cheap models, Jev was 5 to 8 times faster. Against the same DeepSeek model with reasoning switched on it was 81 times faster, and against a flagship it was 25 times in Every’s test and, by my reading, nearly 200 in TypeSafe’s chart. They divide by different things. Between the third and fourth rows the model stayed the same and only the reasoning setting changed, though Lindfors notes that OpenRouter spread his DeepSeek runs over 13 providers, so the latencies are a mix. Cost tells the same story from the other direction: 6 to 14 times cheaper than a small model, 58 times against Gemini, and 444.6 or 580 times against frontier flagships. TypeSafe’s headlines divide by the priciest model and the second slowest; the outside tests mostly compared against models a developer might use for a classification step today, and got smaller numbers.

Accuracy means agreement

The published tests I read grade Jev against another model, or against a key an AI assistant wrote. TypeSafe’s reference is the average of two flagships. Lindfors’ labels come from Claude Fable 5.1, run twice, and he says so directly: “Where the reference is wrong and Jev is right, Jev scores as wrong.” On four stance documents where Jev and the key disagreed, he writes that “Jev was wrong four times” and “Jev disagreed with an LLM four times” are the same number until he settles them. Near Here’s expected decisions “were written by the assistant and fixed before calls” and “have not been independently human-adjudicated,” and the 50 cases in its headline table also informed prompt selection. Every planted its defects in synthetic passages, so its answer key is known by construction, but across 12 passages. The Vercel and Bryo tests TechCrunch relays don’t say how they were scored. I found no published test scored against human-adjudicated ground truth.

Within those limits Jev held its own on narrow tasks. It matched or nearly matched the small models on Near Here’s extra 21 listings (19 of 21, against 19 and 20), tied DeepSeek with reasoning off on Lindfors’ stance question (20 of 24, against 22 with reasoning on), and beat it on the ordered scale (19 of 24 against 14). The samples are small. Lindfors puts the 95 percent interval on a stance number at about 15 points either way with 24 documents, the same kind of interval I went looking for in Google’s voice-agent chart. Reasoning, in his run, bought DeepSeek two stance labels and one respondent type for about ten times the latency.

What “can’t hallucinate” covers

The launch post says Jev “can’t hallucinate,” and its type-error plots put a zero for Jev. The post’s own note under that chart says “our number is not empirical. Schema matching is guaranteed, thus we can confidently add 0% into the plots.” The claim is about shape. Every answer is one of the options defined in advance, and each carries a probability. On Hacker News a common objection was that a valid answer can still be a wrong one, and TypeSafe’s CEO, posting as CompleteSkeptic (“CEO here”), conceded “it’s also possible to be confidently wrong” while defending the term by comparing Jev to a random forest. The company’s FAQ is plainer. Asked whether Jev can still get things wrong, it answers yes: Jev guarantees “the shape of its answers, not that every decision is correct,” and it can choose the wrong one. TechCrunch, reporting the launch, wrote that because users define the outputs in advance, Jev “cannot hallucinate,” without the qualification the founder gave in that thread. Armin Ronacher, quoted in the same piece, put it more exactly: “it delegates the hallucination problem a little bit to the user,” who has to decide what to do with a 50 percent answer.

Whether the probabilities mean anything

That leaves the one property that would make Jev’s answers usable: whether a stated 0.8 comes true about 80 percent of the time. TypeSafe says its training method, RLCD, targets exactly that, and it has declined to publish public benchmarks by design (“Put no weight on public benchmarks”). Lindfors tested it on 192 yes-or-no judgments across his 24 documents.

Jev's stated probability against how often the reference said yes A calibration plot with a dashed diagonal for perfect calibration. In five bins of Jev's stated probability of yes, the share of judgments where the reference model said yes: 0 to 0.1, 0 percent of 14 judgments; 0.1 to 0.3, 4 percent of 56; 0.3 to 0.7, 34 percent of 41; 0.7 to 0.9, 97 percent of 38; 0.9 to 1.0, 98 percent of 43. The low bins sit below the diagonal and the high bins above it, so Jev is underconfident at both ends. Lindfors reports a calibration error of 0.040 with short draft questions and 0.116 with longer, more careful ones; the bins plotted are from the careful run. Data from 24 Norwegian consultation letters, jev-1.13.0, reference labels from Claude Fable 5.1. JEV'S STATED PROBABILITY VS. HOW OFTEN THE REFERENCE SAID YES · 192 JUDGMENTS · LINDFORS Stated 0 to 0.1: reference said yes 0% (14 judgments) Stated 0.1 to 0.3: reference said yes 4% (56 judgments) Stated 0.3 to 0.7: reference said yes 34% (41 judgments) Stated 0.7 to 0.9: reference said yes 97% (38 judgments) Stated 0.9 to 1.0: reference said yes 98% (43 judgments) n=14 n=56 n=41 n=38 n=43 0% 50% 100% 0 0.5 1.0 Jev's stated probability of yes Jev says 0.1 to 0.3 (n=56): reference says yes 4% of the time. Jev says 0.7 to 0.9 (n=38): reference says yes 97%. Lindfors: "a bit underconfident at both ends." ECE, short draft questions: 0.040 ECE, longer careful ones: 0.116 Bins: careful-wording run. Reference: Claude Fable 5.1. jev-1.13.0, 24 documents; a first look, not a benchmark.
The direction is right in every bin, in the careful-wording run. The dashed line is perfect calibration, and Jev sits below it at the low end and above it at the high end, which is underconfidence.

In the careful-wording run (calibration error 0.116; the draft run’s bins are not published), when Jev said 0.7 to 0.9 the reference said yes 97 percent of the time, and when it said 0.1 to 0.3, 4 percent. The direction is right everywhere, and in Lindfors’ words it is “a bit underconfident at both ends.” The second finding is the one I would carry into a deployment. He rewrote his questions to be more careful, with qualifiers about passing mentions and quotations, and the calibration error rose from 0.040 to 0.116, which he attributes to the longer instruction pushing probabilities toward the middle. TypeSafe’s documentation warns that the model reads instructions at face value, and he had read the warning first.

What is still not public

In anything I read, TypeSafe has not published parameter counts or architecture details. Lindfors notes it “does not publish what is inside Jev,” the CEO told Hacker News that “architecture is close to the chest for now, but we have talked about writing a paper,” and TechCrunch reports that the model is transformer-based and, per Almeida, trained exclusively on synthetic data, and adds that outside observers suspect it is built on an open-weight LLM, without citing evidence. The FAQ answers “Is Jev just a smaller LLM?” with “Jev is neither small nor an LLM, hence being off the intelligence Pareto curve.” Access is early access, though several developers outside the company have run it. The company’s own advice is to build your own eval, because “System One tasks are much easier to evaluate.” Near Here did that on the day after launch, with 50 cases and 21 listings. That is the scale at which the claim can be checked today. Near Here’s 5 to 6 times faster are the two smallest speed multiples in the second figure.


References

  1. Almeida, D. Introducing System One Models & Jev. TypeSafe AI blog, September 15, 2026.
  2. TypeSafe AI. Workflow evals. Read September 20, 2026. Chart positions in the first figure are the writer’s readings of this page.
  3. TypeSafe AI. Home page. Headline multiples and FAQ answers quoted above. Read September 20, 2026.
  4. Reed, J. Testing TypeSafe Jev, Mistral and Gemini for local event validation. Near Here, September 16, 2026.
  5. Lindfors, E. An early-access test of TypeSafe’s Jev: calibrated judgments for half a cent. lindfors.no, September 18, 2026.
  6. Taylor, M. Mini-Vibe Check: TypeSafe’s Jev Judged Everything I’ve Written in 0.7 Seconds. Every, September 15, 2026.
  7. Fernholz, T. A new kind of AI model from a ChatGPT inventor is thrilling developers. TechCrunch, September 18, 2026.
  8. Hacker News. Introducing System One Models and Jev. September 15, 2026. Comments by jacobgold and CompleteSkeptic.