Jev, the decision model from TypeSafe AI, is now the fastest-adopted model in the history of Vercel's AI Gateway, according to Vercel's own data. The company's October 2026 production index says Jev reached "35% of paid teams in its first 14 days" (35.3% on its scorecard), nearly double the reach of Sonnet 5.5 in second place. TechCrunch picked up the figure on October 16.
The same index shows Jev took roughly a quarter of one percent of gateway spend. Both facts are true, and the gap between them is the real story: reach, retention and spend measure different things. This post unpacks them, compares Jev with the other September launches, and says what builders should take from it.
TL;DR: the Jev numbers on Vercel
| Question | Answer (source: Vercel October 2026 index) |
|---|---|
| Launch on the gateway | September 17, 2026 |
| Reach in first 14 days | 35% of paid teams (35.3% scorecard) |
| Second place | Claude Sonnet 5.5 at 18.1% |
| Spend share | 0.27% in first two weeks (0.15% full month, 0.3% scorecard) |
| Week-two return rate | 55.6% of teams that tried it in week one |
| Method | Paid teams active from start of launch month; spend at published list prices |
| Does Vercel call it a record? | Yes, fastest adoption of any model in AI Gateway history |
What does 35% of paid teams actually mean?
Vercel's methodology note matters. Reach counts paid teams active from the start of the launch month through the data cutoff, and the figure is the share of those teams that sent at least one request to the model. It is a "tried it" measure. It is not a count of requests, tokens or dollars.
Observability dashboard with an eye, a way to picture reach, retention and spend as separate Jev adoption metrics
That is why a model can top the reach chart and still be a rounding error in spend. A team that routes ten thousand classification calls through Jev in an afternoon counts the same as a team that streams a billion tokens through a flagship model. Reach tells you how many builders were curious or found a fit. Spend tells you how much money moved.
How does Jev compare with other September launches?
Vercel lists reach in the first two weeks for each launch. The top of the table:
| Model | Share of paid teams, first two weeks |
|---|---|
| Jev (TypeSafe AI) | 35.3% |
| Claude Sonnet 5.5 | 18.1% |
| GPT-6 Luna | 14.9% |
| Gemini 3.8 Flash | 13.2% |
| Claude Opus 5.5 | 13.1% |
| GLM-5.3-Flash | 9.5% |
| DeepSeek-V4.1-Flash | 8.4% |
| GPT-6 Sol | 6.9% |
| GPT-6 Astra | 6.6% |
| Claude Fable 5.1 | 5.8% |
| Grok 4.7 | 3.6% |
The pattern is clear. Cheap, fast models earn wide reach because trying one costs almost nothing. The expensive flagships, GPT-6 Astra and Fable 5.1, sit at the bottom of the list even though they are the most capable. If you want to see how those flagships compare on price and performance, read our GPT-6 Astra launch breakdown and the Claude Opus 5.5 launch post.
Why is Jev's spend share so small?
Jev does not generate text. As we explained in our launch coverage, it returns typed decisions, such as a choice, a score or a yes or no with a probability, and TypeSafe AI priced it at $0.042 per million input tokens with free output tokens. A fact-check of its speed and cost claims found those numbers came from the vendor.
At that price, even 35% of teams produce little spend. TechCrunch reports that Vercel itself replaced GPT-6 Luna with Jev for classifying harmful commands. The article says the main stated use is classifying large quantities of data. That is exactly the high-volume, low-value-per-call workload where a decision model beats a chat model on cost.
The 0.27% figure should not be read as failure. It is what a cheap utility looks like on a spend chart. Compare it to a routing layer or a spam filter: it matters to the system and barely registers on the bill.
Is the 55.6% retention a good number?
Vercel says 55.6% of teams that tried Jev in week one returned in week two. Read that carefully.
- It compares two adjacent weeks, so it says nothing about month-three behavior.
- The denominator is teams that tried the model, many of whom may have run a single evaluation.
- Vercel does not publish a retention figure for the other models in the text we read, so we cannot say whether 55.6% is high or low against Sonnet 5.5 or Luna.
A fair reading: more than half of early testers came back within a week, which suggests real use rather than pure curiosity. It does not prove production dependence.
What are the caveats on this data?
- It is one gateway. Vercel's AI Gateway is a unified API, similar in concept to other routers. TechCrunch notes the view may undercount labs like OpenAI and Anthropic, which sell through their own APIs. A team that uses Claude directly never shows up here.
- Spend is estimated at list prices. Discounts, committed-use deals and cache pricing are not reflected.
- The population skews. Teams that route through a gateway tend to be the ones that already switch models, so they adopt new releases faster than the market at large.
- Vercel is a participant. The company reports its own platform data and was itself an early Jev adopter, per TechCrunch. The figures are first-party, not audited.
None of this makes the number wrong. It means "fastest adopted on Vercel's gateway" is the accurate phrasing, not "fastest adopted anywhere."
What did the rest of the October index say?
The same report carries two headline findings that frame Jev's rise. Open-weight models now process more than seven in ten gateway tokens, 71.5% in September, and account for nearly one in four dollars, about 23%. Anthropic's spend share fell from 63.7% in August to 43.9% in September, dropping below half for the first time.
Those shifts line up with the small-and-cheap trend Jev rides. When teams route simple, high-volume work to cheap models, the flagship share of tokens falls even if flagship quality improves. For wider background on the model class, see our guide to what decision models are and the ranked list of decision models.
How does this fit Vercel's earlier gateway reports?
Vercel has published a production index for several months, and the trend line is consistent. The September index, which covered August traffic, reported open-weight models at 56% of token volume and 14% of estimated spend, with Anthropic holding 64% of all spend. The October edition, covering September, moves the same two measures to 71.5% of tokens and about 23% of spend, while Anthropic's share falls to 43.9%.
Read together, the reports describe a market where token volume migrates to cheaper models faster than dollars do. Jev is an extreme case of that: a model with a very high reach and a very small spend share. It also shows why a single leaderboard position, whether by reach or by spend, is a poor guide to what to build on. Reach rewards low price and low switching cost. Spend rewards capability and volume. Retention, if Vercel keeps publishing it, may turn out to be the more honest measure of whether teams found real use.
A practical habit follows: when you read any adoption chart, ask which of the three measures it shows, and what the denominator is. A figure like "35% of paid teams" invites a headline, while "0.27% of spend" invites a shrug, and both describe the same model in the same two weeks.
How should builders use this?
The data suggests a practical pattern, and it is one you can test this week.
A rail switch sending one track to one of three destinations, an analogy for routing classification calls to Jev
- List your repeated, structured calls. Moderation, intent routing, labeling, scoring, and safety checks are all candidates.
- Measure accuracy before price. TypeSafe's own disclosed numbers showed an accuracy gap against the best large model on its workflow set, covered in our Jev versus XGBoost and BERT comparison. Your task may differ.
- Keep a fallback. Route low-confidence decisions to a larger model. Decision models return probabilities, which makes thresholding easy.
- Watch competing offerings. TechCrunch notes OpenAI and Amazon have released competing products. Our developer verdict on OpenAI's decisions API versus Jev is a starting point.
- Track cost per correct decision, not price per token.
What to watch next
- Whether Jev's reach holds in Vercel's November index, and whether retention rises past the first-week figure.
- Whether spend share grows as teams move more volume, or stays a rounding error.
- Whether the competing decision models from large vendors close the reach gap.
- Whether TypeSafe AI publishes updated accuracy figures, as access has widened since the general availability announcement.
What this means for what you build or pay
If you run a high-volume classification step on a flagship model, a decision model could cut that line item sharply, but test accuracy on your data first. If you do not run such a step, the adoption number changes nothing for you. The bigger lesson is structural: gateway data shows teams are splitting work between cheap specialists and expensive generalists, so your architecture should expect to route.
Figures are from Vercel's October 2026 index and TechCrunch as of October 16, 2026, and may be revised.