Published Updated
In one or two sentences, before anything else.
TypeSafe AI is a San Francisco AI lab whose first model, Jev, returns typed decisions with probabilities instead of text, for software to act on. Jev opened in early access on September 15, 2026 with a $40 million seed round, and Bloomberg reported on October 9 a raise of about $870 million at a $7.5 billion valuation.
What it can do, one sentence each, and whether we actually tested it.
Send a state (text or JSON) and questions of four types: Choice picks from a list, Score rates against levels, Noul gives a yes probability and Box finds a region in an image.
Box arrived with jev-1.14.0 on October 14, 2026.
Choice and Score answers carry a probability for each option and a confidence value, so code can act on clear cases and send the rest to a person.
TypeSafe calls these calibrated; nobody has measured that independently on production data.
All questions in a request are answered in parallel against the same state; TypeSafe's GDPR cookbook reports 13 questions in one call at 12.2 times lower cost and 10 times the speed of 13 calls.
TypeSafe's own measurement.
$0.042 per million input tokens ($42 per billion), with output tokens free.
HTTP API, Python and JavaScript SDKs, a browser Playground, an agent skill for Claude Code and Codex, and a LangChain integration package.
Images in the state, instructions or criteria, and a 262k-token context, switched on with request headers TypeSafe labels danger zone.
TypeSafe says these are not ready for production use.
No training on prompts or other input, zero data retention for enterprise customers, and a SOC 2 Type II report on request.
From TypeSafe's privacy policy and enterprise page.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
13 routing scenarios modelled on his team's Slack requests, 8 meant for the weak model and 5 for the strong one, run three times end to end, plus 130 direct API calls.
What happened
Jev routed 39 of 39 correctly against 36 of 39 for the local LLM judge, cut judgment time from about 1.55 seconds to 0.27 to 0.42 seconds, and cost about $0.003 for all 130 calls. The scenarios were short single-turn requests, and request text now leaves the local machine.
Case 2
Input
Score about 6,000 candidates per open role and keep a shortlist of 2,000, tested on 150 live roles against Gemini 3.1 Flash Lite.
What happened
TypeSafe reports similar recall (94.6% against 93.9%, not statistically significant), median time down from 20.3 to 10.3 seconds and cost per 1,000 candidates down from $0.755 to $0.092. TypeSafe published it, and the $500,000 saving is a projection.
Case 3
Input
Ask 13 questions about the Wikipedia article on the GDPR, such as whether it is a US federal law and how heavy its compliance burden is, in a single call.
What happened
TypeSafe says one batched call gave the same answers as 13 separate calls at 12.2 times lower cost and 10 times the speed, because the long article is sent once. This is TypeSafe's own example, not an outside test.
What it does badly, with the evidence, and what we could not settle.
Weak
Not a text model.
Evidence: TypeSafe's docs say Jev is not trained to generate text and it returns no explanation, so it cannot chat, write code or justify a decision; DataCamp notes the missing rationale complicates audits. Anyone who needs written answers or a reason for each decision should not use it now.
Known failure modes on numbers, dates and wording.
Evidence: TypeSafe's jev-1.13 page, reviewed October 2, 2026, says it does not count reliably, misjudges date comparisons, reads instructions literally, leans toward the first Choice option and can be steered by adversarial text in the state.
Best in English.
Evidence: TypeSafe's models page says English is where accuracy is best, and that other languages, including Chinese, Japanese and Korean, are handled less well.
Hosted only, with a short default context.
Evidence: Jev runs only on TypeSafe's API, so data leaves your systems, which Classmethod's tester flagged. The default limit is 64k tokens per request and 32k per question; the 262k option is slower and less accurate by TypeSafe's own account.
Uncertain
The speed, cost and accuracy claims.
Why unsure: The 193.6 times faster and 444.6 times cheaper figures come from four workflows TypeSafe's team wrote, scored against GPT-6 Astra and Claude Fable 5.1 answers; SiliconANGLE, VKTR and DataCamp all say no one has reproduced them.
What zero hallucinations means.
Why unsure: TypeSafe guarantees every answer matches the schema you define; VKTR and SiliconANGLE point out that a well-formed answer can still be wrong.
How many people use it.
Why unsure: TypeSafe told Bloomberg that about a third of Fortune 500 companies use Jev and that it passed a million users, but it named no customers.
The later funding round and the price.
Why unsure: Bloomberg (October 9) and Pulse 2.0 (October 11) report about $870 million at $7.5 billion, but TypeSafe had not announced it on its blog by October 17. TypeSafe also says it cannot yet prove its prices are not subsidized.
Who can get access today.
Why unsure: TypeSafe still calls Jev early access and September testers went through a waitlist; we did not sign up, so we cannot say whether the waitlist still applies.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Maker | TypeSafe AI, Inc., San Francisco; founded in 2024 | Founding year per SiliconANGLE. Source: typesafe.aichecked |
| Founders | Diogo Almeida (CEO, formerly at OpenAI), Erik Gafni (CTO), Sasha Sheng (COO) | Source: vktr.comchecked |
| Launch | Jev opened in early access on September 15, 2026 | Source: typesafe.aichecked |
| Current models | jev-1.14.0 (from October 14, 2026, alias jev-preview) and jev-1.13.0 (alias jev-latest) | Source: docs.typesafe.aichecked |
| Price | $0.042 per million input tokens; output tokens free | Source: docs.typesafe.aichecked |
| Context length | 64k tokens per request and 32k per question; 262k with the long-context header | Source: docs.typesafe.aichecked |
| Answer types | Choice, Score, Noul (yes probability) and Box; text input, images experimental | Source: docs.typesafe.aichecked |
| Seed round | $40 million led by DCVC, announced September 15, 2026 | Source: siliconangle.comchecked |
| Reported later round | About $870 million at a $7.5 billion valuation, led by Andreessen Horowitz | Reported by Bloomberg on October 9, 2026; not announced on TypeSafe's blog as of October 17. Source: news.bgov.comchecked |
| Your data | No training on your inputs; zero data retention for enterprise customers; SOC 2 Type II report on request | Source: typesafe.aichecked |
Answer first, then why, an example, the limits and what to do next.
Jev is TypeSafe AI's decision model: it answers typed questions about your data with probabilities instead of writing text. You send a state and questions and get back a choice, a score, a yes probability or an image box that code can branch on; TypeSafe opened it in early access on September 15, 2026.
No, Jev is not open source as far as we could find. TypeSafe publishes SDKs and an agent skill but no model weights; Jev runs only through TypeSafe's hosted API, and its docs say the same weights serve every account.
Jev costs $0.042 per million input tokens, and output tokens are free. That is $42 per billion input tokens for both jev-1.14.0 and jev-1.13.0; TypeSafe's docs list no free tier, and DataCamp saw no trial credits in September.
Limits
TypeSafe itself says it cannot yet prove the price is not subsidized.
Bloomberg reported on October 9, 2026 that TypeSafe raised about $870 million at a $7.5 billion valuation. Andreessen Horowitz led the round, per Bloomberg, and Pulse 2.0 added Sequoia Capital and DCVC; TypeSafe had not posted it on its blog by October 17, and its $40 million seed came less than a month earlier.
Jev can replace an LLM only for decisions with a fixed set of answers, not for chat or writing. According to TypeSafe's case study, Jack & Jill swapped it in for Gemini 3.1 Flash Lite to score candidates, and Classmethod's tester used it in place of a local LLM judge; TypeSafe's docs say it is not a drop-in model for Claude Code or Cursor.
Jev cannot return an answer outside the format you define, but it can still be wrong. TypeSafe's no-hallucination claim means every answer matches your schema; its own docs list failure modes such as counting, date comparison and a bias toward the first option, so check the confidence and test on your own data.
This page is built from TypeSafe's site, launch post, documentation, privacy policy and terms, from reporting by SiliconANGLE, VKTR, DataCamp, Bloomberg and Pulse 2.0, and from one independent hands-on test by a Classmethod engineer. Little beyond that is independently confirmed, and we say where. We checked everything on October 17, 2026.
TypeSafe AI, Inc. is a San Francisco company founded in 2024 by Diogo Almeida, who worked on RLHF, InstructGPT and ChatGPT at OpenAI, with Erik Gafni and Sasha Sheng (SiliconANGLE, VKTR). After two years in stealth it released Jev, which it calls the first System One Model. Jev does not write text: you send a block of data, the state, and a set of typed questions, and it returns a choice, a score or a yes probability; choices and scores also carry per-option probabilities and a confidence value (TypeSafe). TypeSafe trains it with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD).
| Date | Event | Source |
|---|---|---|
| September 15, 2026 | Jev opens in early access; $40 million seed led by DCVC | TypeSafe, SiliconANGLE |
| September 20, 2026 | Classmethod publishes an independent hands-on routing test | Classmethod |
| October 7, 2026 | TypeSafe publishes a case study with the hiring startup Jack & Jill | TypeSafe |
| October 9, 2026 | Bloomberg reports about $870 million raised at a $7.5 billion valuation, led by Andreessen Horowitz | Bloomberg |
| October 14, 2026 | jev-1.14.0 adds a Box type, experimental images and long context | TypeSafe |
Bloomberg says the round followed a launch video that went viral, and that TypeSafe claims about a third of Fortune 500 companies use Jev, without naming them. Pulse 2.0 reported the same round on October 11 and added Sequoia Capital and DCVC as investors (Pulse 2.0). TypeSafe's own blog had no post about the round on October 17. Its agent-skill repository has 2.7k GitHub stars (GitHub).
The one detailed independent test we found is Reona Shimada's at Classmethod. He replaced the LLM judge that routes a Slack agent's requests between a weak and a strong model. Over 13 scenarios run three times, Jev routed 39 of 39 correctly against 36 of 39 for the local judge, took 0.27 to 0.42 seconds per judgment instead of about 1.55, and 130 direct calls cost about $0.003. He also listed the trade-offs: request text now leaves the local machine, Jev gives no reasoning, and the scenarios were short (Classmethod).
Every's Mike Taylor reviewed Jev on launch day; the free part of his piece says how well it does the job is still an open question (Every). Everything else, including the headline 193.6 times faster and 444.6 times cheaper, comes from four workflows TypeSafe's own team wrote and scored against answers from GPT-6 Astra and Claude Fable 5.1. SiliconANGLE, VKTR and DataCamp all note that nobody has reproduced them (DataCamp).
Jev costs $0.042 per million input tokens, and output is free (TypeSafe docs). You reach it through the API console, Python and JavaScript SDKs or a browser Playground (quick start). TypeSafe still calls Jev early access, and testers in September went through a waitlist. TypeSafe says it does not train on your inputs, offers zero data retention to enterprise customers and shares a SOC 2 Type II report on request (TypeSafe).
Jev is not a chatbot, and TypeSafe's docs say it is not a drop-in model for coding agents such as Claude Code or Cursor (TypeSafe docs). It competes with cheap LLM calls used as classifiers, such as Gemini 3.1 Flash Lite, which Jack & Jill replaced with Jev for candidate scoring according to TypeSafe's case study, and with local classifiers like the judge in Classmethod's test. For writing, code or explanations you still need a generative model.
Sources
Changelog
APIMODELS is a pay-as-you-go AI API relay that sells about 150 models through one key at its own prices, mostly below the makers' list prices, and spends the margin on backup channels: 98.62% of requests succeeded over 30 days, moderation rejections excluded. APIMODELS is our own product: it is made by the team behind GenGrades.
ZenMux is an LLM gateway, operated under its terms by NexaMind Singapore Pte. Ltd., that sells Claude, GPT, Gemini and many Chinese models through OpenAI, Anthropic and Vertex-style endpoints at the makers' list token prices. It adds automatic credits for bad or slow answers, which it calls insurance, and 20 to 200 USD monthly coding plans.
Vercel AI Gateway is Vercel's hosted API for about 480 text, image, video and speech models behind one key, billed from prepaid credits at the providers' list prices with no markup, including when you bring your own keys. It became generally available on August 21, 2025, and gives each Vercel team 5 USD of free credit a month.
Requesty is a London-based LLM gateway, run by Requesty Ltd and seed-funded by 20VC in 2025, that sells OpenAI- and Anthropic-compatible access to more than 600 models at list price plus a 5% top-up fee. It stands out for routing policies, spend controls and an EU endpoint in Frankfurt.
Poe API is Quora's developer API for the bots in its Poe chat app, opened on July 31, 2025, with OpenAI-compatible endpoints and an Anthropic-compatible one for official Claude bots. Calls spend the same points as a Poe subscription, so there is no separate API bill; current Claude bots cost about 1% more than Anthropic's list price.
OpenRouter is a New York company's API relay that sells access to more than 680 models from over 100 hosting providers through one key, at each provider's list price plus a 5.5% fee on credit purchases. Stripe agreed to buy it on August 19, 2026; no closing had been announced by October 21, 2026.