TypeSafe AI
TypeSafe AI (styled TypeSafe) is a San Francisco AI lab founded in 2024 that builds what it calls "machine-native, composable AI": models whose outputs are typed decisions with probabilities for other software to consume, rather than text for people to read. The company came out of stealth on 15 September 2026 with a seed round of approximately $40 million led by DCVC and released its first model, Jev, into waitlisted early access the same day [1][2]. It removed the waitlist and opened Jev to everyone on 20 September 2026 [23][24]. On 9 October 2026 it announced an $870 million Series A led by Andreessen Horowitz at a $7.5 billion valuation, with Sequoia Capital, DCVC and angel investors participating [26][28][29]. Its co-founder and chief executive is Diogo Almeida, a former OpenAI researcher who is the fourth listed author of the InstructGPT paper; the company describes him as a co-inventor of RLHF and ChatGPT [1][15]. The other founders are Erik Gafni (CTO) and Sasha Sheng (COO) [1][6].
TypeSafe's position, set out in a company manifesto under the tagline "Build Prod, Not God", is that today's models are already intelligent enough to create large economic value but are "hard to build on" because they were trained to please people, and that a different training target, calibrated decisions rather than human preference, is what production automation needs [5]. Its multipliers for Jev's speed and cost (ranging from "up to 100 times" in the press release to "193.6x faster, 444.6x cheaper" on its home page) are the company's own measurements on its own workflow evaluations, and the company's evaluation site shows Jev behind several frontier models on accuracy while far ahead on cost and latency [1][7][8]. Its adoption figures are also its own: TypeSafe told Bloomberg that Jev passed one million users within days, and says about a third of the Fortune 500 use it, a share its lead investor puts at 25 percent; no customer names were given with either figure [26][27][30].
History
The company was founded in 2024 and worked in stealth for roughly two years before its public launch; Almeida's launch post says "after two years in stealth" and his announcement tweet says he "spent the last 2 years in stealth building a new way to train models (RLCD)" [4][16]. DCVC's portfolio page lists its first investment in TypeSafe as 2025 [3]. The company's GitHub organization, typesafe-ai, has a repository dating from May 2024, consistent with the founding year in the press release [17].
On 15 September 2026 the company announced the seed round and Jev through a Business Wire release, a launch post by Almeida, a Q&A published by DCVC general partner James Hardiman, and a tweet from Almeida that had drawn about 18.8 million views and 53,600 likes by the following day [1][2][4][16]. Forbes covered the launch under the headline "This $200 Million Startup Wants To Fix AI's Overconfidence Problem"; DCVC linked to that piece as its coverage of the round [2][19]. The Register ran a story the next morning under the headline "TypeSafe AI debuts model for machines that plays Doom" [10]. TypeSafe had shown the model to outside developers a few days earlier: CoreWeave says the company, still in stealth, sponsored its CoreWeave Hacks event in San Francisco on 12 and 13 September 2026 and helped more than 60 teams build on Jev [51].
The following weeks moved quickly. On 24 September 2026 The Information reported that TypeSafe had started talking to investors about raising "$1 billion or even more" and that some unnamed investors had offered to invest at a valuation of "$10 billion or more", adding that talks were early [31]. In early October The Wall Street Journal quoted Almeida as saying that about a quarter of Fortune 500 companies were using Jev and that "we were at a trillion tokens per day about a week ago" [32]. TypeSafe published its first customer case study on 7 October, announced its Series A on 9 October and released Jev 1.14.0 on 14 October [26][33][36]. On 13 October Workday made its Sana assistant, whose interface Jev assembles, the home screen of its HR, finance and IT software [35]. Over the same weeks OpenAI and Amazon Web Services released competing decision models (see Competition) [39][40][41].
| Date | Event | Source |
|---|---|---|
| 2024 | Company founded in San Francisco | Press release [1] |
| May 2024 | First repository created in the typesafe-ai GitHub organization | GitHub [17] |
| 2025 | DCVC's first investment in TypeSafe | DCVC portfolio page [3] |
| 31 Mar 2026 | "Founders You Should Know" video interview with Almeida posted to the company blog | TypeSafe blog [14] |
| 19 Jun 2026 | Almeida's AI Council talk "AI: too good to be true, too bad to be useful" posted to the company blog | TypeSafe blog, AI Council [13] |
| 8 Aug 2026 | system-one-adapter-python repository created (MIT) | GitHub [17] |
| 4 Sep 2026 | typesafe-sdk-python and typesafe-sdk-js repositories created (MIT) | GitHub [17] |
| 10-11 Sep 2026 | "The Bitterest Lesson" and "Lies, Damned Lies, and Benchmarks" essays published | TypeSafe blog [11][12] |
| 12-13 Sep 2026 | Sponsors CoreWeave Hacks in San Francisco while still in stealth; teams build on Jev before launch | CoreWeave on X [51] |
| 15 Sep 2026 | Emerges from stealth; ~$40M seed led by DCVC; Jev enters early access | Press release, DCVC, launch post [1][2][4] |
| 20 Sep 2026 | Waitlist removed; Jev opened to everyone | TypeSafe on X, home page [23][24] |
| 24 Sep 2026 | The Information reports talks about a round of $1 billion or more | The Information [31] |
| 29 Sep 2026 | OpenAI announces a Decisions API at its developer day | TechCrunch [39][41] |
| 1 Oct 2026 | AWS releases the open-source Strands Decider 2B, described by TechCrunch as inspired by Jev | TechCrunch [40] |
| 7 Oct 2026 | First published customer case study (Jack & Jill) | TypeSafe blog [33] |
| 9 Oct 2026 | $870M Series A led by Andreessen Horowitz at a $7.5B valuation; Martin Casado joins the board | TypeSafe blog, TechCrunch, Bloomberg [26][28][29] |
| 13 Oct 2026 | Workday makes its Jev-driven Sana interface the home screen of its software | The New Stack [35] |
| 14 Oct 2026 | Jev 1.14.0 released, with opt-in image input, long context and a bounding-box type | TypeSafe blog, docs [36][37] |
Founders and team
| Person | Role | Background as described by TypeSafe |
|---|---|---|
| Diogo Almeida | Co-founder and CEO | "co-invented RLHF and InstructGPT, the methods that led to ChatGPT and GPT-4"; previously at Google Brain [6] |
| Erik Gafni | Co-founder and CTO | Repeat founder (Ravel, "multi-modal AI for DNA sequencing"); early employee at Invitae and Freenome [6] |
| Sasha Sheng | Co-founder and COO | Former research engineer at Meta AI's FAIR, working on News Feed, AI Experiences and AI Research; published at NeurIPS and ECCV [6] |
| Ke Deng | Chief of Staff | Press contact on the launch release [1] |
The "co-inventor" wording for Almeida is the company's and his own; the press release calls him "co-inventor of RLHF/ChatGPT", the team page says he "co-invented RLHF and InstructGPT", and DCVC's post says he "co-invented some of the core techniques and products that launched the current boom: InstructGPT, ChatGPT, and GPT-4" [1][2][6]. What is independently verifiable is that Almeida is the fourth author of "Training language models to follow instructions with human feedback" (Ouyang et al., 2022), the InstructGPT paper [15]. In his launch post Almeida writes that at OpenAI he "helped build the methods that made language models useful at following instructions and talking with people" and that "that work ended up as the research behind ChatGPT" [4].
The team page says the company is "a close-knit, flat team from OpenAI, Google Brain, Meta/FAIR, Stripe, Airbnb, Plaid, Docker, and more" that works in person five days a week from an office "near the Embarcadero station" in San Francisco [6]. Its LinkedIn page listed the company size as 11-50 employees in September 2026 and still did on 27 October 2026 [20]. As of 16 September 2026 its Ashby job board carried five full-time San Francisco openings: Member of Technical Staff (Model Capabilities), Member of Technical Staff (Backend/Platform), Member of Technical Staff (Infrastructure, Kubernetes), Member of Staff, and Founding Marketer [18]. The job listings describe the stack as primarily Python, with TypeScript, Next.js and Tailwind CSS for front ends and Kubernetes for orchestration [18]. By 27 October 2026 the board had grown to 15 San Francisco openings, 14 full-time and one six-week contract role, including Developer Advocate, Founding Recruiter, Product Manager and Head of Legal [45].
Funding
| Round | Date announced | Amount | Valuation | Lead | Other named investors |
|---|---|---|---|---|---|
| Seed ("Series Seed" in DCVC's wording) | 15 Sep 2026 | "$40 million" (headline); "approximately $40 million" (boilerplate) | Not disclosed by the company; Forbes' headline calls it a "$200 Million Startup", and The Information cited a PitchBook figure of $200 million [19][31] | DCVC (James Hardiman, General Partner) | None named in the press release, DCVC's post, or FinSMEs' summary [1][2][21] |
| Series A | 9 Oct 2026 | "$870 million" | "$7.5B" (pre- or post-money not stated) [26] | Andreessen Horowitz; general partner Martin Casado joins the board [26][27] | Sequoia Capital, existing investor DCVC, and angel investors [26][28] |
DCVC's portfolio page dates its first investment in TypeSafe to 2025, which indicates the firm's involvement predates the September 2026 announcement; neither DCVC nor the company says when the $40 million round closed [3]. Hardiman's quoted rationale is that TypeSafe is "turning increasingly capable models into technology that developers can reliably build into products at scale"; in his own post he frames the bet around "no-humans-in-the-loop automation" that "just hasn't shown up yet" [1][2]. Neither the release nor DCVC's post names other participants or a valuation. The public valuation figures for the seed are the "$200 Million Startup" of Forbes' headline and a $200 million valuation that The Information later attributed to PitchBook; the team page says only that the company is "backed by top-tier investors" [1][2][6][19][31].
Series A
On 9 October 2026 TypeSafe announced that it had raised $870 million at a $7.5 billion valuation in a round led by Andreessen Horowitz, with participation from Sequoia Capital, existing investor DCVC "and the tech illuminati (aka a bunch of the best angel investors)", and that Martin Casado, a general partner at Andreessen Horowitz, would join its board [26][27]. TechCrunch and Bloomberg reported the same amount and valuation [28][29]. Bloomberg added that Andreessen Horowitz wrote the largest check, and that Almeida said the deal came together after the Jev launch video went viral and drew interest from other chief executives, developers and investors [29][30]. The round came in below both figures The Information had reported about two weeks earlier ($1 billion or more, at a valuation of $10 billion or more) [31]. Neither the company's post, TechCrunch's report nor the free portion of Bloomberg's says whether $7.5 billion is a pre-money or post-money figure.
TypeSafe's post, titled "TypeSafe A raises Series AI", says the company finds "fundraising announcements incredibly boring" and instead lists what the money means: for developers, "even more machine-native models" and "everything else to be the best infrastructure for building smart software"; for businesses, the enterprise features customers had asked for; and for recruits, "We are a real company and we're here to stay" [26]. Andreessen Horowitz's own announcement, signed by Jennifer Li, Sarah Wang, Martin Casado, Marc Andreessen and Ben Horowitz, called the launch "the biggest narrative violation we've seen this year" and said the firm was "leading the investment" [27].
Thesis: machine-native, composable AI
TypeSafe's manifesto, titled "Composable AI: Build Prod, Not God", states a mission "to pave the shortest path to an AI-based economic revolution by making intelligence composable to catalyze a Cambrian explosion of intelligent software" [5]. Its argument runs as follows. Current models "have long since crossed the threshold of intelligence needed for creating massive economic value", yet "most software still isn't meaningfully intelligent", so "the bottleneck isn't raw intelligence. It's that today's intelligence is hard to build on" [5]. The manifesto compares chat-style AI to early "horseless carriages" that kept the shape of the thing they replaced, and argues that because RLHF "directly optimizes for human preference", the foreseeable consequence is "AI that requires humans in the loop instead of running in the background" [5]. The alternative it proposes is AI "as a primitive that any programmer can invoke for semantic judgement and decisions, while still using code for what it's best at: exact computation", which the manifesto's own footnote describes as the neuro-symbolic dream "cheekily summarized as 'smart if-statements'" [5].
The document lays out three steps: ship "the shape of machine-native composable AI with the highest possible intelligence-per-dollar"; make that AI "reliable enough to transform the economy via real automation"; then provide "higher-level intelligence abstractions that are stable enough to compose and layer upon" [5]. Its yardstick for an economic revolution is global total factor productivity growth "reaching 3% within five years and holding at that level for ten", and it says it is "not racing towards the ever-moving goalposts of 'AGI'" [5].
In DCVC's Q&A Almeida adds that intelligence should become "a utility, a basic building block of automation" and predicts that once AI has automated most of the economy "over 99% of AI calls will be made by and for software, not humans" [2].
Technology: System One Models and RLCD
TypeSafe calls its model class "System One Models", after the fast, intuitive System 1 thinking in Daniel Kahneman's Thinking, Fast and Slow [4]. According to the launch post, the company built "a new stack entirely focused on automation: with a new model architecture, parallel sampler for maximum efficiency, and training method we call Reinforcement Learning for Calibrated Decisions (RLCD)" [4]. An arXiv search for the phrase "Reinforcement Learning for Calibrated Decisions" returned no paper as of 16 September 2026, and the company's public description of RLCD is limited to the launch post and the home page, which lists "a new architecture, a new sampler, and a new training algorithm" [4][7]. The launch post's FAQ answers the question "Why was a new training algorithm needed?" only at the level of objectives: RLHF optimizes for "the text that a human rater prefers", which the company calls "the wrong task for automation", and RLVR "is great for tasks with simple programmatic verification, but most real-world judgement tasks don't fit into that shape" [4].
Later interviews added some detail. TechCrunch described Jev as based on a transformer architecture but not a large language model, and reported that Almeida says it is trained exclusively on synthetic data; outside observers, it added, suspected the model was built on top of an open-weight LLM [28][42]. On The MAD Podcast in October 2026 Almeida said TypeSafe "did not do pre-training": the company uses "a combination of models", drawing on "the variety of open-weight models that we can link up", and makes all of its training data in-house because its task needed "a new shape of data" [43]. He did not name the base models. He has also told TechCrunch that the synthetic data the company creates is its moat [39].
The developer-facing interface is documented at docs.typesafe.ai. A call sends a "state" (a string or JSON object) plus a set of typed "questions", and the model returns "typed values and probability distributions that your code can branch on, sort by, and route with" [9]. There are three question primitives [8][9]:
| Primitive | Question type | Returns |
|---|---|---|
| Noul | Yes or no | A probability of "yes" |
| Choice | One option from a defined set | A distribution over the options, plus a confidence value |
| Score | A level on a scale | A score, a distribution over levels, plus a confidence value |
The launch post contrasts the approach with existing LLMs in a table whose main claims are the company's own: outputs are "type-safe structured values" (structured output whose shape is "defined in advance"), so "the model never makes type errors"; sampling is parallel, generating "all outputs in a single query" rather than one token at a time; and every answer carries "calibrated probabilities and confidence scores" [4]. The company states that the "no type errors" claim is guaranteed by schema matching and is "not empirical", which is also why it plots Jev's hallucination rate as zero [4]. The Register noted that this "really isn't a fair comparison as its output is not natural language" and "does not preclude the possibility of being incorrect" [10]. The company's home-page FAQ poses the question "Can Jev still get things wrong?" and its launch post asks early-access developers to report "where Jev works, and where it falls short" [4][7]. The launch post also reports that Jev supports a cardinality of up to 255 options per choice, with a two-stage scoring-then-choosing procedure for larger sets [4].
Jev
Jev, named for the economist William Stanley Jevons and the Jevons paradox, is TypeSafe's first public System One Model, released in waitlisted early access on 15 September 2026 [1][4]. VentureBeat reported that, according to Almeida, the company cleared 140,000 people from the waitlist within 36 hours of launch [25]. On 20 September 2026 TypeSafe's X account posted "Jev is now available to everyone. No waitlist." with a link to its console, and the home page added a news item dated that day headed "No more waitlist" [23][24]. VentureBeat reported that new users get $5 in free credit [25]. The company's published price is $0.042 per million input tokens ("$42 per billion tokens") with output tokens free, and it quotes end-to-end response times of 70 to 500 milliseconds [4]. The press release says the model "delivers frontier-level intelligence at less than 100 milliseconds of latency and is up to 100 times faster and less expensive than other frontier models" [1]. TypeSafe concedes that it "can't prove it isn't subsidized" and says it will "need the long-term to prove the sustainability of our pricing" [4]. As of 23 September 2026 the home page FAQ answers the question "Are these prices temporary or subsidized?" with "We can serve Jev profitably at our current prices" [24].
The speed and cost multipliers the company has published vary by venue and by comparison, and all are the company's own numbers:
| Figure | Where it appears | What it compares |
|---|---|---|
| "Up to 100 times faster and less expensive" | Press release [1] | Other frontier models, unspecified |
| "20-200x faster", "40-400x cheaper (w/ output tokens free)" | Almeida's launch tweet [16] | Unspecified |
| "40x-200x faster" | Launch post [4] | Frontier models "for System One shaped queries" |
| "193.6x Faster, 444.6x Cheaper" | Home page [7] | Company workflow evals; launch post calls these "the higher end of real world gains" [4] |
| 0.114 s vs 8.566 s; $0.000081 vs $0.013880 | Home-page side-by-side demo [7] | GPT-5.6 Terra at default reasoning, on a short, simplified query the company says paints its model "in an advantageous light" [4] |
| "238x lower input price than Claude Fable 5.1" | Home page [7] | Input price only, against Claude Fable 5.1 |
The company's evaluation site, evals.typesafe.ai, publishes four "workflow evals" (security incidents, agent-trace observability, invoice processing, customer service) in which every model runs the same decomposed decision workflow, and reference labels are the average of GPT-6 Astra and Claude Fable 5.1 at high thinking [8]. On the site's overview chart as of 16 September 2026, averaged across the four workflows, Jev scores 67.8% agreement with those labels at $0.0004 and 0.4 seconds per case, while GPT-5.6 Sol scores 74.1% at $0.0836 and 23.3 seconds and Claude Opus 5 scores 73.1% at $0.1761 and 37.8 seconds; Claude Sonnet 5 matches Jev's 67.8% at $0.1174 and 78.1 seconds [8]. In other words the company's own data show Jev on the cost and latency frontier rather than at the top of the accuracy column. The launch post lists the caveats: the workflows were written by the company's own model capabilities team, the reference answers are biased toward OpenAI and Anthropic models, and the LLM comparisons run through TypeSafe's own "System One LLM" wrapper [4]. TypeSafe later published the code for reproducing the site as the WorkflowEvals repository (created 28 September 2026) [44]. By late October the site had been reorganized to compare each model run as a structured workflow against the same policy given as a single prompt, with Fireworks-hosted models added, and it states that averaged over the four tasks every model is more accurate, cheaper and faster in the workflow form [52].
The company published two demonstrations alongside the launch: a bot that plays Doom from a structured text representation of game state (not images), making about ten queries a second at a cost the company puts at roughly $7 per hour, and a Wikiracing agent that chooses among "hundreds to thousands of links" per step [4][10]. The Register's headline led with the Doom demonstration [10].
TypeSafe released Jev 1.14.0 on 14 October 2026, first under the jev-preview alias, describing it as "small, subtle improvements to everything" with the largest expected gain in multilingual support, and said jev-latest would move to it a week later [36]. The same post exposed experimental features behind "danger zone" request headers that the company says it does not yet vouch for in production: image input (the post says the previous model, jev-1.13.0, "was multimodal all along"), a longer context window, and a new box type that returns a single bounding box and works only on Jev 1.14 [36]. The post also added an opt-in for developers to share queries that expose "jaggedness", while repeating that the company will "not train on your queries", and promised 60 days' notice before jev-1.13.0 is deprecated and 30 days for jev-1.14.0 [36]. The models documentation lists both versions at $42 per billion input tokens with a default limit of 64,000 tokens per request and 32,000 per question, raised to 262,000 for both with the long-context header; as of 27 October 2026 jev-latest pointed to jev-1.14.0 [37].
Customers and adoption claims
TypeSafe has published one detailed customer case study. Its 7 October 2026 post says Jack & Jill, a talent marketplace run by AI agents, replaced Gemini 3.1 Flash Lite with Jev for "100% of calls" in the initial-scoring stage of its candidate-matching pipeline within 10 days of its first test [33]. In Jack & Jill's offline test on 150 live roles, Jev kept 94.6% of the candidates that hiring managers later asked to meet, against 93.9% for Gemini, a difference the post says was not statistically significant; the cost of scoring 1,000 candidates fell from $0.755 to $0.092, and the median time to screen all candidates for a role from 20.3 to 10.3 seconds [33]. The headline saving of $500,000 a year is a projection that assumes growth; at current volume the post puts it at $265,000 [33]. The company's enterprise page also advertises "classification at 98.7% lower cost", along with a SOC 2 Type II report available through a trust center, zero data retention for enterprise customers, and a commitment not to use customer prompts or other inputs for training [34]. The 98.7% figure links to MotherDuck's 21 September 2026 post announcing a prompt_jev() SQL function, in which Jev classified 100,000 news articles from the AG News dataset in about 40 seconds for 50 cents at 89% accuracy, against 88%, $37.58 and almost 32 minutes for GPT-5.6 Terra [34][53].
Workday also uses Jev. The New Stack reported that Workday's Sana assistant, which became the home screen of Workday's HR, finance and IT software on 13 October 2026, uses Jev to choose from the components of Workday's design system as the user types, with none of the selection logic hard-coded, and that Workday has added TypeSafe to its list of subprocessors for natural language processing, intent classification and confidence scoring across all its product lines [35]. According to TypeSafe, as quoted by The New Stack, Jev runs only on the company's hosted API and processes data in the United States [35].
Usage figures come mostly from the company and its investors and differ in their details:
| Claim | Who made it | Source |
|---|---|---|
| Jev passed one million users "in a matter of days" | TypeSafe, to Bloomberg (October 2026) | [30] |
| About a quarter of the Fortune 500 use Jev; "a trillion tokens per day about a week ago" | Almeida, to The Wall Street Journal (early October 2026) | [32] |
| "1 trillion tokens generated in merely 3 days post launch"; 25% of Fortune 500 enterprises "have integrated Jev" | Andreessen Horowitz (9 October 2026) | [27] |
| "A third of the Fortune 500 are getting their Jev on" | TypeSafe Series A post (9 October 2026) | [26] |
| Used by 35% of paid teams on Vercel's AI Gateway within two weeks of launch, nearly double the reach of the next model, Claude Sonnet 5.5 | Vercel data, reported by TechCrunch (16 October 2026) | [38] |
| TypeSafe's offerings "for now, the most widely-queried on OpenRouter" among decision models | TechCrunch (22 October 2026) | [41] |
Bloomberg reported that TypeSafe declined to name the Fortune 500 companies, and The New Stack pointed out the gap between the company's one-third and a16z's 25 percent [29][35]. TechCrunch noted that Vercel's gateway data may undercount labs such as OpenAI and Anthropic that sell through their own APIs [38].
Competition
Larger companies followed within weeks. At OpenAI's developer day on 29 September 2026, Sam Altman announced a Decisions API that gives the company's Luna model a predefined set of options to choose between; TechCrunch reported that Almeida joked on X about the start of the "clone wars" and said OpenAI's interest could be "a sign...that building in a System One compatible way is the future" [39]. OpenAI made the API widely available on 22 October; Nikunj Handa, its API product lead, told TechCrunch that the team had reconfigured ChatGPT Luna for speed and cost rather than training a dedicated calibrated decision model [41]. On 1 October Amazon Web Services released Strands Decider 2B, an open-source model built on Qwen3.5-2B that TechCrunch described as inspired by Jev, and the outlet wrote that "dozens of similar models" had been produced by researchers since TypeSafe introduced the idea [40].
Almeida has played down the challengers. "I get that people think it's a gold rush, but they might be underestimating the difficulty of making the models actually smart," he told TechCrunch, adding that "the current batch seems more like ML people wanting to implement a cool architecture than a team deeply dedicated to making intelligence useful" [40]. Third-party rankings give a mixed picture: TechCrunch reported on 22 October that on the AIM System One Index and the Artificial Analysis Decision Index the latest Jev ranked seventh and third respectively, while both indices tended to favor D1 Ultra from Liquid and Base Lab's Elephant-1, and OpenAI's decision model placed eighth and fourteenth [41].
Open-source tooling and documentation
TypeSafe publishes client libraries and helpers under the MIT license on GitHub: typesafe-sdk-python and typesafe-sdk-js (both created 4 September 2026), system-one-adapter-python (created 8 August 2026), which the README describes as "a drop-in replacement for typesafe_sdk's system_one evaluation API, backed by LLM APIs instead of TypeSafe" for cost, speed and intelligence comparisons, and skills, agent skills for building with the System One API [17]. Later additions include n8n-nodes-typesafe-ai (created 23 September 2026), an n8n community node with "Evaluate" and "Route" operations; WorkflowEvals (28 September, under the Apache 2.0 license rather than MIT), the evals-site code; and typesafe-public-examples (5 October), content used in the documentation, demos and cookbooks [44]. The adapter is the "System One LLM wrapper" through which the company runs competing models in its evals [4][17]. The documentation site covers the primitives, "patterns" such as speculative fan-out and confidence-gated routing, cookbooks, and an API reference, and a hosted console with a playground is linked from the launch post [4][9]. Access is through the hosted API and console with an API key; the press release said early access was waitlisted, a restriction lifted on 20 September 2026 [23], and no repositories appear under a TypeSafe organization on Hugging Face as of 16 September 2026 [1][9].
On The MAD Podcast, after the host noted that Cloudflare's version of such a model is open source and asked whether TypeSafe was considering the same, Almeida said the company had debated it but had "not really committed to anything", arguing that an API lets it iterate much faster, and that an open release "could definitely be in the cards in the future, especially as we go into smaller and all on-device" [43].
Public positions on evaluation
Two unbylined pre-launch essays on the company blog, written in the first person by an author who describes working on InstructGPT at OpenAI, set out positions that shape how the company reports results. "The Bitterest Lesson" argues that Rich Sutton's bitter lesson is "the tip of an iceberg", and that in practice "doing the right task > data > compute > algorithms"; it cites the InstructGPT result that much smaller models trained on the right objective were preferred over GPT-3 [11]. "Lies, Damned Lies, and Benchmarks" (published at the URL slug "antibenchmaxxing") argues that any public eval a lab can optimize against "inevitably gets benchmaxxed", cites the Llama 4 Arena episode, Vending-Bench behaviour, and the successive revisions of the Artificial Analysis index after GPT-6 Astra's launch as examples, and commits TypeSafe to "no standard benchmark table in our model releases", to evals that are "dated snapshots and immediately retired once posted", and to publishing "the evidence that looks bad for us" [12]. The caveat blocks under each result in the launch post follow that policy [4]. The Jev 1.14 post went a step further by publishing the company's internal intelligence measurement, with caveats that it is "our best guess", is "likely biased towards our models", and compares Jev only with open-weight models because the company is wary of model providers' data policies [36]. TechCrunch remarked that TypeSafe released "its own rankings" despite Almeida's distrust of benchmarks [41]. On The MAD Podcast Almeida praised OpenAI for releasing its Decisions API without benchmarks [43].
Reception
Thomas Claburn of The Register, writing on 16 September 2026, summarized Jev as a model that "doesn't chat" but "produces typed probabilistic decisions", judged it "more likely to be used for sorting customer service problems and other business workflows" than for games, and flagged both the hallucination framing and the open question of whether demand for tokens will be as broad as demand for energy [10]. The Rundown AI noted that agreement with GPT-6 Astra and Claude Fable 5.1 "gives only a partial view of whether Jev makes correct decisions" and that "response times under load and sustained pricing remain open questions" [22]. FinSMEs recorded the round as a $40 million seed led by DCVC with no other investors listed [21].
Industry attention grew in the following weeks. The Information reported that at its AI Agenda Live conference on 23 September 2026 at least two speakers brought up Jev, including Nvidia's senior director of high performance computing, Dion Harris, who called it "incredibly fast" while noting the model's limitations [31]. TechCrunch, covering the Series A, wrote that the size of the round "comes as little surprise given Jev went viral almost instantly" [28].
Researchers took up the model quickly too, and papers on Jev began appearing on arXiv within days of the launch [46]. A review posted on 26 September 2026 counted 28 papers from 19 to 24 September and concluded that in this early literature the typed readout had "not shown an independent accuracy advantage" over comparable label-probability readouts, with Jev's clearest gains in latency and cost [46]. A zero-shot evaluation of jev-1.13.0 on 37 datasets found that it beat the open-weight Qwen3.8-27B on 27 of them and Gemma-4-E4B on all 37, with well-calibrated choice probabilities, while all three models degraded on low-resource languages and rubric-based judgments [47]. Other preprints found that Jev and flash-tier LLM judges, grading against the same rubrics, tend to be wrong in the same places [48]; that malicious injected content shifts Jev's action probabilities but rarely makes it pick the attacker's target, so typed outputs change but do not remove prompt injection risk [49]; and that type-correct outputs can still follow an option's name rather than the definition bound to it [50].
Continue reading
- Jev (AI model)Jev is an AI model released in waitlisted early access on 15 September 2026, and opened to all developers on 20 September, by TypeSafe AI, a San Francisco startup founded in 2024 by former OpenAI researcher…Mentioned 62 times above
- Diogo AlmeidaDiogo Almeida (full name Diogo Moitinho de Almeida) is a machine learning researcher and the co-founder and chief executive of TypeSafe AI, a San Francisco lab that came out of stealth on 15 September 2026…Mentioned twice above
- Intern-DecisionIntern-Decision is a family of three open-weight multimodal decision models (Intern-Decision-0.8B, Intern-Decision-2B and Intern-Decision-4B) released in late September 2026 by the InternLM team of Shanghai AI…Listed in See also
- Decision 2.0Decision 2.0 is a family of open-weight models for structured decisions from the vLLM Semantic Router project.Listed in See also
- Microsoft-Decision-1Microsoft-Decision-1 is a Microsoft model for scoring predefined choices in software workflows. Announced on 9 October 2026, it targets classification, model routing, response evaluation, and agent control.Listed in See also
- Calibration (machine learning)Calibration in machine learning is the property that the probability scores produced by a probabilistic classifier match the empirical frequency of the predicted event: a model that assigns a confidence of 0.8…Listed in See also
More from See also
More to read
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16TypeSafe AI Emerges From Stealth With $40M in Funding With New Model for Composable AI, Business Wire, 15 September 2026 (syndicated copies at Yahoo Finance and Morningstar).
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9James Hardiman, TypeSafe emerges from stealth with a new way of doing AI, DCVC News & Insights, 15 September 2026.
- ^1 ^2 ^3TypeSafe, DCVC portfolio page (First Investment: 2025), accessed 16 September 2026.
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23Diogo Almeida, Introducing System One Models & Jev, TypeSafe AI blog, 15 September 2026.
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Composable AI: Build Prod, Not God, TypeSafe AI manifesto, accessed 16 September 2026.
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Our Team, TypeSafe AI, accessed 16 September and 27 October 2026.
- ^1 ^2 ^3 ^4 ^5 ^6TypeSafe AI home page, accessed 16 September 2026.
- ^1 ^2 ^3 ^4Workflow evals, TypeSafe AI, accessed 16 September 2026.
- ^1 ^2 ^3 ^4Introduction, TypeSafe AI documentation, accessed 16 September 2026.
- ^1 ^2 ^3 ^4 ^5Thomas Claburn, TypeSafe AI debuts model for machines that plays Doom, The Register, 16 September 2026.
- ^1 ^2The Bitterest Lesson, TypeSafe AI blog, 10 September 2026.
- ^1 ^2Lies, Damned Lies, and Benchmarks, TypeSafe AI blog, 11 September 2026.
- ^AI: too good to be true, too bad to be useful, Diogo Almeida talk, AI Council; video embedded on the TypeSafe blog, 19 June 2026.
- ^Diogo Almeida - Founders You Should Know, TypeSafe AI blog (embedded YouTube interview from the Founders You Should Know channel), 31 March 2026.
- ^1 ^2Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida et al., Training language models to follow instructions with human feedback, arXiv:2203.02155, 2022.
- ^1 ^2 ^3Diogo Almeida (@CompleteSkeptic), launch tweet, X, 15 September 2026.
- ^1 ^2 ^3 ^4 ^5 ^6typesafe-ai GitHub organization: typesafe-sdk-python, typesafe-sdk-js, system-one-adapter-python, skills; repository metadata via the GitHub API, accessed 16 September 2026.
- ^1 ^2TypeSafe AI job board, Ashby, accessed 16 September 2026.
- ^1 ^2 ^3This $200 Million Startup Wants To Fix AI's Overconfidence Problem, Forbes (The Prompt), 15 September 2026 (headline; article body is paywalled).
- ^TypeSafe AI, LinkedIn company page, accessed 16 September and 27 October 2026.
- ^1 ^2TypeSafe AI Raises $40M in Seed Funding, FinSMEs, September 2026.
- ^TypeSafe launches Jev for AI decisions inside software, The Rundown AI, September 2026.
- ^1 ^2 ^3 ^4TypeSafe AI (@typesafeai), "Jev is now available to everyone. No waitlist.", X, 20 September 2026.
- ^1 ^2 ^3 ^4TypeSafe AI home page ("No more waitlist" news item dated 20 September 2026; FAQ), accessed 23 September 2026.
- ^1 ^2Companies are putting Jev in charge of AI agent decisions, and prompt injection can influence the verdict, VentureBeat, 21 September 2026.
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10TypeSafe A raises Series AI, TypeSafe AI blog, 9 October 2026.
- ^1 ^2 ^3 ^4 ^5Jennifer Li, Sarah Wang, Martin Casado, Marc Andreessen and Ben Horowitz, Investing in TypeSafe AI, Andreessen Horowitz, 9 October 2026.
- ^1 ^2 ^3 ^4 ^5 ^6Marina Temkin, The maker of non-text AI model Jev valued at $7.5B just weeks after launch, TechCrunch, 9 October 2026.
- ^1 ^2 ^3 ^4 ^5Natasha Mascarenhas and Rebecca Torrence, Andreessen Horowitz Backs Jev Maker at $7.5 Billion Value, Bloomberg, 9 October 2026 (free excerpt; remainder paywalled).
- ^1 ^2 ^3TypeSafe AI raises $870 million led by Andreessen Horowitz - Bloomberg, Investing.com (summary of Bloomberg's report), 9 October 2026.
- ^1 ^2 ^3 ^4 ^5 ^6Stephanie Palazzolo, Cory Weinberg and Valida Pau, Jev Fervor Leads to Talk of Big Valuation Boost, The Information, 24 September 2026 (free excerpt; remainder paywalled).
- ^1 ^2Elias Schisgall, The Wall Street Journal, article on TypeSafe and Jev copycats (paywalled), as summarized and quoted by Techmeme, 2 October 2026.
- ^1 ^2 ^3 ^4 ^5How Jack & Jill Cut Candidate Screening Costs By 88% With Jev, TypeSafe AI blog, 7 October 2026.
- ^1 ^2Jev for enterprise, TypeSafe AI, accessed 27 October 2026.
- ^1 ^2 ^3 ^4 ^5Frederic Lardinois, Workday is using TypeSafe's Jev to build user interfaces on the fly, The New Stack, 13 October 2026.
- ^1 ^2 ^3 ^4 ^5 ^6Jev: For the builders AND dreamers, TypeSafe AI blog, 14 October 2026.
- ^1 ^2Models, TypeSafe AI documentation, accessed 27 October 2026.
- ^1 ^2Russell Brandom, Jev is Vercel's fastest-adopted AI model ever, TechCrunch, 16 October 2026.
- ^1 ^2 ^3 ^4Tim Fernholz, OpenAI's Jev clone could help the frontier lab stop its swarming agents, TechCrunch, 30 September 2026.
- ^1 ^2 ^3 ^4Tim Fernholz, Amazon releases its own Jev clone as decision models flood the web, TechCrunch, 1 October 2026.
- ^1 ^2 ^3 ^4 ^5 ^6Tim Fernholz, OpenAI makes its Jev competitor widely available as decision model competition heats up, TechCrunch, 22 October 2026.
- ^Tim Fernholz, A new kind of AI model from a ChatGPT inventor is thrilling developers, TechCrunch, 18 September 2026.
- ^1 ^2 ^3Matt Turck, Jev: The AI That Makes Decisions, Not Text - and Why It Changes Everything (Diogo Almeida), The MAD Podcast (episode page with transcript), 15 October 2026.
- ^1 ^2typesafe-ai GitHub organization: n8n-nodes-typesafe-ai, WorkflowEvals, typesafe-public-examples; READMEs and repository metadata via the GitHub API, accessed 27 October 2026.
- ^TypeSafe AI job board, Ashby (listings via Ashby's public job-posting API), accessed 27 October 2026.
- ^1 ^2Lijuan Tang and Yuemeng Zheng, Typed Decision Models: An Early Evidence Audit and Evaluation Checklist, arXiv:2609.32160, 26 September 2026.
- ^Tobias Deußer, Lorenz Sparrenberg and Rafet Sifa, Evaluating and Benchmarking the System One Model Jev, arXiv:2609.37647, 29 September 2026.
- ^Delip Rao and Chris Callison-Burch, Decision Models vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places, arXiv:2609.29769, 24 September 2026.
- ^Tiantong Wu and Wei Yang Bryan Lim, Decision Hijacking: Prompt Injection Attacks on Jev's Typed Probabilistic Decisions, arXiv:2609.28613, 23 September 2026.
- ^Yu Sun, Junhao Xu, Jiajia Shi and Zijin Yang, Type-Safe Is Not Error-Free: Typed Decision Models Follow the Option Name, Not the Definition Bound to It, arXiv:2609.26758, 22 September 2026.
- ^1 ^2CoreWeave (@CoreWeave), post on CoreWeave Hacks and Jev, X, 19 October 2026.
- ^Workflow evals, TypeSafe AI, accessed 27 October 2026.
- ^Hamilton Ulmer, Till Döhmen and Jacob Matson, Introducing prompt_jev(): bringing Jev to Motherduck SQL, MotherDuck blog, 21 September 2026.
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
6 revisions · v6 · 6,469 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Funding rounds, Jev releases, customer case studies, open-source repositories and early research papers checked against company posts, GitHub, arXiv and TechCrunch, Bloomberg and The Information coverage, October 2026.
Cite this page: AI Wiki. "TypeSafe AI." aiwiki.ai, updated 28 Oct 2026, fact-checked 27 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/typesafe_ai