# TypeSafe raises $40M for Jev, its AI model built to skip chat > Source: > Published: 2026-09-18 16:33:37+00:00 # TypeSafe raises $40M for Jev, its AI model built to skip chat **TypeSafe AI's Jev made 47 Pong decisions in 12 seconds in an Ably demo; Gemini, Claude and GPT made two or three, while still choosing correctly in most runs.** By [Ryan Merket](https://runtimewire.com/author/ryan-merket) ยท Published Primary source: [Ably Labs](https://jev-pong.ably.dev/) ## Why it matters TypeSafe is betting developers will buy narrower models that make bounded decisions faster and cheaper than chat models. Pong proves latency; production will test calibration. [Diogo Almeida](https://openai.com/contributions/gpt-4/?ref=runtimewire), Erik Gafni and Sasha Sheng took [TypeSafe AI](https://typesafe.ai/?ref=runtimewire) out of stealth on September 15th with a $40 million seed round led by [DCVC](https://www.dcvc.com/news-insights/typesafe-emerges-from-stealth-with-a-new-way-of-doing-ai/?ref=runtimewire) and a model built to give software bounded decisions instead of prose. The San Francisco lab's first public model, Jev, is Almeida's attempt to address a limitation in work he helped create. OpenAI credits him under "Foundational RLHF and InstructGPT work" in its [GPT-4](https://runtimewire.com/models/openai/gpt-4) contributor record. Those methods helped turn next-token predictors into assistants that could follow instructions and converse with people. Almeida now argues that the same human-preference objective leaves general chat models poorly suited to autonomous, machine-speed decisions. "People can't be the only consumers of intelligence," Almeida said in TypeSafe's launch announcement. His wager is that most useful AI will eventually run inside ordinary software, making small judgments that users never see. Gafni and Sheng bring the production side of that thesis. [TypeSafe's team page](https://typesafe.ai/team?ref=runtimewire) describes Gafni as a repeat founder who built Ravel around machine learning for DNA sequencing and worked at Invitae and Freenome. Sheng previously worked as a research engineer at Meta and FAIR across News Feed and AI research. Together, the founders spent roughly two years developing a separate architecture, sampler and training process rather than wrapping another language model in a stricter prompt. ### A model that refuses to write In [TypeSafe's September 15th launch post](https://typesafe.ai/blog/introducing-system-one-models-and-jev?ref=runtimewire), Almeida describes Jev as a "frontier-intelligence function call." Developers submit unstructured state and define the allowed response types in advance. Jev returns choices, scores or boolean probabilities, along with confidence estimates that an application can use when deciding whether to act or escalate. The distinction matters inside a software loop. A chat model usually generates tokens sequentially, even when an API constrains the final response to JSON or another schema. TypeSafe says Jev samples multiple typed outputs in parallel and gives up open-ended string generation entirely. That design also explains TypeSafe's claim of "zero hallucinations." TypeSafe is using a narrow, software-oriented definition: Jev cannot invent an off-schema answer or malformed tool call because its possible outputs are fixed. Schema compliance does not guarantee a correct decision. Jev can still select the wrong allowed answer, which puts the burden on TypeSafe's separate claim that its probabilities are calibrated well enough for developers to set useful confidence thresholds. ### Pong makes the latency visible A [playable Pong comparison](https://jev-pong.ably.dev/?ref=runtimewire) built by [Matthew O'Riordan](https://ably.com/blog/author/matthew?ref=runtimewire), co-founder of realtime infrastructure provider Ably, gives TypeSafe's speed pitch a legible form. Four lanes receive the same numerical game state and the same instruction: move a paddle up, down or leave it in place. Each answer advances the ball by one step, so slow responses produce visibly slow games. The recorded run was captured on September 17th from Vercel's iad1 region through Vercel AI Gateway. According to the demo, Jev averaged 227 milliseconds per decision, with a 400 millisecond p95. [Gemini 3.8 Flash](https://runtimewire.com/models/google/gemini-3.8-flash) averaged 3.2 seconds, [Claude Haiku 4.5](https://runtimewire.com/models/anthropic/claude-haiku-4.5) averaged 2.5 seconds and GPT-5.6 Sol averaged 3.5 seconds. During the first 12 seconds, Jev returned 47 decisions. The three chat models returned three, two and two, respectively. O'Riordan published the [source code and recorded statistics](https://github.com/ably-labs/jev-pong?ref=runtimewire). The chat models received structured-output requests at temperature zero, with reasoning disabled and no retries. All four lanes used the same Vercel AI Gateway key. The latency shown beside each lane measures the server-side model call. O'Riordan calls it a latency demo. That boundary is important. He says the chat models selected the correct Pong move 95% to 100% of the time across Ably's runs. Jev's advantage in the recording came from answering far more frequently, rather than demonstrating better strategy. The game is still a relevant stress test. A 2024 [Atari-GPT study](https://arxiv.org/html/2408.15950v2?ref=runtimewire) found that multimodal language models struggled as zero-shot, low-level game controllers, with latency and spatial reasoning among the problems. The Jev demo removes most visual and strategic complexity by handing every model an intercept coordinate. It isolates the question TypeSafe wants developers to consider: what becomes practical when a semantic decision fits inside a realtime loop? Ably also has an incentive to ask that question. O'Riordan has been positioning Ably's realtime channels as infrastructure for live, stateful agents. In the Pong implementation, the AI controller joins the same channel as the player and spectators, reads state and publishes moves. Faster decisions make both Jev and Ably's transport thesis look more useful. The public code makes the setup inspectable, while the results remain one recorded run selected and configured by the builder. ### TypeSafe's benchmarks carry their own caveats TypeSafe separately claims Jev was 193.6 times faster and 444.6 times cheaper than language models across selected automation workflows. Those figures come from [TypeSafe's own evaluations](https://evals.typesafe.ai/?ref=runtimewire), rather than an independent benchmark. Almeida's launch post includes unusually direct limitations. TypeSafe says its shorter inputs can favor Jev, its workflow authors may have introduced bias, and the advertised gains probably sit near the high end of what users should expect. The reference answers also come from the average predictions of large external models instead of independently established ground truth. The Pong page avoids some of those evaluation problems because it measures round-trip time and exposes the prompt, code and replay statistics. It leaves accuracy, calibration under changing conditions, sustained throughput and production reliability largely untested. Pricing is less mysterious than the published units initially make it appear. TypeSafe advertises $42 per billion input tokens, while [Vercel lists Jev](https://vercel.com/ai-gateway/models/jev?ref=runtimewire) at $0.04 per million. The two figures are consistent after rounding: $42 per billion equals $0.042 per million. Vercel lists no separate output-token charge because Jev returns typed evaluations rather than generated passages. ### What the $40 million is buying DCVC's $40 million Series Seed gives TypeSafe substantial room to train, serve and test a new model architecture before the usual revenue milestones associated with application software. Jev remains in early access, so TypeSafe's immediate task is proving that a specialized decision model can earn a durable place between deterministic code and general-purpose language models. The positioning steers TypeSafe away from a direct chatbot contest with OpenAI, Anthropic and Google. Jev is designed for routing, verification, classification, approvals, tool selection and agent control, where the response space can be defined before inference. General models retain the advantage when a task requires open-ended writing, broad reasoning or code generation. Almeida's core argument is that the AI market has over-applied the chat interface because chat produced the industry's first mass-market success. Jev turns that critique into a product with a deliberately smaller output surface. Pong shows why speed can matter once every answer becomes a software action. TypeSafe's larger test will be whether Jev's confidence scores remain trustworthy when the decisions carry financial, operational or safety consequences instead of moving a paddle.