ChatGPT Co-Creator's New Jev Model Ditches Text for 200x Faster AI Decisions
News Summary
A startup founded by one of the researchers behind ChatGPT has launched an AI model that skips natural language entirely, returning structured probabilistic decisions instead of text — and early testers say it is dramatically faster and cheaper than mainstream large language models for the tasks it targets. The model, called Jev, comes from TypeSafe AI, a company started by Diogo Almeida, a former OpenAI researcher credited as a co-inventor of reinforcement learning from human feedback (RLHF), the training technique widely credited with making ChatGPT and its successors usable at scale.
What TypeSafe AI Announced
TypeSafe AI emerged from two years of stealth work on Tuesday, September 16, 2026 Eastern Time, unveiling what it calls a "System One model." Rather than generating conversational text one token at a time like GPT-style large language models, Jev is designed to be consumed directly by other software: it takes structured inputs and returns typed, probability-scored outputs — for example, classifying a request, scoring a record, or flagging a policy violation — without producing any prose at all. The company says it raised $40 million in funding to pursue what Almeida described as "an alternative path for AI research, focused on machine-native AI."
How Jev Differs From Traditional LLMs
Jev is trained using a technique TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD), which the company positions as a successor to the RLHF and RLVR (reinforcement learning from verifiable rewards) methods that underpin most current chat-oriented models. Instead of predicting output sequentially, token by token, Jev samples all of its possible outputs in parallel within a single query. Because developers define a fixed output schema in advance, TypeSafe says the model is mathematically constrained from producing malformed or "hallucinated" results — it can only return values that fit the schema it was given.
Almeida framed the launch as addressing a gap he sees in the industry: "Models have been superhuman at chat for years, so where is all the automation?" The pitch is that most software still relies on brittle, hand-written rules or slow, expensive LLM calls for decisions like routing requests, sorting content, or screening outputs for jailbreak attempts — tasks TypeSafe argues are better served by a purpose-built decision engine than by a general-purpose chatbot.
Performance and Pricing Claims
TypeSafe published benchmark comparisons showing Jev responding in roughly 70 to 500 milliseconds, which the company describes as 40 to 200 times faster than frontier chat models on comparable structured-decision tasks. In one demonstration, the company had Jev play the classic video game Doom by feeding it structured game-state data instead of images; TypeSafe said Jev responded in about 0.114 seconds, compared with roughly 8.566 seconds for a comparably positioned OpenAI model in the same test setup.
On pricing, TypeSafe set input tokens at $0.042 per million tokens with output tokens offered free of charge, a structure the company contrasts with the metered input-and-output pricing typical of conversational LLMs. TypeSafe cites internal workflow evaluations claiming Jev is roughly 193.6 times faster and 444.6 times cheaper than comparable large language model approaches on certain production-grade automation workflows, though these figures come from the company's own benchmarking rather than independent third-party testing.
Developer Reception
The reaction from early adopters has been notably enthusiastic. Engineers at Vercel reported that swapping out an OpenAI conversational model for Jev in part of their infrastructure produced results five to eighteen times faster while maintaining or improving accuracy. Engineers at Bryo AI said they found Jev roughly ten to twenty times cheaper than comparable offerings from Google's Gemini family, while also providing genuine calibrated probability scores that made it easier to build reliable workflow-automation logic rather than parsing free-text responses for signals.
Demand following the announcement was strong enough that TypeSafe said its API briefly could not keep up with incoming traffic, an outage the company attributed directly to the volume of developer interest rather than any planned capacity constraint.
Why the Framing Matters
TypeSafe's name for Jev nods to 19th-century economist William Stanley Jevons, whose paradox holds that as the cost of using a resource falls, total consumption of that resource tends to rise rather than fall. Almeida has invoked that idea to describe his vision for where AI decision-making is headed: not a smaller number of very large, general-purpose chat models handling every task, but a much larger number of small, fast, inexpensive decision models embedded throughout software infrastructure. "We think that there's just going to be smart software all over the place in a way that's emergent and distributed," Almeida said, describing System One models as a complement to — rather than a replacement for — the large language models that power conversational AI products.
What Comes Next
TypeSafe has positioned Jev primarily for backend and infrastructure use cases: model routing, content moderation and safety screening, workflow automation, and large-scale data classification, rather than consumer-facing chat applications. Because the model's outputs are schema-constrained rather than open-ended text, it is not intended to compete directly with chatbots or writing assistants. Independent researchers and rival labs have not yet published third-party benchmarks confirming TypeSafe's performance claims, and how the System One category performs across a broader range of real-world production workloads — beyond the use cases TypeSafe itself has highlighted — remains to be seen as more developers put the model into production.