A ChatGPT Co-Inventor Says the Missing Piece of AI Is a Model That Won't Talk

On September 15, a two-year-old startup called TypeSafe AI came out of stealth with $40 million in seed funding and a blunt diagnosis of the AI industry: chat models are very good at talking, and that is exactly why so much promised automation has not arrived.

Related reading: Anthropic's Claude Haiku 5.5 Is Its Cheapest Small Model Yet · SpaceXAI Ships Grok 4.7 for Coding at Grok 4.6 Prices, With Vendor Benchmarks to Match.

The company's answer is a model called Jev, the first of a class it calls System One Models. Jev does not write sentences. It takes unstructured state as input and returns typed, structured values with calibrated confidence scores attached to each one. In the company's own framing, it is a "frontier-intelligence function call": something a program can drop into a decision point where a hand-written rule would be too brittle and a language model would be too slow, too expensive, or too willing to make things up.

TypeSafe was founded in 2024 by Diogo Almeida, a former OpenAI researcher who co-invented the reinforcement-learning methods behind InstructGPT and, by extension, ChatGPT. His co-founders are Erik Gafni and Sasha Sheng. The seed round was led by DCVC, and the company says Jev is available now in early access, with developers pulled off a waitlist as capacity allows.

Why a model would give up strings

The pitch rests on a distinction that is easy to miss. A conventional large language model returns text, and almost everything it does downstream depends on someone parsing that text back into a decision. Its training rewards answers that human raters prefer, or answers whose correctness can be checked automatically. Neither objective produces a model that is honest about how sure it is.

Almeida's team argues that this is the bottleneck. "If a model can do a task 95% of the time but doesn't say when it's in the 5%, it can't automate that task," the company wrote in its launch post. A system that can be right most of the time and cannot flag the exceptions is not a component you can wire into an automated workflow, because the failures are invisible until something downstream breaks.

TypeSafe's training method is called Reinforcement Learning for Calibrated Decisions, or RLCD. Where RLHF optimizes for human preference, RLCD optimizes for decisions that come with epistemically honest probabilities. The model's outputs are constrained by a schema defined in advance, so the possible values and their structure are fixed before the model runs. TypeSafe says that makes type errors impossible, and that a model which never generates free-form text cannot hallucinate text.

The second half of the design is the sampler. Ordinary language models generate one token at a time, each conditioned on the last, which is why a hard request can take seconds and why cost scales with how much the model writes. Jev generates all of its outputs in a single query, in parallel. The company says a System One response lands in 70 to 500 milliseconds, against the 3 to 329 seconds it measures for frontier chat models, and it prices input at $0.042 per million tokens with output tokens free.

On its homepage, TypeSafe reports gains of up to 193.6 times on speed and 444.6 times on cost reduction for the kinds of workloads it targets, while noting those figures sit at the optimistic end of what it expects customers to see.

The benchmark is the weak point

The interesting part of the launch is not the speed claim. It is how carefully the company hedges its intelligence claim, and how quickly outside reviewers have found the seam.

TypeSafe published a set of workflow evaluations covering 711 cases across four jobs: security incident triage, observability over agent traces, invoice processing, and customer service. Instead of scoring against a hand-labeled ground truth, it treats the average judgment of two frontier models, GPT-6 Astra and Fable 5.1, as the reference answer. Against that reference, Jev owns the efficiency frontier by a wide margin, often responding in a fraction of a second for a fraction of a cent.

The accuracy picture is more mixed. On TypeSafe's own dashboard, Jev's aggregate agreement with the reference policy is about 67.8%, below the 74.1% of the best visible comparator. Its strongest showing is customer service, at 76.0%, close to the best model and above another. Its weakest is invoice processing, where it reaches 61.8% against 79.1% for the top comparator, the kind of gap that would matter in production. A detailed review published the same day by the analysis site Kingy concluded that the public evidence "supports the systems story" around typing, batching, latency and cost, but "does not yet support the strongest intelligence story," in part because the reference labels are model-generated rather than verified.

TypeSafe does not hide these caveats. It flags that the workflows were built by its own capabilities team, that averaging Astra and Fable biases the reference toward OpenAI and Anthropic, and that its headline multipliers come from a workload-specific test rather than a universal measure. It also concedes that some comparisons are not flattering, including one where a reviewer found Jev missing a writing defect that a frontier model caught.

The counterweight is a small independent test. Mike Taylor of the newsletter Every ran 37 documents and 777 judgments through Jev, which returned results in under 0.7 seconds at a cost of roughly a quarter of a cent. In a head-to-head on defect detection, Jev found six of seven intended flaws where Fable 5.1 found all seven, while running far faster and cheaper. That is the shape of the trade the company is selling: not better than the smartest model, but good enough on narrow decisions at a price and speed that make high-volume automation viable.

Why confidence is the whole argument

The calibration claim deserves more attention than the speed charts, because it is the part that changes what software can do.

An agent that has to decide whether to issue a refund, flag a transaction, or route a support case needs two things from a model: an answer and an honest estimate of how much that answer can be trusted. Today's chat models tend to supply the first and invent the second. Ask one for a confidence level and it will usually return something plausible and unmoored from its actual error rate. The consequence is that teams building automation end up bolting on human review for every decision, which erases much of the savings they were chasing.

TypeSafe is betting that a model trained to produce calibrated probabilities can clear that bar, letting software act on high-confidence answers and escalate the rest. That is a testable promise, and it is also a fragile one: calibration measured on a vendor's own workflows says little about how the model behaves when the inputs shift. The public material does not include the tail latency, rate limits, or retry behavior a production team would want, and the confidence curves that would substantiate the central claim are not published in full.

There is a broader reason the calibration question is landing now. The past year has made clear that capable models acting with autonomy can fail in ways that are expensive, and that the failures are often confidence problems rather than capability problems. Systems that can say "I do not know" clearly, and mean it, are the ones that get to operate without a person watching every step.

What Jev is not

It is worth being precise about the limits, because the framing invites over-reading. Jev cannot write a summary, draft an email, or explain its reasoning in prose. It is not a smaller or cheaper chatbot, and the company does not claim it should replace one. Any workflow that needs to produce words still needs a separate generator, which means the true cost of a finished task includes a large model on top of Jev's cheap decisions. The company's own demos lean on the surrounding code to constrain the model's freedom, which is honest about where the reliability comes from.

What Jev is, if the numbers hold, is a fast and inexpensive decision primitive for narrow, high-volume questions that software already has to answer: is this message abusive, does this record match that entity, should this request be forwarded or rejected. Those are unglamorous jobs, and there are a great many of them.

The fun demos, and the serious ones

To make the point concrete, TypeSafe shipped demos that are meant to be remembered rather than deployed. One drives a bot through Doom, feeding it structured game state and asking for decisions at about ten queries per second, a workload the company says costs roughly $7 an hour. Another plays wikiracing, navigating from one Wikipedia page to another using only links, a task where a single wrong choice derails the run and where Jev's inability to hallucinate a link is an advantage.

The intended commercial uses are more mundane and more valuable: classifying and routing support tickets, scoring and extracting fields, guarding against jailbreaks, and verifying the inputs or outputs of other models. These are the fuzzy if-statements inside software that teams have been approximating with brittle rules or with prompts to a large model. Jev's bet is that a small, fast, schema-bound model is a better fit than a general one.

That bet is crowded. OpenAI, Anthropic and Google have all pushed structured outputs, and the past year has brought a wave of small and specialist models aimed at cheap inference. What TypeSafe is arguing is subtler: that calibrated confidence, not raw capability, is the thing that decides whether a model can be trusted to act without a human in the loop. That is a claim customers can test, and the company is inviting them to.

What to watch

Early access will be the real test, because the workflow evals are vendor-run. Watch whether independent teams publish ground-truth comparisons that hold up the cost and speed claims, and whether Jev's confidence scores stay calibrated when the input drifts away from the demos. Watch, too, whether the accuracy gap closes on harder tasks such as invoice processing, where the model currently trails.

The name is a hint about the strategy. Jev is short for Jevons, after the nineteenth-century economist who observed that making a resource more efficient tends to increase how much of it is used, not less. TypeSafe's thesis is that intelligence has been too expensive to be sprinkled through ordinary software, and that a hundredfold drop in cost turns it into a building block rather than a destination. If that happens, the most useful AI in a company may end up being the model nobody chats with.