TypeSafe AI's Jev answers with choices, never sentences

Ask Jev a question and it will not write you a sentence back: it picks one option from a list you defined in advance and staples a confidence score to it. TypeSafe AI came out of two years of stealth with the model on 15 September, in a post from founder Diogo Almeida, and as Testingcatalog reports, output tokens cost nothing.

At a glance

• TypeSafe AI has opened early access to Jev by pulling developers off a waitlist, calling it the first of what the company labels System One Models, a fast decision layer for software.
• Input tokens cost $0.042 per million, output tokens are free, end-to-end replies land between 70 and 500 milliseconds, and TypeSafe claims 40 to 200 times the speed on comparable queries.
• The four published workflows came from TypeSafe's own model-capabilities team and were scored against the average probabilities of GPT-6 Astra and Fable 5.1 rather than ground-truth labels.

If you missed the earlier chapters: TypeSafe AI was founded in 2024 by Almeida, who according to SiliconANGLE worked on reinforcement learning from human feedback, InstructGPT, ChatGPT and GPT-4 at OpenAI, alongside co-founders Erik Gafni and Sasha Sheng. The same outlet reports a $40 million seed round led by DCVC, and says Forbes put the company's valuation at $200 million, citing a person familiar with the transaction. Its founding thesis, per SiliconANGLE, is that the traits making chat models pleasant company — plausible invention, varying methods between requests, confident uncertainty — turn against you inside production software.

Input runs $0.042 per million tokens and output is free

TypeSafe prices Jev at $0.042 per million input tokens, which it also states as $42 per billion, and charges nothing at all for output. End-to-end response times run 70 to 500 milliseconds, and the company says Jev can be 40 to 200 times faster on similarly intelligent System One-shaped queries.

The homepage examples go further: 193.6 times on speed and 444.6 times on cost, numbers TypeSafe says likely sit near the upper end of real-world results. The company also says long-term pricing sustainability remains unproven. According to SiliconANGLE, TypeSafe's website lists 39 cents per 1,000 workflows for Jev against $3.31 for OpenAI's GPT-5.6 Luna.

Up to 255 choices go straight through, and one demo is Doom

Jev takes unstructured context, including text and structured program state, and returns probabilistic decisions with confidence scores inside a schema the developer fixes beforehand. It handles up to 255 choices directly; for larger sets, one stage scores candidates independently and a second selects among them.

TypeSafe says schema matching is guaranteed and type errors are mathematically impossible. The claim that Jev cannot hallucinate covers only that constrained output space, and does not mean every factual decision comes out right.

The listed uses are the plumbing of most applications: classification, routing, scoring, extraction, branching, large-dataset processing, real-time applications and guardrails on LLM output. The demos are odder — Doom, driven by structured game state and text rather than images, and Wikiracing.

Why does every answer arrive with a confidence score?

Because that is what the training method optimises for. According to MindStudio, RLCD, or Reinforcement Learning for Calibrated Decisions, trains a model to pair a decision with a confidence score reflecting how likely the decision is to be correct, where RLHF trains a model to produce answers human raters approve of.

Almeida's public criticism of RLHF, per the same source, is that it rewards how convincing an answer looks to a rater in the moment rather than whether it holds up when acted on, which produces mode dropping and confidence disconnected from accuracy. Jev combines a new architecture, a parallel sampler and RLCD.

Instead of emitting tokens one after another, the model evaluates the decision in parallel. Think of an essay exam versus a multiple-choice sheet: the essay has to be composed word by word, the sheet only has to be scored. TypeSafe's documentation describes three question types, Noul, Choice and Score, with confidence between 0 and 1 derived from the underlying probability distribution.

The comparisons were built and run by TypeSafe

The workflow evaluations use identical compute graphs and compare outputs against the average probabilities from GPT-6 Astra and Fable 5.1, not ground-truth labels. Four workflows are published, created by the company's own model-capabilities team, which leaves room for selection bias.

The LLM baseline runs through TypeSafe's System One LLM wrapper, which constrains conventional models to compatible structured decisions and may affect speed and cost. Most tests were run from laptops on the US West Coast, where the service is based. TypeSafe notes its Wikiracing comparisons mostly use non-reasoning LLM modes, making those rivals look weaker than they might with reasoning enabled. SiliconANGLE reports the company's near-194-times speed and roughly 445-times cost figures have not been independently verified and will vary by workload, network location and comparison method.

So the headline multiples come from a harness the vendor built, judged against what two other models believed on average, measured from machines sitting close to the vendor's own servers. In our view the more telling disclosure is the price one: a service that bills nothing for output while admitting its long-term pricing sustainability is unproven is signalling that $0.042 per million is a bet rather than a floor.

When the waitlist clears

Early access is opening now as TypeSafe brings developers off the waitlist, and that is as far as the published schedule goes: no general availability date, no commitment past the current rate. The numbers worth waiting for are runs from outside the US West Coast, workflows picked by someone other than TypeSafe's own team, and Wikiracing comparisons with reasoning modes switched on.