TypeSafe AI debuts model for machines that plays Doom

A launch rewrite. The piece is built on a company statement, a company demo and vendor-supplied figures, and it hedges accordingly: TypeSafe claims LLMs have variable response times that can stretch into minutes, and Jev's response time supposedly ranges from 70 ms to 500 ms, which it puts at 40x to 200x faster. No third party ran the model here, so treat every number as a claim with a source attached.

Structured game state, not pixels. The article says Jev can play Doom when fed structured data describing the player's game state, which sidesteps perception entirely. That makes the demo evidence about decision latency on a constrained action set, not about a model that perceives a game world, and it is the same shape as the support-routing example the article leads with instead.

The article raises that objection itself. Because the output is structured, framing it as hallucination-free is not a fair comparison, and structured output still does not preclude an incorrect answer. Type safety removes the parsing and validating step you would otherwise need on text responses; it says nothing about whether the chosen option was the right one.

The argument is efficiency, not capability. The piece concedes you can wrap natural-language responses in code to produce a constrained set of answers, then makes the case on cost and speed: a chat model predicts tokens sequentially while Jev evaluates independent questions in parallel, and the vendor positions it for high-volume business workflows such as sorting customer service problems rather than for conversation.

The report says the startup raised $40 million and that co-founder and CEO Diogo Almeida is a former OpenAI researcher and a co-inventor of RLHF and ChatGPT. Jev relies on an architecture TypeSafe calls Reinforcement Learning for Calibrated Decisions.

$0.042 per million input tokens with output free for Jev, against $2.00 input and $12 output per million tokens for GPT-5.6 Terra in the article's comparison.

The article cites a Doom demo at 0.114 s for Jev versus 8.566 s for GPT-5.6 Terra, and a general Jev response range of 70 ms to 500 ms, which TypeSafe puts at 40x to 200x faster than LLMs.