TypeSafe AI is an AI company that came out of stealth on September 15, 2026 with Jev, the first model in a class it calls System One Models — models built to return structured, typed decisions that software can consume directly, rather than the strings that language models generate. The company was founded by Diogo Almeida, who states that at OpenAI he "helped build the methods that made language models useful at following instructions and talking with people" and that "that work ended up as the research behind ChatGPT" (Source: typesafe.ai).
Everything recorded below is drawn from the company's own launch post and site. No independent evaluation of its claims has been located, and the page is rated confidence: low on that basis; the company's own stated qualifications are carried alongside its claims.
Infobox
| Field | Value |
|---|---|
| Founder | Diogo Almeida (previously OpenAI) |
| Emerged from stealth | September 15, 2026, after a stated two years in stealth |
| First model | Jev, in early access at launch |
| Model class | "System One Models" |
| Training method | Reinforcement Learning for Calibrated Decisions (RLCD) |
| Stated pricing | $0.042 per million input tokens; output tokens free |
System One Models and Jev
The company describes System One Models as a class distinguished from existing language models on four axes it sets out in a comparison table: the optimization target, the input emphasis, the output type, and the sampling method. Where the company places existing models as optimized by reinforcement learning from human feedback or with verifiable rewards, for human preference or programmatically checkable outputs, it states that RLCD optimizes for "calibrated decisions: answers with epistemically honest probabilities on System One tasks." Where existing models take sequential messages and emit strings that must be parsed and validated, it states that System One Models take structured program state and emit type-safe structured values whose possible outputs and structure are defined in advance, accompanied by calibrated probabilities and confidence scores. And where existing models generate one token at a time, it states that Jev "generates all outputs in a single query" in parallel (Source: typesafe.ai).
The company states that the model gives up string generation and in exchange "can't hallucinate," describing Jev as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." It states that Jev reaches intelligence on System One tasks comparable to existing language models while being two orders of magnitude faster and more efficient, gives end-to-end response times of 70 to 500 milliseconds against a cited range of 3 to 329 seconds for frontier models, and supports output cardinality up to 255, above which it applies a two-stage scoring-then-choice system (Source: typesafe.ai).
The company states that its stack comprises a new model architecture, a parallel sampler, and the RLCD training method, and that it built the stack specifically for automation rather than for conversation. It attributes the class name to Daniel Kahneman's distinction between fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning, and the model name to the economist William Stanley Jevons, stating that it expects "machine intelligence to follow a similar path to coal, after steam-engine efficiency led to an increase in demand" (Source: typesafe.ai).
Evidence offered, and its stated limits
The launch post presents three categories of evidence and appends the company's own qualifications to each, which are recorded here because they bound what the claims establish.
On speed and cost, the company states the figures are verifiable but notes that its published evaluations are "generally run from our laptops on the West Coast," where its service is based, and that it cannot prove its pricing is not subsidized. On type errors, it states that a single counter-example would falsify the claim but that schema matching is guaranteed and type errors are therefore "mathematically impossible"; it also states that its plotted hallucination figure "is not empirical" and is entered as 0% on that guarantee, and that the comparison figures for language models are drawn from OpenRouter, where "more complex queries might be routed to better models" (Source: typesafe.ai).
On capability, the company describes a workflow evaluation in which every model is given the same fixed compute graph and scored against reference probabilities taken from the average of GPT-6 Astra and Fable 5.1, rather than against a ground-truth label. It states that this biases the reference toward OpenAI's and Anthropic's models and that it therefore likely underestimates its own model's and DeepSeek's relative performance; that the four published workflows were "not deliberately chosen nor constructed to make our model look good" and are outside its training distribution, but were made by its own model-capabilities team, "so some bias could exist"; and that the "193.6x faster, 444.6x cheaper" figures on its home page come from these evaluations and are "on the higher end of real world gains." In its side-by-side demonstration against GPT-5.6 Terra, it states that the query is "highly simplified" and that the relatively short input "paints our model in an advantageous light" (Source: typesafe.ai).
Products and availability
Jev was opened to early access on September 15, 2026, with developers being taken off a waitlist. The company publishes documentation at docs.typesafe.ai, a console with a playground, a workflow-evaluation site at evals.typesafe.ai, and an open-source adapter that constrains existing language models to emit structured decisions compatible with its API, which it uses for its own comparisons (Source: typesafe.ai).
Open questions
- Whether any party outside the company has reproduced the speed, cost or accuracy comparisons.
- The company's funding, headcount, headquarters and investor base, none of which the launch post states.
- Whether the "cannot hallucinate" claim, which the company grounds in schema conformance, extends to the correctness of the values returned within a valid schema — the post distinguishes type errors from wrong answers but does not report an error rate for the latter.
- Whether Jev receives a separate model page; it is documented here for now because the company's public material treats the model and the model class together.
Relationships
- instance-of: Agentic AI — the company positions its models as decision components inside automated software rather than as conversational assistants
- related: OpenAI — the founder's prior employer and the origin he claims for the instruction-following research behind ChatGPT
- contradicts: Sycophancy and Hallucination — the company asserts that constraining outputs to a predeclared schema removes a class of hallucination; the claim is the company's own and is untested here
- related: Post-Training — RLCD is presented as an alternative post-training objective to RLHF and RLVR
- related: Enterprise AI Deployment Gap — the company frames the gap between chat capability and deployed automation as its founding question