TypeSafe AI Releases Jev as First Public System One Model

Jev is a non-LLM System One decision model from TypeSafe AI that accepts unstructured program state and typed questions to return structured outputs without hallucinations or type errors.

TypeSafe AI announced the early access availability of Jev on September 15, 2026. The model addresses limitations in existing frontier systems by prioritizing machine-usable decisions over text generation.

What background led to the introduction of System One Models?

Diogo Almeida, founder of TypeSafe AI, identified that prior optimization for human language created barriers for automation. Computers require structured data rather than natural language responses to function reliably in software pipelines.

The development occurred over two years in stealth with multiple research advancements. These efforts produced a category of models built specifically for fast, structured decisions that software systems can consume without further interpretation.

What capabilities set the Jev release apart from prior offerings?

Jev takes unstructured program state and typed questions as inputs. It produces outputs limited to Choice, Score, or Noul/Boolean formats that include calibrated probabilities and confidence scores.

Predefinition of output structures by the user removes the possibility of hallucinations or type mismatches. Results become immediately actionable in code without additional validation steps.

Availability extends through integrations with Vercel AI Gateway, Netlify AI Gateway, LiteLLM, and Cloudflare AI. These connections support broader developer adoption of the early access release.

What technical design choices enable Jev performance?

The architecture incorporates a parallel sampler and RLCD training method. All questions undergo simultaneous evaluation against the shared program state inside one API request.

Reported end-to-end response times fall between 70 milliseconds and 500 milliseconds. These latencies support time-sensitive decision loops in automated environments.

Performance metrics drawn from TypeSafe AI workflow evaluations for System One tasks.

Metric	Jev	Traditional LLMs
Speed	Up to 193.6x faster on System One tasks	Baseline sequential processing
Cost	444.6x cheaper at $42 per billion input tokens	Input prices up to 238x higher
Output Format	Predefined structures with probabilities	Unstructured text requiring post-processing
Error Handling	No hallucinations or type errors	Frequent need for correction
Query Mode	Parallel in single call	Typically one query per call

What market and stakeholder implications follow from Jev adoption?

Cost reductions and speed gains open automation opportunities in sectors constrained by LLM variability. Organizations can deploy decision logic at higher volumes with greater consistency.

Software teams gain tools that align with programmatic requirements through typed outputs. This alignment may expand AI use in backend services and multi-step agent workflows.

1. Predefined structures remove post-processing requirements in automation sequences.
2. Confidence scores allow direct incorporation into conditional code logic.
3. Parallel single-call evaluation lowers cumulative API latency and expense.
4. Token pricing supports sustained high-throughput decision workloads without proportional cost increases.

What reactions accompanied the announcement of the Jev model?

After two years in stealth, countless technical challenges, and research breakthroughs… I am beyond excited to announce that today, TypeSafe AI is releasing our first System One Model: a new class of frontier models built to make fast, structured decisions that software can use directly.Diogo Almeida, Founder, TypeSafe AI

The statement positions Jev as the initial entry in a distinct model class. Reports from integration partners such as Vercel confirm the reported efficiency metrics from internal testing.

What steps may follow the early access phase for System One Models?

TypeSafe AI has positioned Jev as the first public model in its System One line. Continued application of RLCD training and parallel sampling techniques is expected to inform subsequent releases.

Expanded integration support and additional model variants could extend the approach to wider automation scenarios based on the initial release framework.