TypeSafe Jev: a revolutionary approach to AI decision-making?
TypeSafe introduces Jev. Speed and pricing figures come solely from the vendor. Independent measurements are lacking.
With Jev, TypeSafe offers a model that supplies programs with a decision and its probability. Running software can immediately process this value and take the next step. The vendor cites 70 to 500 milliseconds end to end.
Key takeaways
- TypeSafe introduced Jev on 15 September 2026 as a model for machine decisions. Almeida published the launch post.
- Jev turns unstructured state into typed output. Programs read the signal and take the next step.
- Customer service intents can be routed to code, a language model or people. The documentation lists order_status, product_question and complaint.
- TypeSafe supplies its own latency and token pricing figures. The vendor ran the tests using laptops on the US West Coast.
Related:Serving infrastructure: how multi-agent systems make IT operations manageable / OpenAI pulls the plug on Cursor: your own key as Plan B
Jev delivers typed decisions
What is Jev? Jev is TypeSafe’s decision model. An application passes in its current state and the permitted responses. It receives probabilities or a yes/no value. The program can use these values to take the next step. It need not dissect a chat response or parse an essay.
TypeSafe introduced Jev on 15 September 2026 in its own launch post. Diogo Almeida leads the company and cites InstructGPT as part of his career. TypeSafe was founded in 2024 and then operated in stealth. The company raised 40 million dollars in a seed round. Almeida authored the post dated 15 September 2026 and speaks for TypeSafe there.
Jev accepts unstructured state and returns typed decisions with probabilities to running programs. Programs receive the decision as a processable signal for the next step in their own code. TypeSafe describes dispensing with text generation and presents Jev as a model for machine decisions.
The probability accompanies the decision, allowing software to set its own thresholds. TypeSafe targets workflows where programs need a decision and pass it on in typed form. Unstructured input becomes a fixed format that code can evaluate directly.
TypeSafe opened Jev for early access in September 2026 and says its training methods include RLCD, a method for calibrated decisions. Running software feeds the decision directly into existing workflows and processes the typed signal locally. Early access in September 2026 follows the company’s stealth phase.
An order query can go to ordinary program code
TypeSafe demonstrates intent routing as a customer service pattern, routing decisions to existing destinations. A simple order status query can go to ordinary program code, a product question to a language model and a complex or uncertain case to a person. TypeSafe’s documentation describes this pattern and names the corresponding intents.
The documentation assigns cases to the intents order_status, product_question and complaint, making the decision machine-readable. A customer service stack passes unstructured input to Jev and expects a typed response for the next step. Jev supplies downstream code with the decision and its probability. Downstream code executes the selected intent, keeping the workflow within its own stack.
Programs read the decision and branch according to the intent. The order_status intent stays within their own code, resolving the status query there. The product_question intent goes to a language model to retrieve product information. The complaint intent reaches a person for manual handling. TypeSafe describes this mapping in its intent routing pattern.
The figures come from the vendor
| Metric | Value according to TypeSafe |
|---|---|
| Response time | 70 to 500 milliseconds |
| Input price | Around 4 US cents per million tokens |
| Output | No separate charge, according to the vendor |
Source: Launch post, 15 September 2026. Also 17 September 2026. Independent measurements are lacking.
TypeSafe conducted the evaluations itself and published the results in its launch post. Measurements ran on laptops on the US West Coast, tying them to the vendor’s test environment. Pricing may be subsidised and should be viewed with that caveat.
Independent measurements are lacking. External laboratories have yet to test Jev’s speed and costs or publish their own latency measurements. The range comes solely from the vendor and carries that caveat. TypeSafe remains the source for latency and pricing.
TypeSafe is keeping early access open in September 2026, allowing teams to work with the vendor’s figures during this phase. Latency and pricing remain figures set by TypeSafe itself.
Frequently asked questions
What does Jev return to a running program?
The model delivers typed output plus a probability. The stack processes this signal within its own process.
How are support queries routed?
Status queries stay within the application’s own code. Catalogue questions use a language model. Unclear or difficult cases go to staff.
How reliable are the speed and pricing figures?
The range and token price come from TypeSafe. External laboratories have yet to confirm these figures.
Editor’s Picks
- Nvidia Acquires Hugging Face-Open Models Remain an Option
- Kubeflow: CNCF Declares AI Platform Mature
- OpenAI Buys Macs for AI Agents That Click Like Humans
More from the MBF Media Network
MyBusinessFutureBitkom Finds AI Widely Adopted-but Its Potential Still UntappedDigital ChiefsSAP Ties AI Agents to Cloud Migration and S/4HANA AdoptionSecurityTodayAttackers Access GitLab Files Without LoginImage Source: AI-Generated (September 2026)
Translated from the German original using artificial intelligence. The German version is authoritative.
Also available in
