Everything below is labeled: vendor claim = said by TypeSafe, unverified; observed = demonstrably true; analysis = this site's reasoning. No invented benchmarks.
| LLM + function calling | LLM + JSON mode | System One Model (Jev) | |
|---|---|---|---|
| What it returns | Arguments for a function call (text-as-JSON) | A JSON string it wrote | Probabilities over your options — no generated text observed shape |
| Passes | Autoregressive token-by-token | Autoregressive | Single pass vendor claim |
| Can it hallucinate? | Yes — may emit invalid args/enum values | Yes — malformed or invented fields | Vendor claims no: output is a distribution over options you defined vendor claim |
| Probabilities | Not calibrated (token likelihoods ≠ truth) | Not calibrated | Calibrated by training (RLCD) vendor claim |
| Best at | Open-ended multi-step tool use | Extracting structured fields from text | High-volume closed-set decisions: routing, moderation, scoring, triage analysis |
| Weak at | Cost & latency at scale; silent failures | Same + format drift | Anything needing prose or novel output by design |
Read as: not a replacement — a new slot in the stack. Keep System 2 for reasoning steps; put System One in the hot loop. analysis
Can Jev write text?
No. It gives up string generation entirely — that's the design trade that buys speed and the no-hallucination property. vendor claim
Who's behind it?
TypeSafe AI, founded by Diogo Almeida — previously at OpenAI, where he worked on instruction-following research that ended up in ChatGPT. Two years in stealth before the Sep 15, 2026 announcement. observed
How do I get access?
Waitlist/early access at TypeSafe, and it's listed on Cloudflare Workers AI as typesafe/jev. observed
What does it cost?
No public pricing as of 2026-09-20. Use the calculator to model scenarios with your own assumed price.
Are the 100× speed and "can't hallucinate" claims verified?
Not independently, as of this page's date. The launch post says "extraordinary claims require extraordinary evidence — see the receipts" and includes evaluations; independent replication hasn't shown up yet. Treat numbers accordingly. analysis
Is there an open-source version?
Only toy-scale research starters (see ecosystem below). Nothing Jev-equivalent is open as of 2026-09-20. observed
Is this site official?
No — independent and unaffiliated. See the footer on every page.
| Project | What it is | Status (2026-09-20) |
|---|---|---|
| TypeSafe AI / Jev | The first commercial System One Model | Early access; HN launch post ~1.9k points |
| jevlike | Open-source (MIT) "starter model with the same input/output shape" — picks among N text options in one pass. Independent; TypeSafe's design unpublished | Research demo scale (~1k stars): demos are Doom button-picking & chess move choice; NOT a Jev substitute — no intelligence claims made by its own README |
| jev-model-router-demo | Community demo routing coding tasks to different models | Early demo |
| Respawn / Skillsync / Radio | Adjacent AI-agent tooling shipped the same week | Multi-builder cluster signal, not System One models themselves |
This table is updated as the category evolves. Last check: 2026-09-20.
typesafe/jev; response nesting quirk observed in real calls 2026-09-19.Corrections welcome: [email protected]. Anything we got wrong, we fix with a dated note.