System One models and Jev: intelligence as a function call
TypeSafe AI announced its first System One model, Jev, on 2026-09-15: no string generation, only type-safe structured values with calibrated probabilities. The official post prices input at $0.042/MTok, output free, with 70ms-500ms end-to-end latency. Here is what it actually is, what the vendor itself evidences, and how to do the same class of work today given that QCode does not offer it.
Updated 2026-09-20
Four checkable numbers
input price (vendor list)
The blog's price line: input $0.042/MTok ($42 per billion tokens), output free because it is "too cheap to meter". That is TypeSafe's own number, not a QCode tariff.
end-to-end latency (vendor claim)
The same page puts frontier LLMs at 3 to 329 seconds and claims 40x-200x speedup; the homepage's 193.6x faster / 444.6x cheaper figures are also vendor numbers, and the post itself says those sit at the high end of real-world gains.
parallel sampling, not token by token
Per the post: existing models generate one token at a time conditioned on the last; System One generates all outputs in a single query with a parallel sampler, collapsing a structured decision into one call.
type errors (by construction)
The set of possible outputs and their shape are fixed before the call, so the post says the model cannot make type errors - while also noting that 0% in that chart is not measured but guaranteed by construction.
What a "System One model" is supposed to be
TypeSafe's definition: a class of models built for fast, structured decisions software can use directly, trained with RLCD (Reinforcement Learning for Calibrated Decisions) rather than the RLHF / RLVR line. The vendor's comparison table keeps unstructured text in - weighted toward program state - but forces every output to be a type-safe structured value carrying a calibrated probability, sampled in parallel. The trade-off is stated plainly: Jev gives up string generation, so it is not a chat or code-writing model. The first public model, Jev, entered early access on 2026-09-15.
Release cadence and where access stands
The 2026-09-15 post announced Jev with same-day early access and shipped a workflow-evaluation site (examples, disagreements, full queries). It also states the published evals were mostly run "from our laptops on the West Coast" and that the pricing cannot be proven unsubsidised - time will show whether it is sustainable. As of this page's check (2026-09-20), access is still TypeSafe's own waitlist / early access, not a third-party API catalog.
Timeline
2026-09-15: TypeSafe AI publishes "Introducing System One Models & Jev", signed by founder Diogo Almeida; the first public model Jev enters early access.
The same 2026-09-15 post sets out pricing and latency: input $0.042/MTok, free output, 70ms-500ms end to end, against a table putting frontier LLMs at $0.20-$10/MTok and 3-329 seconds. The docs add the operating parameters: model id `jev-1.13.0`, one endpoint `POST /v1/systemone` selected by the model field, 64k context per request (32k for state plus the longest question), rate limits of 250,000 tokens/second and 1,200 requests/minute with 429 over either, text-only input.
Check date for this page (2026-09-20): searching QCode's public model catalog for Jev / TypeSafe / System One returns no match, so nothing on this page claims the model is usable here.
Confirmed vs claimed
Verifiable yourself
The 2026-09-15 date, the author, the price and latency figures, the name of the training method (RLCD), each row of the vendor comparison table, and the three caveats the vendor writes about itself (laptop-run evals, LLM numbers sourced from OpenRouter, the 193.6x/444.6x multipliers likely being high-end) are all in the blog post and can be read directly.
Not yet safe to state as fact
As of 2026-09-20, "Cannot hallucinate", "two orders of magnitude faster", "444.6x cheaper" are vendor claims: no independent re-run is published, and the reference answer in the workflow evals is the average of GPT-6 Astra and Claude Fable 5.1, which the post itself notes biases toward OpenAI and Anthropic models. Enterprise readiness, quotas, SLA, open weights and region availability are not described, so this page does not infer them.
Jev vs doing structured output with a general model
Jev (per the vendor)
Outputs are type-safe structured values with calibrated probabilities, produced by one parallel sampling pass; claimed 70ms-500ms, $0.042/MTok in ($42/Btok in the docs), free out, one endpoint, 64k context, 250k tokens/s plus 1,200 req/min limits, text-only input. The cost: it never emits strings, so chat and code generation are out of scope.
General frontier models (including tiers sold on QCode)
The same question can be answered with JSON schema or tool calling in a mature ecosystem, with long context and prose you can log; the price is autoregressive latency plus per-output-token billing, confidence has to be self-reported or proxied, and zero type errors is not guaranteed.
Doing the "smart if-statement" today on QCode
Decompose the task into several small independently answerable questions, have the model return enums and scores as JSON via schema or tool calling, and let the surrounding code validate types, apply thresholds and retry. Structurally that mirrors TypeSafe's own shape - one endpoint plus a declared schema - except here the schema goes to a general model via tool calling or response_format. Price-sensitive paths can push classify/route work to the cheapest tiers in the family and keep hard reasoning on the top tier; when you need a "can I trust this call" signal, approximate it with agreement across samples or a cheap judge call rather than treating a self-reported confidence as calibrated. Current tiers and prices come from /pricing and the live catalog.
The actual QCode situation
QCode does not offer Jev, nor any TypeSafe model: the 2026-09-20 check of the public model catalog returns no match for Jev / TypeSafe / System One. This page is not a purchase path and does not quote vendor prices as ours. For the structured-decision work described above, the available route on QCode is a general model plus schema/tool calling with external validation, billed at official price times our service rate; see /pricing and the model catalog for what is on the shelf now.
Questions
Is Jev a chatbot?
No. By the blog's own definition it only emits pre-declared structured values with probabilities and does not generate strings, so it cannot serve as a conversation or writing model.
Why does it claim to be faster and cheaper?
Two reasons: one parallel sampling pass instead of token-by-token generation, and giving up string output. The stated figures are 70ms-500ms, $0.042/MTok in, free out - and the post itself says it cannot prove the price is unsubsidised.
Should I believe "it cannot hallucinate"?
In the vendor's framing that means no type errors: the value set is fixed before the call, which is guaranteed by construction. Whether the decision content is right and the probabilities calibrated rests on TypeSafe's own workflow evals, with no independent replication yet.
Can I call it through QCode?
No. The 2026-09-20 catalog check has no such model and QCode provides no TypeSafe service. The equivalent work is done here with a general model plus structured output.
Is it the same thing as a small tier like GPT-5.6 Mini or Nano?
No. Small tiers are still autoregressive general models whose output is a string (constrainable to JSON); Jev is a different design: parallel sampling, typed outputs, calibrated probabilities, and no text generation.
Are there rate limits?
Yes: the docs state 250,000 tokens/second and 1,200 requests/minute, with 429 Too Many Requests over either. The model id is `jev-1.13.0` and everything goes through `POST /v1/systemone`. Access remains early access / waitlist; general availability is not specified.
Sources
TypeSafe AI blog, "Introducing System One Models & Jev", 2026-09-15, https://typesafe.ai/blog/introducing-system-one-models-and-jev (fetched for this page 2026-09-20); price, model id, endpoint, context and rate limits come from https://docs.typesafe.ai/models.md (same day). Prices, latency, speed/cost multipliers and the workflow-eval methodology all come from that page, including the vendor's own caveats: laptop-run evals, LLM figures taken from OpenRouter, and the 193.6x/444.6x claims likely sitting at the high end.
No "Jev ready now" promise
QCode does not offer Jev. To ship the same class of structured decision, use a general model we do sell plus schema/tool calling, billed at official price times our service rate.
Read next
GPT-5.6 family guide
How the general models we do sell handle tiers, context and structured output.
Claude Haiku 4.5 guide
A low-latency small tier, useful for splitting off classify/route work.
AI agent eval readiness
Building reviewable evaluations for smart if-statements instead of one demo.
Not affiliated with TypeSafe AI. Every price, speed and benchmark figure here restates the vendor's public material and was not re-run by us; QCode does not offer Jev and nothing on this page is an availability promise.