System1 Models

What is System1 Models?

System1 Models hosts decision models behind an API that accepts TypeSafe Jev's request format. Instead of generating text, a model answers typed questions about a state you send: pick one option from a list (choice), give the probability of yes (noul) or rate on a scale you define (score), with a probability for every allowed answer. Output tokens are free because nothing is generated.

It is run by productivity-boost.com Betriebs UG (haftungsbeschränkt) & Co. KG in Passau, Germany. The models are third-party rather than trained by System1. The company states it is not affiliated with TypeSafe or OpenAI.

Models

• s1-fast (Plumb-4B): short text, one question per request.
• s1-pro: up to three questions per request. On the EU tier it runs Surogate Rune 26B-A4B v3, with Winnow-12B Q8 as the fallback when EU capacity is busy.
• s1-vision: one question about one PNG, JPEG or WebP image, with Gemma 4 12B as the image fallback.
• s1-llm-auto-router: a CPU classifier that labels a prompt's category, difficulty and stakes, for routing between LLMs. It reads the first 512 tokens.

Each request takes up to 4,096 input tokens, counting the state, the questions and any image. Longer requests are rejected with HTTP 400 and not billed, rather than cut off. The S1-Model response header names the model that served each request, which tells you when a fallback answered.

Two tiers

The EU tier runs on GPU servers in Finland (Verda, plus UpCloud for overflow since 6 October 2026), with the API gateway and database at Hetzner in Helsinki. A DPA is accepted at signup. The Peer-to-Peer tier costs slightly less and starts on spare EU capacity. Only keys with "Allow worldwide processing" switched on may run on GPUs rented through Lium outside the EU; without that opt-in, Peer-to-Peer requests queue or get rejected when EU capacity is busy. The tier is set per key or per request with the S1-Region header.

Request content stays in memory for the duration of the request. It is never logged, written to disk or used for training; logs keep only metadata such as model, token count and latency. The ISO/IEC 27001 certificates belong to the hosting providers (Hetzner, Verda, UpCloud). Verda also holds SOC 2 Type II. System1 does not claim a certification of its own.

Pricing

Per 1M input tokens, output free, as of October 2026:

EUR rates are published too, from €0.028 for s1-fast. Credit is prepaid from EUR 1, with a free allowance by invitation. The prepaid balance is also the spend cap.

Integration

Requests go to POST /v1/systemone on api.system1models.ai, documented in an OpenAPI spec. A hosted MCP server lets an agent call a model and check its own usage and balance (decide, list_models, get_usage, get_balance).

Benchmarks

The benchmarks page collects results from five sources, comparing s1-fast and s1-pro with Jev. Two of them, JevBench and ImageJevBench, are run by System1's founder, which the page discloses. Run your own cases before relying on any of them.

Who it is for

Teams that want decision models with EU-only processing and a DPA, for slightly less than Jev's $0.042 per 1M input tokens. The trade-offs are the 4,096-token limit, one question per request on s1-fast and s1-vision, third-party models with fallbacks under load and a small company behind the service.