What is System1 Models?

What System1 Models is, and what it is not

System1 Models is a German company’s service that hosts open-weight decision models. You send a state and a typed question. You get back an answer with a probability for each option, the same contract Jev made popular.

It is not Jev. System1 says so on its own front page: it is a separate service, it does not resell Jev, and it is not affiliated with TypeSafe AI. The models that answer are different models, so the answers can differ from Jev’s on the same input.

s1-fast

Plumb-4B. Short text, one question per call, 4,096 input tokens.

s1-pro

Surogate Rune 26B-A4B v3 on the EU tier, with Winnow-12B Q8 as the fallback. Up to three questions per call, 32,000 input tokens.

s1-vision

The same Rune model with one image attached. One question per call, 32,000 input tokens.

s1-llm-auto-router

A CPU classifier that returns category, difficulty and stakes, so your own router can pick an LLM. It reads at most 512 tokens.

Source: System1’s model list. A request over the input limit gets HTTP 400. It is not truncated and it is not billed.

System1 Models pricing next to Jev's list price

Every System1 model bills input tokens only. Output is free, as it is on Jev. These are the launch rates on System1’s pricing page, in US dollars per million input tokens, before tax.

The image rate applies to every input token of a request that contains an image, text included. System1 gives the reason: an image takes about three times the GPU time per token. It says an image usually counts as 300 to 450 tokens.

System1 also measured the cost of 1,000 decisions on medium-length text with three questions per call: $0.0087 on Peer-to-Peer, $0.0093 on the EU tier and $0.0109 on Jev. Read the Peer-to-Peer figure with its footnote. It covers successful calls only, and in that test 14 of 30 Peer-to-Peer calls from Finland succeeded.

There is no subscription. You top up prepaid credit from EUR 1. For what the Jev side of that table buys in practice, read Jev pricing.

System1 Models latency and benchmark scores, as System1 publishes them

System1 timed its own API against Jev on 2 October 2026, from Nuremberg, with 30 measured calls per service on the same synthetic support tickets and one question per call.

Both System1 models have a lower median than Jev in that run. s1-pro has a higher p95, so its slow calls are slower. System1 calls the sample small and says results change with workload and network.

On quality, the model behind s1-pro scores close to Jev on two public boards. Decision Index v0.2.1 gives Jev 57.91 and Surogate Rune 57.44. JevBench Capability Score v1.5.5 gives Jev 80.01 and Rune 79.02. Plumb-4B, the model behind s1-fast, scores 71.65 on the same JevBench board.

Two notes from System1’s benchmark page belong next to those scores. The scores are for the open models, not for the hosted System1 service, which has not been scored yet. And JevBench is run by System1’s founder. System1 prints both facts itself.

The front page also says “up to ~40× cheaper” and “~18× faster”. Those figures compare a decision model with a 27B reasoning LLM, not with Jev. They are JevBench estimates, not measurements of the service.

System1 Models tiers: Peer-to-Peer or EU

The tier decides where your request runs. You set it per API key, or per request with the S1-Region header.

EU

Inference runs in Helsinki, Finland, on contracted EU operators. Inference data does not leave the EU. EU requests get priority. A Data Processing Agreement comes with signup.

Peer-to-Peer

Starts on spare EU capacity. Lower price. When it is busy, a request can be queued or refused with a retryable HTTP 503.

Worldwide opt-in

A switch on each Peer-to-Peer key. With it on, a busy request can run on GPUs rented through Lium, from independent providers in several countries, some outside the EU.

Both tiers

Prompts and outputs are not stored and not used for training. Usage records such as token counts are kept for billing.

The name needs one clarification. Peer-to-Peer rents its extra capacity on demand today. The open network where anyone contributes compute is System1’s stated plan, and System1 says it carries no customer traffic yet.

System1 also warns that the worldwide switch is not a legal basis for a data transfer. Do not send personal data on a key with the switch on unless you have that basis. The trust page and the sub-processor list name who handles what.

How to move a Jev call to System1 Models

The body of the request keeps its shape: a state and a map of questions, each of type noul, choice or score. If you have made a first Jev call, this one will look familiar.

curl https://api.system1models.ai/v1/systemone \
  -H "Authorization: Bearer $SYSTEM1_API_KEY" \
  -H "Content-Type: application/json" \
  -H "S1-Region: eu" \
  -d '{
    "model": "s1-fast",
    "state": {
      "message": "I need to update my invoice address.",
      "account_type": "business"
    },
    "questions": {
      "route": {
        "type": "choice",
        "instructions": "Choose the support queue that best fits the request.",
        "criteria": {"billing": null, "technical": null, "account": null}
      }
    }
  }'

{
  "id": "dec_route",
  "model": "s1-fast",
  "answers": {
    "route": {
      "type": "choice",
      "choice": "billing",
      "confidence": 0.87,
      "probabilities": {"billing": 0.87, "technical": 0.09, "account": 0.04}
    }
  },
  "usage": {"input_tokens": 351, "output_tokens": 0, "decisions": 1},
  "tier": "eu"
}

The response is the illustrative one in System1’s quickstart, which has the same call in seven languages. Four things change when you come from Jev, per System1’s migration guide:

1. Base URL and key. Requests go to api.system1models.ai with a System1 key. A TypeSafe key does not work there.
2. Model IDs. Use s1-fast, s1-pro or s1-vision. GET /v1/models returns the live list.
3. Questions per call. Three on s1-pro, one on the others. One question too many returns HTTP 400 request_limit_exceeded. The service does not split the call for you.
4. Score rubrics. A score question needs at least two levels in its criteria array.

So the work is small if your calls ask one to three questions, and larger if you batch many questions into one Jev request. Batched calls must be split, and each split call sends the state again.

System1 Models over MCP, in Claude Code and Cursor

System1 hosts an MCP server, so an agent can ask for a typed decision while it works. It has four tools: decide, list_models, get_usage and get_balance.

claude mcp add --transport http system1models https://api.system1models.ai/mcp \
  --header "Authorization: Bearer $SYSTEM1_API_KEY"

System1’s MCP page has the config for Cursor and Claude Desktop. For the servers that do the same job on Jev, read Jev MCP servers.

Where System1 Models is the wrong choice

These limits are all on System1’s own pages. They decide the fit more than the price does.

• Large batches. A call with more than three questions does not fit any System1 model.
• Long context. s1-pro stops at 32,000 input tokens and s1-fast at 4,096.
• Work that cannot retry. Peer-to-Peer can refuse a request when it is busy. Three-question calls on that tier are best-effort. Use the EU tier where a refusal is a problem.
• Images with no fallback. On the EU tier, s1-vision returns a retryable 503 if its model is down. There is no second model behind it.
• Exact repeatability. s1-fast runs on a batched engine, so two identical requests can return slightly different probabilities.
• Hard reasoning. System1 says a decision model scores below a frontier LLM on accuracy, and tells you to test your own cases first. For where that line sits on Jev, read Jev vs an LLM.

Who System1 Models fits

• Teams that must keep data in the EU. The EU tier names its location and its operators, and comes with a DPA.
• Decisions about a picture. Jev takes text only, so an image needs another model in front of it. The ways to do that are in Jev and images. s1-vision takes the image in the same call.
• High volumes of short text. Ticket routing, lead scoring and escalation checks ask one question about a short message, which is what s1-fast is priced for.
• Anyone who wants a second provider. The request shape is shared, so a one-question call can run on either service.

To weigh it against the other services and the models you can run yourself, read Jev alternatives and open-source Jev.

Common questions

Is System1 Models the same as Jev?

No. System1 Models is a separate service that hosts open-weight decision models behind a Jev-compatible API. It does not resell Jev and is not affiliated with TypeSafe AI. Keys, balances, model IDs and limits are its own.

Is System1 Models cheaper than Jev?

On list price, yes. System1 publishes $0.032 to $0.040 per million input tokens for its text models, with output free. Jev's list price is $0.042 per million input tokens, with output free. A request that contains an image costs more: $0.217 to $0.228 per million input tokens.

Is System1 Models a drop-in replacement for Jev?

Not fully. The request shape is shared: a state and typed questions of type noul, choice or score. The limits are different. s1-pro takes at most three questions per call, and s1-fast and s1-vision take one. System1 says it does not claim byte-for-byte compatibility.

Where does System1 Models run?

The EU tier runs inference in Helsinki, Finland, on contracted EU operators. The Peer-to-Peer tier starts on spare EU capacity. It leaves the EU only for API keys where you turned on "Allow worldwide processing".

Can System1 Models read images?

Yes, with s1-vision. It takes one PNG, JPEG or WebP image per request and one question. Jev itself takes text only.

Does System1 Models have a free tier?

There is a free allowance, but System1 gives it by invitation. Without an invitation you top up prepaid credit, from EUR 1.