Key takeaways
- A System One model takes an input, a question and a fixed list of options, and returns a probability for each option in one pass, without writing text.
- TypeSafe coined the name for its Jev model after Kahneman's fast System 1 thinking, and Google launched Gemini Decision 1 on November 11, 2026 at $0.10 per million input tokens.
- In the SO-Arena benchmark the best System One models scored 70.2 against 79.7 for GPT-6 Luna and answered within 0.3 seconds, but the most accurate one ranked only seventh on reliability.
A System One model is an AI model that makes a decision instead of writing an answer. You give it some input, a question and a fixed list of possible answers, and it returns a probability for each answer in a single pass, usually in well under a second. The term comes from TypeSafe, the startup behind a model called Jev, and on November 11, 2026 a team of researchers published SO-Arena, a benchmark of how good and how reliable these models are. The best ones came within about ten points of a top reasoning model, answered tens of times faster, and were not the most reliable.
Where the name comes from
The name borrows from Daniel Kahneman's book Thinking, Fast and Slow, which splits human thought into fast, intuitive System 1 and slow, deliberate System 2. TypeSafe founder Diogo Almeida cites the book in the post that introduced Jev, dated September 15, 2026. He admits the catch: in Kahneman's work, System 1 is the part that makes mistakes. TypeSafe argues its models can be more reliable than the alternatives.
The other half of the analogy is the reasoning model, such as OpenAI's GPT-6 Sol and Luna, which writes out reasoning tokens before answering (see how reasoning models work). That makes it better at math and code but slower and more expensive. The SO-Arena paper calls this System Two.
How a System One model works
In the SO-Arena paper's definition, each task has three parts: the supplied information (text, records, passages or a history of actions), the decision to make, and the allowed outcomes. The model returns a probability for each outcome, from a decision head trained on top of a language model or from the language model's own scores for each answer. It never writes free text, so the answer is always one of the options. TypeSafe says Jev "never makes type errors", a guarantee about format, not about being right.
Three question types from TypeSafe's API docs have become standard, and Google's API uses the same names:
| Type | What it asks | What it returns |
|---|---|---|
choice |
Pick one option from a list | The option, a probability for each, a confidence |
noul |
A yes or no question | A number from 0 (no) to 1 (yes) |
score |
Rate on an ordered scale | A level, with probabilities and a confidence |
Typical jobs are routing a support ticket, choosing an agent's next tool, checking whether a task was finished, or flagging a policy violation. TypeSafe says Jev uses a new architecture with a parallel sampler and a training method it calls Reinforcement Learning for Calibrated Decisions (RLCD), aimed at confidence scores that match how often the model is right. The post does not describe either in detail.
Who makes System One models
TypeSafe opened Jev in early access with its September 2026 post. It lists input at $0.042 per million tokens, output free, and claims 70 to 500 milliseconds end to end.
Google launched Gemini Decision 1 through the Gemini Decisions API on November 11, 2026. Google doesn't say System One, but it is the same idea: options scored in a single forward pass, with a confidence from 0 to 1. As of November 11 it costs $0.10 per million input tokens and nothing for output, takes up to 1,000 questions per request, reads text, images, audio, video and PDFs, and has a free tier in Google AI Studio. Google claims 150 to 250 milliseconds end to end.
OpenAI has a Decisions API, which the SO-Arena paper cites as an October 2026 release and tests as "OpenAI Decisions".
Open-weight models are the largest group: Fastino's GLiDE, PelaAI's KnowLine-31B, Perplexity's Decider, Cloudflare's Clef, H2O.ai's H2O Lightning, Jared Palmer's Kev and others, from 0.4 billion to 31 billion parameters. Latent Space reported that Glean uses System One models alongside LLMs and that UiPath tested Jev for ranking document chunks.
What SO-Arena found
SO-Arena, from a team led by Xinchao Wang of the National University of Singapore, has 2,919 tasks across 11 domains, from sentiment and safety to math, code, tool use and web browsing. It tests 16 open-weight System One models, two closed decision APIs, StartLux Auto (which can switch thinking on), and GPT-6 Sol and Luna as System Two references. The score counts each domain equally and uses its hardest tasks.
| Model | Maker | Type | Score (0-100) |
|---|---|---|---|
| GPT-6 Sol | OpenAI | Reasoning | 82.3 |
| GPT-6 Luna | OpenAI | Reasoning | 79.7 |
| GLiDE (27B) | Fastino | System One | 70.2 |
| KnowLine-31B | PelaAI | System One | 70.2 |
| Sieve-31B-v3 | Sthanika AI | System One | 67.7 |
| Decider-27B | Perplexity | System One | 66.7 |
| Kev-27B | Jared Palmer | System One | 61.4 |
| Jev API 1.14 | TypeSafe | System One | 60.4 |
| OpenAI Decisions | OpenAI | System One | 53.6 |
| Laya (0.4B) | ConvAI Innovations | System One | 30.9 |
Speed is the trade. Every open-weight System One model answered within 0.3 seconds per task on one Nvidia H100 GPU. GLiDE averaged 113 milliseconds, while GPT-6 Luna averaged 2.7 seconds.
The gap depends on the job. On knowledge, sentiment and search ranking, System One models scored about as well as Luna. On routing, safety, task verification, tool use and web tasks, GLiDE came within 4 points of Luna, while the best 4B model trailed by up to 29. On math, science and code, both fell well behind. Thinking helped most there: StartLux Auto's math score rose from 50.8 to 92.5 with thinking always on, but in automatic mode it averaged about 3 seconds per task.
Accuracy is not reliability. The team checked whether a model keeps its answer when options are shuffled, and whether it notices when the right answer is removed. 14 of 18 System One models scored above 0.8 out of 1 on shuffling. Spotting a missing answer ranged from near chance to about 0.8. The two most reliable models, JevAny-27B and TypeSafe's Jev API 1.14, ranked only eighth and sixth of the 18 on accuracy, and the most accurate System One model ranked seventh on reliability.
Limits to keep in mind
- Google's Gemini Decision 1 came out the same day as the paper and is not in it.
- Difficulty tiers are set within each domain, and one leaderboard domain rests on only 20 tasks.
- Reliability measures consistency, not correctness. A model can be consistently wrong.
- Closed APIs were timed over the network and open models on a local GPU, so their speeds don't compare directly.
- Jev's and Google's prices and speeds are the companies' own figures.
The code is on GitHub under the MIT license, and the project page has the leaderboard.
Why it matters
Many AI calls inside software are small decisions: which tool, which queue, is this done, is this allowed. Running a reasoning model for each one is slow. System One models answer in about a tenth of a second at prices from $0.042 per million input tokens, and they return a probability you can act on, such as sending low-confidence cases to a person.
SO-Arena's result suggests a practical split: use a System One model for routing, ranking and checks, and keep a reasoning model for anything that needs several steps of math or code.
Frequently asked questions
What is the difference between a System One model and an LLM?
A regular LLM writes its answer one token at a time and can say anything. A System One model only scores a fixed list of options in a single pass, so it is much faster and its answer is always one of the allowed outputs, with a probability attached.
Are there open-source System One models?
Yes. The SO-Arena benchmark tests 16 open-weight ones, including Fastino's GLiDE, PelaAI's KnowLine-31B, Perplexity's Decider-27B and Cloudflare's Clef, from 0.4 billion to 31 billion parameters.
Which System One models score best on benchmarks?
In SO-Arena, GLiDE and KnowLine-31B tied for the top System One score at 70.2 out of 100, ahead of Sieve-31B-v3 at 67.7 and Decider-27B at 66.7. On reliability, JevAny-27B and TypeSafe's Jev API 1.14 ranked first and second.
What System One models are there besides Jev?
Google's Gemini Decision 1, OpenAI's Decisions API, and open-weight models such as GLiDE, KnowLine, Sieve, Decider, Clef, Kev and H2O Lightning all return probabilities over fixed options the way Jev does.
Sources
- SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One ModelsarXiv (Li, Dong, Feng, Li, Wang)
- SO-Arena full textarXiv
- SO-Arena project pageSO-Arena authors
- SO-Arena codeGitHub
- Introducing System One Models & JevTypeSafe
- Primitives (questions)TypeSafe
- Gemini Decisions APIGoogle
- The Jev Revolution: What Engineers Are Already Building with System One ModelsLatent Space


