Which System One model should I use?
Answer five questions and you get up to five models that fit, each with the numbers that put it there. The picker covers the 34 decision and extraction models on this site that are generally available or in early access, and it starts with hosted APIs. Choose "Weights I run myself" or "Either" in question 3 to bring in open-source models.
Your shortlist
The top 5 of 23 matching models.
Celeris DecisionCeleris, hosted API
System One Score 85.9 from 2 of 3 boards; #1 of 19 overall on the ranking, which also weighs price and access; $0.004 per 1M input tokens.
d1-ultraLiquid AI, hosted API
System One Score 95.9 from 2 of 3 boards; #2 of 19 overall on the ranking, which also weighs price and access; $0.042 per 1M input tokens.
GLiDEFastino Labs, hosted API
System One Score 99.7 from 3 of 3 boards; #3 of 19 overall on the ranking, which also weighs price and access; $0.04 per 1M input tokens.
Gemini Decision 1Google, hosted API
System One Score 91.5 from 2 of 3 boards; #4 of 19 overall on the ranking, which also weighs price and access; $0.10 per 1M input tokens.
mammoth-1Baseten, hosted API
System One Score 97.2 from 3 of 3 boards; #5 of 19 overall on the ranking, which also weighs price and access; $0.10 per 1M input tokens.
Quick answers
Cheapest hosted decision models
- Celeris Decision (Celeris)$0.004 per 1M input tokens; System One Score 85.9 from 2 of 3 boards; #1 of 19 overall on the ranking, which also weighs price and access.
- Span-02 (Respan)$0.01 per 1M input tokens; System One Score 80.9 from 2 of 3 boards; #15 of 19 overall on the ranking, which also weighs price and access.
- Cicero (Butler Compute)$0.01 per 1M input tokens; no independent board score yet.
Highest-scoring hosted decision models
- GLiDE (Fastino Labs)System One Score 99.7 from 3 of 3 boards; #3 of 19 overall on the ranking, which also weighs price and access; $0.04 per 1M input tokens.
- SubQ Snap (Subquadratic)System One Score 98.1 from 2 of 3 boards; #12 of 19 overall on the ranking, which also weighs price and access; $0.04 per 1M input tokens; early access.
- mammoth-1 (Baseten)System One Score 97.2 from 3 of 3 boards; #5 of 19 overall on the ranking, which also weighs price and access; $0.10 per 1M input tokens.
Best-calibrated hosted decision models
- Gemini Decision 1 (Google)Calibration error 0.0132 on the Artificial Analysis Decision Index; System One Score 91.5 from 2 of 3 boards; #4 of 19 overall on the ranking, which also weighs price and access; $0.10 per 1M input tokens.
- d1-ultra (Liquid AI)Calibration error 0.021 on the Artificial Analysis Decision Index; System One Score 95.9 from 2 of 3 boards; #2 of 19 overall on the ranking, which also weighs price and access; $0.042 per 1M input tokens.
- d1 (Liquid AI)Calibration error 0.0116 on the Jev Decision Index and 0.0585 on the Artificial Analysis Decision Index; System One Score 75.3 from 3 of 3 boards; #9 of 19 overall on the ranking, which also weighs price and access; $0.02 per 1M input tokens.
Best open-source models to run yourself
- H2O-Lightning-31B (H2O.ai)System One Score 97.1 from 2 of 3 boards; you run the weights, so no per-token price; licence: H2O Model Non-Commercial Licence v1.1.
- Jeffan-31b-1710 (Trong Vu)System One Score 95.5 from 2 of 3 boards; you run the weights, so no per-token price; licence: Apache-2.0.
- Quyet-1.0-Large (Quyet)System One Score 89.4 from 2 of 3 boards; you run the weights, so no per-token price; licence: Apache-2.0.
Hosted models that read images
- Celeris Decision (Celeris)System One Score 85.9 from 2 of 3 boards; #1 of 19 overall on the ranking, which also weighs price and access; $0.004 per 1M input tokens; reads images.
- d1-ultra (Liquid AI)System One Score 95.9 from 2 of 3 boards; #2 of 19 overall on the ranking, which also weighs price and access; $0.042 per 1M input tokens; reads images.
- GLiDE (Fastino Labs)System One Score 99.7 from 3 of 3 boards; #3 of 19 overall on the ranking, which also weighs price and access; $0.04 per 1M input tokens; reads images.
How the shortlist is ranked
Every filter and number comes from the model pages on this site and is read when the site is built. With no priority set, hosted decision models follow the ranking of decision models and extraction models come after them. Weights you run follow the System One Score, this site's average of a model's scaled results on the Artificial Analysis Decision Index, the Jev Decision Index and JevBench. A model measured by two or more of those boards ranks above one measured by a single board.
Lowest price sorts by the hosted price per million input tokens. Best calibration sorts by expected calibration error, the gap between how sure a model says it is and how often it is right. Errors are only compared within the board that measured them, so a model can sit above one whose error on another board is slightly lower. Longest context sorts by the largest input the model accepts. Announced and deprecated models never appear.
If a model lacks the data a filter needs, it is left out of that filter. If it lacks the data a priority sorts by, it goes to the end of the list. Either way the result names it. Image, audio, video and PDF input come from the input types each model page records from the maker's docs or model card. Some models read an input only once you turn it on with a header or server flag, and the result says so.
FAQ
Which System One model should I use?
It depends on what you send it and where it has to run. With the default answers (a hosted API and no other preference), the picker starts with Celeris Decision, d1-ultra and GLiDE, the highest-ranked hosted models on this site's ranking of decision models. Each board tests a fixed set of tasks, not yours, so try two or three on your own data before you commit.
What is the cheapest decision model?
Among hosted decision models, Celeris Decision at $0.004 per million input tokens. Open weights you run yourself have no per-token price: you pay for the hardware instead.
What are the alternatives to Jev?
21 other commercial decision models on this site have a hosted API, as Jev does. By System One Score, the highest are GLiDE, SubQ Snap and mammoth-1. Read the Jev alternatives guide.
Are there open-source System One models?
Yes. 21 models on this site publish weights you can download and run yourself, and 11 of them were released as open source. Of those, the highest by System One Score are H2O-Lightning-31B, Jeffan-31b-1710 and Quyet-1.0-Large. Pick "Weights I run myself" or "Either" in question 3 to include them. H2O-Lightning-31B and Torchcast Decision 27B have licences marked non-commercial, so read the terms before you ship.