Best AI decision models in October 2026: 13 ranked

Each model in one line

1. 1.

   pplx-decider91.8/100Best overallCheapest

   Perplexity's 27B model has the top System One Score, 97.4 from all three boards, and the highest Jev Decision Index Full score (62.75) and JevBench Capability Score (82.8) in the ranking; it costs $0.02 per million input tokens with free output, and its weights are on Hugging Face under Apache 2.0.

   Compare: Jev vs pplx-decider, Microsoft-Decision-1 vs pplx-decider

2. 2.

   Kev90.9/100Best calibrated

   Jared Palmer's open-weights family has a System One Score of 94.5 from two of the three boards, and Kev 27B's calibration error of 0.022 on the Jev Decision Index is the lowest of any ranked model there; Kev-4B is also hosted on OpenRouter at $0.042 per million input tokens.

3. 3.

   Jev84.7/100

   TypeSafe AI's Jev 1.13 has a System One Score of 96.3 from all three boards and the highest Artificial Analysis Decision Index score in the ranking, 57.5. It costs $0.042 per million input tokens with free output, reads a 64,000-token context, and defines the /v1/systemone API that Kev, Drex and CLM also serve.

   Compare: Jev vs BERT, Jev vs Claude, Jev vs Clef, Jev vs GLiNER, Jev vs GPT, Jev vs Laya, Jev vs Microsoft-Decision-1, Jev vs OpenAI Decisions API, Jev vs pplx-decider

4. 4.

   GLiDE84.1/100

   Fastino's GLiDE has a System One Score of 94.2 from all three boards, and its no-thinking checkpoint scores 60.21 on the Jev Decision Index, just above Jev's 60.11; it costs $0.15 per million input tokens, and the thinking tokens it reports are billed at $0.

5. 5.

   Microsoft-Decision-180/100

   Microsoft-Decision-1 has a System One Score of 89.5 from two of the three boards: 53.8 on the Artificial Analysis Decision Index and 70.78 on JevBench 1.7.31, from a native Azure Foundry run. It costs $0.042 per million input tokens through Microsoft Foundry, the same as Jev, and the weights are closed.

   Compare: Jev vs Microsoft-Decision-1, Microsoft-Decision-1 vs Clef, Microsoft-Decision-1 vs OpenAI Decisions API, Microsoft-Decision-1 vs pplx-decider

6. 6.

   Clef77.9/100

   Cloudflare's 27B Clef has a System One Score of 86.7 from all three boards and scores 53.08 on the Jev Decision Index with a 0.040 calibration error; it reads images as well as text and costs $0.24 per million input tokens on Workers AI, the highest hosted price in the ranking.

   Compare: Jev vs Clef, Microsoft-Decision-1 vs Clef

7. 7.

   d175.5/100

   Liquid AI's d1 has a System One Score of 70.4 from all three boards: the hosted d1 scores 50.3 on the Artificial Analysis Decision Index, and the open d1-3B scores 40.08 on the Jev Decision Index, though Liquid submitted that board's public-suite run itself. The hosted d1 costs $0.04 per million input tokens on OpenRouter.

8. 8.

   Clef-omni73.1/100

   Cloudflare's Clef-omni has a System One Score of 75.4 from all three boards, with 49.32 on the Jev Decision Index; it reads audio and video as well as text and images, and costs $0.15 per million input tokens on Workers AI, with Apache 2.0 weights.

9. 9.

   Celeris-1 Decision89.6/100

   Celeris-1 Decision has a System One Score of 91 from one board, the Artificial Analysis Decision Index alone, where it scores 52.3 and has the fastest median time per task of any ranked model, 63 ms; it costs $0.04 per million input tokens with free output, and the weights are closed.

10. 10.

    Mercury Decide82.7/100Cheapest

    Inception's Mercury Decide has a System One Score of 91.7 from one board, the Artificial Analysis Decision Index alone, where it scores 52.7; at $0.02 per million input tokens, half its regular $0.04, it ties pplx-decider as the cheapest ranked model, and a free tier stays on OpenRouter.

11. 11.

    Drex71.9/100

    Nace.AI's Drex 1.5 has a System One Score of 81 from one board, the Jev Decision Index alone, where it scores 50.84; it reads states up to 131,072 tokens and costs $0.05 per million input tokens, and its 9B weights carry a licence that limits commercial use.

12. 12.

    OpenAI Decisions API65.1/100

    OpenAI's Decisions API has a System One Score of 84.3 from one board, the Artificial Analysis Decision Index alone, where its GPT-6 Luna row scores 48.5; it has been in public beta since 6 October 2026 at $0.10 per million input tokens, with free output.

    Compare: Jev vs OpenAI Decisions API, Microsoft-Decision-1 vs OpenAI Decisions API

13. 13.

    Tev138.4/100

    Together AI's Tev1-4B-experimental has a System One Score of 54.4 from one board, the Jev Decision Index alone, where it scores 34.14; it answers Choice questions only, with 2 to 24 options, and is still early access at $0.042 per million input tokens.

Not yet independently measured (4)

These hosted models have no independent run on the Artificial Analysis Decision Index, the Jev Decision Index or JevBench, so the rubric cannot rank them without leaning on the vendors' own numbers, which it never does. They join the ranking when one of the boards measures them.

How the System One Score works

The System One Score is this site's combined quality score. It reads three independent boards, each with its own tests and its own scale:

• The Artificial Analysis Decision Index, run by Artificial Analysis on hosted APIs. The score reads its Index score, out of 100.
• The Jev Decision Index, a Hugging Face board run by multimodalart. The score reads its Full score, where 0 is random guessing and 100 is perfect. It counts 20% public benchmarks and 80% private tests.
• JevBench, run by Benchmark Heaven. The score reads its Capability Score.

How it is worked out:

1. On each board, take the model's independent result from its page. Vendor-run results are left out. When a page lists several editions, the newest is used; when it lists several sizes of one family, the first listed is the headline model.
2. Scale that result so the top result on the board is 100: 100 times the model's result over the top. The top is taken from every model on this site, hosted or self-hosted, so a model has the same score on every page that shows it.
3. Average the scaled results from the boards that measured the model, and round to one decimal. A model needs one board to get a score. The board count is shown beside every score.

Calibration is combined the same way. The Artificial Analysis Decision Index and the Jev Decision Index both publish an expected calibration error. On each, the ranked model with the lowest error gets 1 and the highest gets 0, in a straight line. The calibration factor is 15 times the average across the boards that measured the model. JevBench publishes no comparable error, so it does not count toward calibration.

Each ranked model, board by board

What the score cannot tell you

• The boards test different things, so a model can lead one and trail another. The score averages them; it does not settle which board is right.
• Half of the Artificial Analysis Decision Index is knowledge and reasoning tasks, not decision tasks alone.
• The Artificial Analysis Decision Index covers hosted APIs only. Open weights you run yourself can be scored on the other two boards at most.
• A score from one board is less certain than a score from three. Read the board count beside it.
• Scaling to the top result says how close a model is to the best one here on each board, not how good it is in any absolute sense.

How this ranking is scored

Six factors add up to 100. Each one reads a field on the model's page and applies the rule below, so two people running the rubric on the same pages get the same table. Totals are rounded to one decimal, and a tie keeps alphabetical order.

• Only models a production app can call by API are ranked: every commercial model, and open weights that a paid host serves. Open weights you run yourself are listed in the next section, by System One Score, and never ranked against the hosted models.
• Only an independent run on at least one board makes a model rankable. Quality is 40 times the model's System One Score over the top System One Score among ranked models.
• Calibration uses expected calibration error, the average gap between a model's stated confidence and how often it is right, from the Artificial Analysis Decision Index and the Jev Decision Index. It is combined across the two as described above.
• Access reads the status on the model's page. Generally available is 15; early access is 6; announced is 0.
• Price is per million input tokens. The cheapest model among those ranked gets 15, and the rest get 15 times that price over their own. A model with no published price gets 0.
• Primitives are the three question types a decision model can answer: Choice, Score and Noul. Three points each, plus one for all three.
• Five points for a published SDK or an API endpoint, none otherwise.

What the ranking cannot tell you

• Each board tests a fixed set of tasks. A model that scores well there can still miss on your data, and a low score does not rule a model out for one narrow job.
• Every vendor-run benchmark is left out, including rows a vendor submitted to a board. A model with only vendor numbers sits unranked, not scored low.
• Price without latency. A model that costs less per token but answers slower, or needs more tokens of state, can cost more per decision. Latency figures on the model pages are vendor-run.
• Calibration is measured across each board's task groups. A model can be well calibrated on average and poorly calibrated on the one task you care about.

Open-weights decision models you run yourself (8)

These models have no hosted API, so they are not ranked against the hosted models above: self-hosting changes the price, latency and access factors. They are listed by System One Score instead, on the same site-wide scale, with each board's scaled and published result; vendor-run numbers are left out. Open weights that a paid host also serves are ranked above as hosted models.

Best open weights Quyet-1.0-Large by Quyet, System One Score 96.1 from 2 of 3 boards.

1. 1.

   Quyet-1.0-LargeQuyet

   System One Score 96.1 from 2 of 3 boards: Jev Decision Index 93.6 (58.71 on 0.3.1, calibration error 0.0778) and JevBench 98.7 (81.7 on 1.7.46).

2. 2.

   Torchcast Decision 27BTorchcast AI

   System One Score 95.5 from 1 of 3 boards: Jev Decision Index 95.5 (59.91 on 0.3.1, calibration error 0.0235).

3. 3.

   H2O-Lightning-4BH2O.ai

   System One Score 90.6 from 1 of 3 boards: JevBench 90.6 (75 on 1.7.46).

4. 4.

   Surogate Rune 26B-A4BSurogate

   System One Score 90.1 from 2 of 3 boards: Jev Decision Index 91.5 (57.43 on 0.3.1, calibration error 0.1196) and JevBench 88.8 (73.5 on 1.7.46).

5. 5.

   Strands-Decider 2BStrands Labs

   System One Score 34.7 from 1 of 3 boards: Jev Decision Index 34.7 (21.75 on 0.3.1, calibration error 0.0441).

6. 6.

   LayaConvai Innovations

   System One Score 23.2 from 2 of 3 boards: Jev Decision Index 7.1 (4.43 on 0.3.1, calibration error 0.1402) and JevBench 39.3 (32.5 on 1.7.46).

7. 7.

   CLMContrastive-LM

   System One Score 17.1 from 2 of 3 boards: Jev Decision Index 9.9 (6.24 on 0.3.1, calibration error 0.3227) and JevBench 24.3 (20.1 on 1.7.46).

8. 8.

   GLiNER2.5-DecideFastino Labs

   System One Score 14.7 from 1 of 3 boards: Jev Decision Index 14.7 (9.24 on 0.3.1, calibration error 0.0884).

FAQ