Which Decision Model to Choose: A Practical Guide
Analyze top decision models by classification accuracy, tool routing, epistemic awareness, open weights, and data residency.
👁 115 reads
Selecting a decision model requires balancing task complexity, latency limits, licensing, and compliance requirements. Artificial Analysis data from October 13, 2026, combined with producer technical specifications, outlines distinct trade-offs across current architectures.
Moderation and Tool Routing
For input moderation, safety screening, and classification, Fastino Labs’ GLiDE achieved the top WildGuardTest mark with an F1 of 95.1. Perplexity’s Decider v1.1 27B and TypeSafe’s Jev 1.13 followed closely at 93.8 and 93.6 respectively. For latency-sensitive filtering, Decider v1.1 costs $12.1 per million WildGuardTest tasks compared to GLiDE’s $84.3.
On NVIDIA’s When2Call tool-invocation suite, GLiDE and Celeris-1 Decision tied for the highest score at 83.7. Celeris-1 Decision offers a median latency of 63 ms and supports an optional one-sentence explanation field at no added charge, though it lacks open weights or fine-tuning capabilities. GLiDE uses dynamic reasoning passes that raise its median task time to 217 ms.
Knowledge Boundaries and Calibrated Declines
Knowledge-heavy tasks expose systematic weaknesses in decision models. Liquid AI’s d1 posted the highest knowledge accuracy at 41.0 percent on AA-Omniscience-Decision, but chose ‘I don’t know’ on only 1.7 percent of incorrect answers, indicating substantial overconfidence. Organizations requiring models to decline unknown queries see better epistemic calibration from Microsoft-Decision-1 (39.8 percent ‘I don’t know’ rate), Jev 1.13 (35.2 percent), and Mercury Decide (32.0 percent).
For structured multi-step reasoning, Jev 1.13 achieved 20.9 on HLE-MultiChoice, leading all competing decision models, none of which exceeded 15.4. However, Jev 1.13 showed weak non-English ticket routing accuracy in a test on 50 customer-support tickets in a non-English language (Polish) - a small indicative sample - scoring 41 out of 47 unambiguous cases versus 44 for Microsoft-Decision-1, Decider, Mercury Decide, and Clef.
Deployment, Licensing, and Compliance
Infrastructure constraints often determine model feasibility:
- Decider v1.1 27B: Open weights under Apache-2.0, requiring ~49 GiB GPU memory; supports multimodal inputs and batch queries up to 128 questions.
- d1: Open weights under LFM Open License v1.0 (free commercial use below $10M revenue); edge deployment reaches 8 ms on an RTX 4090 and 50 ms on a Jetson Orin Nano.
- Microsoft-Decision-1: Post-trained Qwen3.5-9B available via Microsoft Foundry under GlobalStandard or DataZoneStandard tiers, guaranteeing data processing stays within specific zones such as the EU.
- Jev 1.13: Closed-source system hosted in the United States with standard contractual clauses, lacking dedicated EU processing regions.
- GPT-6 Luna Decisions: OpenAI beta requiring dedicated JSON schema formats, with a 10 percent surcharge for EU/EEA workloads.
For general efficiency on hosted cloud APIs, Decider v1.1 27B, Microsoft-Decision-1, and Jev 1.13 define the cost-quality Pareto frontier on the Decision Index v1.0.
More: AI decision models - Decision Index ranking, prices and tests
More on this from QUANTSLANT
magazine·Decision Model Pricing and Cost BenchmarksAn overview of token rates, benchmark task expenditures, and empirical support routing costs for leading decision models.