Official results · AILuminate Safety 1.2.1 · English (US) · published October 29, 2026

MLCommons began running the Safety 1.2.1 benchmark on September 22, 2026. As of October 29, 2026, this document reports Official results for 27 systems, and the run is ongoing. Each system has an overall percentile score and a score for each of the 12 hazard categories, all on a scale from 0 to 100 calibrated against a fixed set of reference systems. The composition of the reference set is fixed before a benchmark version is released, so new AI systems may be added to the results page without shifting the percentile scores of previously tested systems. Higher scores mean a lower tendency to give violating responses. 

Citing these results. “[System] scored in the [Nth] percentile overall on AILuminate Safety 1.2.1, English (US).[1]” [1] Official result retrieved from mlcommons.org on [date]. The AILuminate name and logo are trademarks of MLCommons Association. All rights reserved. Unauthorized use is strictly prohibited. See www.mlcommons.org for more information. All citation rules  AILuminate Benchmark Use & Messaging Policy

How to read these scores

• A score places a system on the Safety 1.2.1 scale and describes its measured behavior within the benchmark’s scope: single-turn, English-language, text-to-text requests in 12 hazard categories.
• The overall score weights all 12 categories equally. A hazard-category score shows where a system falls short.
• A score is not a certification, and it does not set a threshold for acceptable safety.
• Only results labeled AILuminate Safety 1.2.1, English (US), compare directly with these scores; grades from Safety 1.0 are not comparable.

Results

Detailed reports for all 27 systems give each system’s percentile score in each of the 12 hazard categories, with a citation ready to copy.

Open-weight models

These are standalone models with no external guardrails or other moderation logic filtering prompts or responses. They are typically presented as a single network of model weights.

Guardrail systems

These are safety classifiers that check the prompts sent to an AI system, the responses it returns, or both, against a hazard policy. They are typically deployed alongside one or more models as part of a complete system, rather than used to answer users directly.

Complete systems

These are systems that may include one or more models, guardrails, algorithmic filters, and other moderation techniques. They are typically presented through an API or as a set of software components. Preserving the integrity of our benchmarks is of the utmost importance to MLCommons. These initial results are primarily for open models we can run locally on our infrastructure. Results for a broader range of models and systems will be available soon.

Systems selected but not published

These are systems that MLCommons selected for benchmarking, but the system providers have chosen to opt out. To date no systems have opted out of being included in this benchmark’s public results.

How MLCommons ran this round

• Release notes: The Safety 1.2 benchmark introduces three new components: a new sample of test prompts (Problem Set 2.0), a new automated evaluator (the system that judges whether each response violates the Assessment Standard), and a new scoring method that reports percentile scores and grades against a fixed panel of reference systems rather than a single reference system. Safety 1.2 release notes are available here and 1.2.1 release notes are available here. 
• Notice: MLCommons made a good-faith effort to notify each provider before benchmarking, as section 3.3 of the AILuminate Benchmark Use & Messaging Policy requires.
• System release dates: The release date shown for each system is the month in which the tested version first became publicly available, as downloadable weights or through a hosted service, whichever came first. For an alias that always points to a provider’s newest model, the date is that of the model the alias served during testing. MLCommons took each date from public sources such as the provider’s announcement, model card, or model repository.
• Technical labels: AILuminate Safety 1.2.1, Problem Set 2.0 official prompts, evaluator v1.4, English (US).

Visit our FAQ for the most common questions about the AILuminate benchmark.