Power BI DAXBench v2 shows stronger AI DAX
SQLBI’s webinar explains DAXBench v2, where AI models are tested on real Power BI DAX tasks and the top scores stay short of perfect.
Key insights
- DAXBench v2 uses prepared Power BI models and separate scores from v1.
- The webinar features Marco Russo and Maxim Anatsko.
- The benchmark evaluates execution and correctness, not just whether code looks plausible.
- The current leaderboard puts GPT-6 Astra Auto first with 91 of 94 tasks solved.
- The benchmark covers tasks such as aggregation, time intelligence, filtering, iterators, and budget allocation.
SQLBI is hosting a webinar titled Inside DAXBench v2: Results and Your Questions, where Marco Russo and Maxim Anatsko explain the benchmark, discuss its results, and answer questions live. The session is scheduled for Wednesday, November 4, 2026, and the recording will remain on the channel after the live event. The main takeaway is simple: DAXBench v2 checks whether AI models can write DAX that works correctly in realistic Power BI scenarios, not just DAX that looks convincing.
What's new
DAXBench is described as a public leaderboard for testing AI models on DAX tasks. Version 2 introduces a new set of tasks that run against prepared Power BI models, and its scores are separate from version 1. That separation matters because it keeps the newer results from being mixed with the older benchmark.
The benchmark is meant to answer a practical question: which models can reliably produce correct DAX in real reporting situations? The research notes say v2 tests 74 models across 94 DAX tasks and validates generated code against Power BI data rather than relying only on human judgment. It also gives additional credit for solutions that are efficient, readable, and aligned with common DAX patterns.
Why it matters
This is most relevant to Power BI developers, analysts, and makers who are considering AI assistance for DAX authoring. DAX is the expression language used in Power BI and related Microsoft data-platform technologies, and small mistakes in filter context, relationships, or context transition can produce wrong results without any syntax error. A benchmark that checks actual results is more useful than one that only checks whether code runs.
The material also shows why leaderboard results should be read carefully. The top entries are strong, but none of the tested models solves every task. In reporting work, a single incorrect measure can affect decisions, so even a few failures still matter when the task is complex or business-critical.
How it works
The benchmark uses the Contoso sample dataset and asks models to solve a set of practical DAX tasks. The task mix includes basic and advanced patterns, which helps expose where models cope well and where they fail. According to the research notes, the categories include aggregation, time intelligence, filtering and filter context, calculation groups, table manipulation, iterators, context transition, advanced business patterns, ABC classification, new and returning customer analysis, and budget allocation.
The evaluation is based on two checks. First, the DAX must execute successfully against the model. Second, it must return the correct result. That second step is important because a measure can run without errors and still be logically wrong. The benchmark also awards extra credit for solutions that are efficient and readable, which encourages answers that are practical for Power BI work rather than just technically valid.
The current leaderboard cited in the research notes places GPT-6 Astra Auto in first place with 91 of 94 tasks solved and a score of 95.91%. Claude Opus 5.5 Auto follows with 90 of 94 tasks solved and 95.38%, and GPT-5.6 SolMED is third with 87 of 94 tasks solved and 91.46%.
Limitations and open questions
The webinar page does not provide the full technical methodology in the description itself, only a high-level summary and links to the leaderboard, a blog post, and an article on how DAXBench works. The material also does not spell out every scoring rule or every test case in the video description.
The benchmark is useful, but it is still a benchmark. The research notes say that performance can vary with the semantic model, prompt quality, metadata given to the model, and the complexity of the business requirement. That means a strong leaderboard result is not a guarantee that the same model will work equally well in every Power BI environment.
We also could not verify from the provided material how the live Q&A was handled, whether the video contains extra demonstrations beyond the benchmark overview, or whether any model rankings changed after the notes were compiled. For those details, the linked sources are the place to check.
Sources
Frequently asked questions
What is DAXBench v2 measuring?
It measures how well AI models can solve DAX tasks for Power BI models. The benchmark checks both whether the code runs and whether it returns the correct result.
Why does version 2 matter compared with version 1?
Version 2 uses a new set of tasks against prepared Power BI models, and its scores are separate from v1. That keeps the two result sets from being mixed together.
What kinds of DAX tasks are included?
The research notes mention aggregation, time intelligence, filtering, calculation groups, iterators, context transition, ABC classification, customer analysis, and budget allocation.
Can I treat the leaderboard as a guarantee?
No. The notes say results can vary with the model, prompt quality, metadata, and the business scenario. The leaderboard is a useful comparison point, not a promise.