Snorkel AI tested SpaceXAI's Grok 4.5 on 2,000 expert-graded workplace tasks — and the new model leads GPT 5.5 and Claude Opus 4.8 on mean pass rates across most sectors.
Snorkel AI has published Grok 4.5 benchmark results that put SpaceXAI's newest model through roughly 2,000 expert-graded professional work tasks — and the numbers show Grok 4.5 outperforming both GPT 5.5 and Claude Opus 4.8 on real workplace deliverables. Using Snorkel's GDPval+ dataset, part of the Snorkel Data Series (SDS), domain experts authored prompts and rubrics spanning legal documents, spreadsheets, presentations, and other outputs that mirror actual U.S. economy occupations. Each frontier model ran once per task through Harbor, Snorkel's open-source LLM and agent evaluation framework, with LLMaJ verifiers scoring outputs against expert-created criteria.
Grok 4.5 achieved a 29% mean pass rate on the evaluation sample, ahead of GPT 5.5 at 22% and Opus 4.8 at 21%. The gains clustered in work that demands deep professional judgment: legal tasks (40% versus 27–28% for rivals), education (58% versus 35–42%), healthcare (35% versus 23–25%), and QA analysis (37% versus 19–27%). GPT 5.5 still led in construction, and Opus 4.8 topped financial-manager scenarios, but across most occupations and sectors Grok 4.5 led — often by wide margins. At SpaceXAI's request, Grok 4.5 was paired with the Grok Build agent inside Harbor; GPT 5.5 and Opus 4.8 used a tuned Stirrup evaluation agent optimized for GDPVal-style tasks.
Failure analysis reinforced the headline. Grok 4.5 posted the lowest prevalence in all six tracked error categories. The most common failure mode — missing domain analysis — fell to 40% of samples versus 51–52% for competitors. Concrete examples include a wage-and-hour liability task where Grok 4.5 alone satisfied every expert criterion, and a commercial real estate assignment where it delivered differentiated site-level recommendations instead of generic guidance. Even so, expert-level work remains hard: no model passes more than one in three criteria, underscoring how much room remains on the professional AI frontier.
Read more: Snorkel AI — Grok 4.5 testing results.
Original Sources
Disclaimer
This is an AI-summarized article from authoritative sources. If you want your content removed, please .