GPT-5.6 Sol tops MMEnterprise-Bench at 68%, multi-image at 27.5%

GPT-5.6 Sol, the strongest model tested, answers 68.0% of questions on a new enterprise-agent benchmark when given 50 retrieved records, and 52.6% when given 10. That is the top line across ten proprietary and open-source multimodal LLMs put through MMEnterprise-Bench, a 1,613-question suite introduced in a paper posted to Hugging Face.

The benchmark grounds its questions in 30 simulated collaboration projects built from 12 research papers and 18 government publications, with 7,140 retrieval records (5,725 textual, 1,415 visual) stretched across Slack messages, Notion entries, email, meeting notes, PDF pages, slides and standalone images. Another 287 record versions preserve the temporal evolution of the workspace. The point is messiness: the same information gets shared, corrected and revised.

Two gaps are stark. On questions requiring a single text record, GPT-5.6 Sol hits 83.3%; on multiple image records it falls to 27.5%, and to 25.6% on mixed text and image records. Conflict resolution is lower still: the top model lands at 36.3% on cross-source conflicts and 32.5% on temporal conflicts. The paper reports that 'even the strongest models struggle when evidence contains cross-source or temporal conflicts or spans multiple images.'

Throwing more retrieval at the problem does not close the gap. 'Increasing the retrieval budget or employing a larger retriever improves evidence recall but does not consistently translate into higher answer accuracy,' the authors write. Gemini 3.8 Flash, meanwhile, beats GPT-5.6 Sol on recognizing unanswerable questions, 87.4% to 65.5%. The paper joins a steady run of agent-evaluation work our agents tracker has logged.