This Open-Weight AI Agent Controls Your Computer? Holo4 vs Claude Opus 5.5 & ChatGpt 6.1 Sol

A local Holo4 agent, Claude Code, and Codex get the same six real computer tasks. Saved work and independent checks decide the result. CHAPTERS 0:00 Six real computer tasks 1:12 The local setup 2:34 The scoring contract 4:00 Files: the archive must be where you asked 4:55 Spreadsheet: formulas that actually update 5:50 Document: the unsaved window is not the handoff 6:45 Round four — GUI only 7:33 Blender: a scene file must produce the requested image 8:23 App repair: two correct edits can introduce another bug 9:10 The results of this local trial 10:03 Scope and infrastructure 10:40 Repeat the check 11:30 Keep the evidence CHECK THE HANDOFF ☐ Pin the setup: Save the exact checkpoint, quantization, model identifier, and agent harness. ☐ Freeze the request: Keep identical prompts and input hashes; use a fresh workspace for every attempt. ☐ Record the work: Capture genuine native prompt entry and private GUI actions. Keep infrastructure retries. ☐ Save before inspection: Copy the produced files at the time limit. Check a separate immutable copy. ☐ Open and test: Recalculate formulas, reload stored state, open the scene, and inspect the real export. ☐ Report the scope: Publish the criteria and artifacts. State failures and harness differences alongside the score. SOURCES H Company Holo4 announcement: https://hcompany.ai/newsroom/holo4 Holo4 35B-A3B model card: https://huggingface.co/Hcompany/Holo4... Official GGUF checkpoint: https://huggingface.co/Hcompany/Holo4... Holo agent conventions: https://hub.hcompany.ai/models-api/bu...