Wow! Computer Use (CU) Models Are Ready for the Enterprise!

We started to create a platform to help enterprise customers rapidly build GUI (Graphical User Interface) Agents using the CU (Computer Use) capability that OpenAI and Anthropic announced in the summer of 2025.

Until January 2026, the CU capabilities were just not good enough for us to solve any meaningful problems. The primary problem was accuracy - the GUI agents barely worked. Even simple things like “open notepad and write x” worked only occasionally.

Fast forward to April, 2026, and the Opus/Sonnet models were multiple orders of magnitude better. Agents became 20x more accurate - they completed accurately almost always. We were able to iterate on our GUIDE Build product so that we could build an agent in minutes. Our GUIDE Harness enhancements allowed us to increase performance 2-3x by specifying app specific skills, understanding the best prompts, etc. But it was still expensive - $1/run.

Then GPT 5.6 Luna was launched and the cost problem was solved. The #1 question in our conversations with enterprise customers is token cost (and ROI). When a run cost $1, then customers would do the math and say if the agent was run a thousand times a day, then the agent would cost $1000/day. With Luna, that same run now costs less than 1 cent. Now this is for 1 agent, but we are seeing similar results for other agents we are building.

With GPT 5.6 Luna, CU models are ready for the enterprise from a basic capabilities perspective. We have been adding a bunch of enterprise capabilities - observability, governance, sovereignty, scalability, and ubiquitous execution. I will share more in the coming weeks.

Also, it’s a race! Opus was our best performing model by price/performance for 3 months, but now Luna is the best. Stay tuned for more race updates!