AGI Was Achieved in February 2026 | Bernardo Aceituno

Open-weight models reach frontier as enterprises start switching

Open-weight models caught up with frontier models on enterprise tasks, a moment Aceituno dates to Claude Opus 4.6 in February. Cheap, controllable inference now runs agents inside real business processes. Stack AI sells into the back-office half of enterprise AI, where customers pick their own model, so the switch happened there first. The gap left is trust, since benchmarks are gamed and no one can say when a model is measurably good enough.

Takeaways

• Aceituno says AGI arrived in February 2026. Hundreds of open-weight models now beat Claude Opus 4.6, the release he dates as general intelligence for enterprise automation. Chinese labs and Meta have shipped open-weight models that score above frontier models. Cheap frontier-grade inference is the real impact.
• Enterprises adopt AI along two paths. Assistive AI for employees runs on tools including Copilot and ChatGPT Enterprise. Back-office automation, Stack AI's business, places agents inside persistent processes or turns a whole process into agents. Those systems let customers pick their own model, so open weights spread quickly.
• Token bills and data residency trigger switches. Enterprise token bills crossing seven figures raise alarms. Closed models run in a cloud and region where the enterprise cannot see what data is used. Military clients wanted full residency in their own data centre, trading off performance to get it.
• Benchmark trust has corroded. Over the past three months providers tuned open-weight models to score well on benchmarks, and enterprise trust in them corroded. His favourite is BiTES. Frontier labs keep the trust for state-of-the-art performance until better benchmarks and community consensus arrive.
• Memory and deep vision remain unsolved. Domains needing long-term memory or very deep visual understanding are still out of reach for open-weight and frontier models alike. He gives no timeline for closing that gap.
• He predicts 90% open-weight enterprise usage. Aceituno compares the shift to Linux against Mac and Windows. Enterprises will soon give employees a personal companion agent and will choose open weights for trusted performance and control over hosting and cost. OpenClaw is the first open-weight personalised agent he names.
• Deployment is easy, evaluation is the hard part. Hyperscalers and neo clouds made open-weight deployment simple. A top defense contractor runs entire data centres of GPUs on open-source models such as Mistral and Muse with no internet access. Knowing when a model is measurably good enough is the remaining challenge.

Four stages of enterprise AI agents