GPT-5.4 Native Computer Use — Human-Level Desktop Tasks

GPT-5.4 launched with a Thinking variant that integrates test-time compute and native computer use (OS-level navigation: files, browsers, terminals). On OSWorld-Verified (desktop task completion benchmarks), GPT-5.4 Thinking surpassed human-level performance. Three variants: Standard, Thinking (test-time compute), Pro. This is distinct from prior Codex/agentic-coding capabilities — GPT-5.4 can autonomously navigate any desktop application, not just code editors. GPT-5.5 ("Spud") pretraining completed March 24; Polymarket assigns 95%+ probability of release by June 30.

Native computer use at the frontier model level (not a specialized "computer use" product) signals that desktop automation is becoming a standard capability of leading models, not a niche add-on. For GTM and revenue operators: any workflow that currently lives inside a GUI — CRM data entry, ad platform management, reporting dashboards — is now in scope for autonomous agent execution. The OSWorld benchmark crossing human-level performance is the leading indicator that this isn't just demo capability. The practical GTM implication: "I can't automate X because it's only in a GUI" is no longer a valid blocker.

• "GPT-5.4 Thinking just crossed human-level performance on desktop task benchmarks. Any workflow that lives in a GUI is now in scope for AI automation. The 'it's only in a UI' objection is gone."
• "Three months ago, native computer use was a demo. Today GPT-5.4 ships it as a standard model variant. The speed from research capability to product-grade deployment is accelerating."
• "GPT-5.5 ('Spud') pretraining is done. Polymarket puts 95%+ probability on release by June. The frontier is moving faster than product roadmaps can absorb."