OpenAI's GPT-5.4 Scores 83% on Knowledge Work Benchmark — Matching Human Professionals Across 44 Occupations

OpenAI released GPT-5.4 on March 5, 2026, calling it "our most capable and efficient frontier model for professional work." The model ships in three variants: standard GPT-5.4, reasoning-optimized GPT-5.4 Thinking, and high-performance GPT-5.4 Pro. Its API context window extends to 1 million tokens — the largest OpenAI has offered. On the GDPval benchmark, which measures knowledge work across 44 occupations, GPT-5.4 scored 83%, up from 70.9% for GPT-5.2 (OpenAI, March 2026). On OSWorld-Verified, a computer-use benchmark, it scored 75% — above the 72.4% human baseline (OpenAI, March 2026). The jump from GPT-5.2 to GPT-5.4 represents the steepest single-generation improvement OpenAI has shipped on professional task completion.

What can GPT-5.4 actually do that previous models couldn't?

Native computer-use capability is the headline addition. GPT-5.4 can operate desktops, browsers, and software applications autonomously — clicking buttons, filling forms, switching between tabs, and completing multi-step workflows without human intervention. The model combines the coding strengths of GPT-5.3-Codex with improved general reasoning and introduces a new Tool Search system that reduces token consumption when working with large tool sets. OpenAI reports 33% fewer individual claim errors and 18% fewer response-level errors compared to GPT-5.2 (OpenAI, March 2026). Token efficiency improved substantially: GPT-5.4 solves the same problems with fewer tokens, reducing cost per task even before accounting for its $2.50-per-million input token pricing.

How does GPT-5.4 affect everyday consumers?

When an AI model can autonomously operate a computer and complete multi-step workflows, the implications reach well beyond enterprise software. Shopping comparison, travel booking, financial research — tasks that currently require consumers to open multiple browser tabs and manually cross-reference information — could be handled by an AI agent navigating sites on their behalf. This capability makes agentic commerce protocols like Google's Universal Commerce Protocol and OpenAI's own Agentic Commerce Protocol more practical. An agent that can read a product page, add items to a cart, and apply a coupon code needs exactly the kind of computer-use ability GPT-5.4 demonstrates. Comparison platforms structuring data for machine readability stand to benefit directly, because agents performing autonomous browsing will prioritize sources where product information is consistently formatted and accurate.

Where does GPT-5.4 sit in the competitive field?

GPT-5.4 trades blows with Claude Opus 4.6 on SWE-bench coding benchmarks, with neither model holding a definitive lead across all task categories. Pricing tells a different story: GPT-5.4's $2.50 per million input tokens runs at roughly 40% of Claude Opus 4.6's output token cost. OpenAI followed up on March 17 with GPT-5.4 mini and nano — smaller variants targeting high-volume workloads. GPT-5.4 mini approaches the full model's performance while running 2x faster (OpenAI, March 2026). OpenAI has surpassed $25 billion in annualized revenue (Crescendo AI/industry reports, 2026), and the efficiency gains in GPT-5.4 suggest the company is optimizing for margin alongside capability. The model race is shifting from raw benchmark scores to deployment cost and token efficiency — the metrics that determine whether AI agents are economically viable at consumer scale.