OpenAI's GPT-5.4, released on March 5, 2026, scored 75% on OSWorld-Verified — surpassing the average human benchmark of 72.4% — and hit a record 83% on the company's own GDPval test for professional knowledge work. Five days on, it's worth looking at what these numbers actually mean and why this launch is different from the incremental updates we've come to expect.
What GPT-5.4 Actually Is
GPT-5.4 ships in three flavors: the standard model, GPT-5.4 Thinking (a reasoning-first version), and GPT-5.4 Pro (optimized for peak performance). The API version supports context windows up to 1 million tokens — by far the largest context window OpenAI has ever offered, and a major jump from GPT-5.2's limit.
The model was also built to be cheaper to run. OpenAI says GPT-5.4 solves the same problems using significantly fewer tokens than its predecessor, which matters when you're running it at scale. Pricing starts at $2.50 per million input tokens.
One under-the-radar change: the API's new Tool Search system. Previously, every API call would include full definitions for every available tool in the system prompt — burning tokens just to describe tools the model might never use. Tool Search lets the model look up definitions on demand. In large agentic systems with dozens of tools, that's a meaningful cost and latency win.
The Benchmarks That Matter
OpenAI is leaning hard on three numbers:
- OSWorld-Verified: 75.0% — tests the model's ability to navigate and operate a desktop computer. Human average: 72.4%. GPT-5.2, for comparison, scored 47.3%.
- GDPval: 83% — OpenAI's internal benchmark for professional knowledge work tasks. Record high.
- APEX-Agents (Mercor): #1 — tests AI professional competency in law and finance. GPT-5.4 took the top spot.
Mercor CEO Brendan Foody put it plainly: GPT-5.4 "excels at creating long-horizon deliverables such as slide decks, financial models, and legal analysis — delivering top performance while running faster and at a lower cost than competitive frontier models."
The OSWorld number is the big one. Hitting 75% while the average human scores 72.4% means GPT-5.4 can now navigate a computer desktop more reliably than a typical person. That's not just a benchmark bragging point — it's a threshold that makes fully autonomous computer use viable for real workflows.
Error Reduction and Safety
OpenAI claims GPT-5.4 is 33% less likely to make errors in individual claims compared to GPT-5.2, and 18% less likely to produce responses containing errors overall. For anyone using these models for anything important, that's a significant reliability improvement.
The company also added a new safety evaluation testing for chain-of-thought deception — the concern that reasoning models might hide their true reasoning process. OpenAI's results suggest the Thinking version of GPT-5.4 is less prone to this kind of obfuscation, meaning chain-of-thought monitoring remains a viable safety tool.
What This Means for Builders
Three immediate implications:
1. Computer-use agents just got real. The gap between "can technically navigate a desktop" and "can actually do useful work without constant babysitting" just closed significantly. If you're building agentic workflows that need to interact with legacy software, this changes the economics of the build.
2. Context windows matter more now. 1 million tokens isn't just a bigger number — it enables workflows like "analyze this entire codebase" or "review these 50 legal contracts at once" without the gymnastics of chunking and stitching.
3. Cost efficiency compounds. Fewer tokens per task + cheaper per token = meaningful cost reduction at scale. For products already running on GPT-5.2, this is an immediate margin improvement without changing anything else.
What's Next
We're in a phase where the frontier labs are converging on agentic capabilities. Anthropic launched computer use in late 2024. Google followed with Gemini. Now OpenAI has closed the gap with a model that actually outperforms humans on desktop navigation benchmarks.
The next six months will reveal whether these capabilities translate to real-world adoption. Benchmarks are one thing; reliable operation in messy production environments is another. But the trajectory is clear: we're approaching the point where AI agents can handle complex computer-based tasks with minimal supervision.
The builders who figure out how to deploy this reliably will have a significant advantage over those still treating AI as a chat interface.
FOLLOW NeuralWire on X for daily AI signal — what matters, why it matters, what to do about it. →