GPT-5.4 Thinking Ships — It Shows You the Plan Before Solving

33% Fewer Tokens, 33% Fewer Mistakes

OpenAI released GPT-5.4 on March 5 across ChatGPT (as GPT-5.4 Thinking), the API, and Codex. This isn't a minor version bump — it's a frontier model that unifies reasoning, coding, and agentic workflows into a single package.

What Happened

Two headline features stand out.

First, upfront planning. GPT-5.4 Thinking now shows you its reasoning plan before diving into the solution. Think of it like a math tutor writing their approach on the whiteboard first. If the model is heading in the wrong direction, you can redirect mid-response.

Second, dramatic efficiency gains. It uses 33% fewer reasoning tokens compared to GPT-5.2 while simultaneously cutting factual errors by 33%. Faster and more accurate.

Why It Matters

On the GDPVal benchmark — which measures how well AI performs economically valuable tasks — GPT-5.4 scored 83.0%, reaching human expert level. This is the most practical test of "can AI do work that actually makes money," and GPT-5.4 just cleared the bar.

Meanwhile, OpenAI has surpassed $25B in annualized revenue and is reportedly exploring a late-2026 IPO.

Going Deeper

GPT-5.4 comes in two flavors: Thinking (reasoning-optimized) and Pro (general high-performance). OpenAI also shipped Mini and Nano variants for high-volume enterprise workloads. Building on GPT-5.2's perfect AIME 2025 math score, GPT-5.4 adds token efficiency on top of raw reasoning power.

Bottom Line

GPT-5.4 Thinking = "think less, get more right." The most practically important upgrade: lower inference costs with higher accuracy.