GPT-5.4 Thinking: the model that surpasses human experts
OpenAI just launched GPT-5.4 with Thinking mode, and the benchmark numbers are historic: 83% on GDPVal (Graduate-level Diverse Professional Valuation benchmark), placing the model at or above human expert level on economically relevant tasks like legal analysis, medical diagnosis, software engineering, and financial research.
Having used GPT-4.5 for daily work for months, the difference with Thinking mode is tangible: the model now "thinks out loud" before responding, showing its reasoning step by step — similar to how OpenAI o1 works but with the conversational versatility of GPT-5.
GPT-5.4 Thinking vs previous models
| Model | GDPVal | MMLU | HumanEval (code) | Access |
|---|---|---|---|---|
| GPT-4o | 61% | 88% | 90% | Plus/Free limited |
| GPT-5 | 74% | 91% | 95% | Plus/Pro |
| GPT-5.4 Thinking | 83% | 94% | 98% | Plus/Pro |
| Human expert level | ~80% | ~90% | ~85% | — |
What is Thinking mode and how does it work
Thinking mode is an internal reasoning chain the model runs before generating the final response. When enabled, you see the model's "thought process": hypotheses, verifications, corrections, and conclusion.
Practical example: ask it to review a legal contract, and Thinking mode will first analyze risky clauses, verify known legal precedents, identify ambiguities, and finally deliver its assessment with complete reasoning.
How to access GPT-5.4 Thinking right now
# Access via ChatGPT web:
1. Go to https://chat.openai.com
2. Make sure you have ChatGPT Plus ($20/mo) or Pro ($200/mo)
3. In the model selector (top left), choose "GPT-5.4 Thinking"
4. Type your prompt normally
# Via OpenAI API:
import openai
client = openai.OpenAI(api_key="your-api-key")
response = client.chat.completions.create(
model="gpt-5.4-thinking",
messages=[
{"role": "user", "content": "Analyze this contract: [text]"}
],
reasoning_effort="high" # "low", "medium" or "high"
)
print(response.choices[0].message.content)Where Thinking mode excels
After testing the model for days, these are the areas where the improvement is most noticeable:
- Math and statistics: Solves olympiad-level problems with near-100% accuracy
- Complex code: Generates multi-file algorithms with correct dependency logic
- Legal analysis: Identifies abusive clauses in contracts with real legal references
- Medical research: Synthesizes scientific literature with evidence quality evaluation
- Strategic planning: Builds detailed business plans with thorough SWOT analysis
How it affects you as a ChatGPT user
If you use ChatGPT for work, Thinking mode changes what you can ask of it. You no longer need to fragment complex tasks into 10 messages — the model can handle long, multi-step instructions in a single prompt:
- Plus ($20/mo): Access to GPT-5.4 Thinking with monthly token limits
- Pro ($200/mo): No token limits in Thinking mode
- API: Per-token pricing, more expensive than standard GPT-5 but cheaper than o1-pro
The jobs concern
Stanford's SIEPR summit report this week shows AI has already cut entry-level software developer hiring by 20% and call center jobs by 15%. GPT-5.4 Thinking accelerates this trend.
Economists' recommendation isn't to avoid AI — it's to use it to do higher-value work. In my experience, professionals who adopt these models as a "copilot" are the ones maintaining competitive edge.
Common issues using GPT-5.4 Thinking
Issue 1: "I don't see GPT-5.4 Thinking in the model selector." The rollout is gradual. Solution: sign out and back in. If it's still not there, wait 48 hours. Long-time Plus users get access first.
Issue 2: "Thinking mode is very slow." Internal reasoning takes 10-60 seconds depending on complexity. Solution: use reasoning_effort="low" in the API or simplify your prompt for tasks that don't need deep reasoning.
Issue 3: "Responses are less accurate in non-English languages." GDPVal is primarily evaluated in English. Solution: for critical tasks, write the prompt in English and request the response in your language at the end.
