Why GPT-5.4 Thinking Changes the Game

On March 5, 2026, OpenAI released GPT-5.4—and the headline feature isn't just another benchmark improvement. The Thinking mode introduces something genuinely new: an AI that shows you its reasoning plan before committing to an answer, and lets you steer it mid-response. That's not incremental. That's a different way of working with AI.

Whether you're a developer integrating via the API, a researcher tackling complex analysis, or a business user trying to automate multi-step workflows, GPT-5.4 Thinking offers meaningful upgrades over its predecessors. This guide covers everything you need to know—from the reasoning effort parameter to computer use capabilities, pricing, and real-world best practices.

The GPT-5.4 Model Family at a Glance

GPT-5.4 ships in five variants, each targeting different workloads:

Standard (gpt-5.4) is the general-purpose flagship at $2.50/$15 per million input/output tokens, with a massive 1,050,000-token context window. Thinking layers interactive reasoning on top of Standard—available to ChatGPT Plus subscribers and above. Pro (gpt-5.4-pro) costs $30/$180 per million tokens (12x the standard price) but delivers superior accuracy on the hardest problems, scoring 38% on FrontierMath versus Thinking's 27.1%. Mini (~$0.40/$1.60 per MTok) targets high-volume, cost-sensitive workloads, and Nano is designed for edge and embedded deployments.

The key insight: you don't need the most expensive variant for most tasks. Standard with Thinking enabled covers the vast majority of use cases.

Reasoning Effort: The Most Important New Parameter

The reasoning.effort parameter is what makes GPT-5.4 Thinking fundamentally different from previous models. Think of it as a dial that controls how deeply the model thinks before responding—and directly impacts both quality and cost.

Five levels are available:

• none — No reasoning chain. Fastest and cheapest. Behaves like a standard non-thinking model.
• low — Minimal reasoning. Good for latency-sensitive real-time applications where a small accuracy boost matters.
• medium — The default sweet spot. Balanced reasoning for general coding, analysis, and content work.
• high — Extended reasoning chains for complex debugging, multi-step problem solving, and architectural decisions.
• xhigh — Maximum depth at 3-5x the base cost. Reserve this for mathematical proofs, advanced research, and high-stakes accuracy needs.

Here's how it looks in the API:

from openai import OpenAI
client = OpenAI()

response = client.chat.completions.create(
    model="gpt-5.4",
    reasoning={"effort": "high"},
    messages=[
        {"role": "user", "content": "Refactor this function to handle edge cases..."}
    ]
)

Practical advice: Most teams should default to the none through medium range and only escalate to high or xhigh for genuinely complex tasks. Always set max_completion_tokens explicitly to prevent runaway costs—reasoning tokens count toward your bill even though they're not visible in the final output.

Steerability: Adjusting the AI Mid-Thought

This is GPT-5.4 Thinking's signature capability. When you submit a complex query, the model first generates a preamble—a short outline of how it plans to approach the problem. While the model is still thinking, you can inject additional instructions or redirect its approach entirely.

Imagine asking it to design a system architecture. The preamble reveals it's heading toward a microservices approach. You can type "Actually, use a monolithic architecture with modular boundaries" and the model adjusts its reasoning without requiring a new conversation turn.

This eliminates the frustrating cycle of generate → review → re-prompt → repeat that defined earlier models. For professionals in law, finance, research, and software architecture, steerability translates directly into time saved and better first-attempt results.

In ChatGPT, Plus subscribers get Standard and Extended thinking tiers. Pro subscribers unlock four tiers: Light, Standard, Extended, and Heavy—giving fine-grained control over how much compute the model dedicates to each response.

Computer Use: AI That Operates Your Desktop

GPT-5.4 is the first general-purpose OpenAI model with native computer use capabilities. It can view screenshots, issue mouse clicks, type on keyboards, navigate browsers, fill forms, and manage files—all autonomously.

The numbers speak for themselves: on the OSWorld benchmark, which measures desktop automation proficiency, GPT-5.4 scores 75.0%. That's a 27.7-point jump from GPT-5.2's 47.3%, and it exceeds the human expert baseline of 72.4%. This isn't a gimmick—it's a generational leap in agentic capability.

Enabling it in the API is straightforward:

response = client.chat.completions.create(
    model="gpt-5.4",
    tools=[{"type": "computer_use"}],
    messages=[
        {"role": "user", "content": "Open the browser, navigate to GitHub, and create a new repository called 'my-project'"}
    ]
)

The model can also write code to operate computers via libraries like Playwright, making it suitable for both direct UI manipulation and programmatic automation pipelines.

Benchmark Performance: The Numbers

GPT-5.4 posts strong results across multiple evaluation suites. On SWE-bench Verified (standard coding tasks), it hits approximately 80%. On SWE-bench Pro (novel codebases), it scores 57.7%, up from GPT-5.3-Codex's 55.6%—confirming that coding specialist capabilities have been successfully folded into the general-purpose model.

On GDPval (knowledge work and research), GPT-5.4 reaches 83%, a significant jump from GPT-5.2's 70.9%. Factual accuracy has also improved measurably: individual claims are 33% less likely to be false, and full responses are 18% less likely to contain any errors compared to GPT-5.2.

For the hardest mathematical reasoning, GPT-5.4 Pro leads with 38% on FrontierMath, though the standard Thinking model's 27.1% still represents a capable reasoning system for most practical applications.

Pricing and Access: What You'll Pay

ChatGPT subscriptions: Plus ($20/month) includes Thinking mode with 80 messages per 3-hour window. Pro ($200/month) unlocks the Pro variant with higher limits and dedicated GPU allocation. Business plans start at $25/user/month.

API pricing has one critical detail that's easy to miss: the 1M-token context window comes with a surcharge. Standard input pricing is $2.50 per million tokens up to 272K tokens. Beyond 272K, the rate doubles to $5.00 per million tokens. If you're processing large documents or entire codebases, staying under that 272K threshold when possible can halve your input costs.

Maximum output is 128,000 tokens. API access requires a paid account with at least $5 in prior spend (Tier 1).

Best Practices for GPT-5.4 Thinking Workflows

After extensive testing and reviewing OpenAI's official prompt guidance, here are the strategies that matter most:

Define your output contract explicitly. GPT-5.4 performs best when you specify the exact format, tool-use expectations, and completion criteria. Tell the model what "done" looks like.

Use the RACE framework for system prompts. Structure them with Role, Action, Context, and Expectation. GPT-5.4 responds exceptionally well to structured system prompts compared to free-form instructions.

Encourage pre-tool reasoning. Adding "Before you call a tool, explain why you are calling it" to your system prompt boosts tool-calling accuracy without inflating reasoning overhead.

Give full goals, not single steps. GPT-5.4 is built for multi-step tasks. Instead of micromanaging each step, provide the complete objective and let the model plan the workflow.

Front-load critical context. The model pays strongest attention to the beginning and end of its input. Place the most important information early in your prompt.

GPT-5.4 Thinking vs. Pro: Which Should You Use?

The decision framework is simpler than it might seem. Choose Thinking for research, code architecture, analytical writing, and any task where steerability matters. It's available on the $20/month Plus plan, and with reasoning effort tuning, you can balance cost and quality precisely.

Choose Pro when you need maximum accuracy on the hardest problems—advanced mathematics, frontier research, high-stakes enterprise decisions—and cost is secondary. The 12x price premium is significant, so Pro should be reserved for tasks where the accuracy difference demonstrably matters.

For most developers and knowledge workers, Standard GPT-5.4 with Thinking enabled at medium reasoning effort is the right starting point. Upgrade to high or xhigh effort selectively, and reach for Pro only when the stakes justify the cost.

Looking Ahead

GPT-5.4 Thinking represents a genuine shift in how we interact with AI models. The ability to observe and redirect a model's reasoning process in real time, combined with fine-grained cost-quality controls via reasoning effort, makes this the most usable reasoning model released to date. Add native computer use that exceeds human performance on desktop automation benchmarks, and you have a model that's moved well beyond chatbot territory into genuine AI assistant capabilities. If you're still on the fence, start with medium effort on a task you know well—and see how the thinking process changes your workflow.