Why GPT-5.4 Changes the Game

On March 5, 2026, OpenAI launched GPT-5.4—and it's not just another incremental update. With a 1 million token context window, native computer use capabilities, and a five-level reasoning effort system that developers can tune in real time, GPT-5.4 represents the moment AI models shifted from sophisticated text generators to genuine digital coworkers.

Whether you're a ChatGPT Plus subscriber exploring Thinking mode or a developer integrating the API into production workflows, this guide covers everything you need to get the most out of GPT-5.4—from reasoning effort tuning to prompt engineering best practices and cost optimization strategies.

What Exactly Is GPT-5.4?

GPT-5.4 is OpenAI's most capable frontier model, inheriting the industry-leading coding capabilities of GPT-5.3-Codex while dramatically improving reasoning, agentic workflows, and tool utilization. Here are the key specs:

• Context Window: 1,050,000 tokens (~750,000 words—roughly seven Harry Potter books)
• Max Output Tokens: 128,000
• Knowledge Cutoff: August 31, 2025
• Reasoning Effort Levels: none, low, medium, high, xhigh
• Modalities: Text in/out, image input
• Model Variants: gpt-5.4 (standard), gpt-5.4-pro (maximum performance), gpt-5.4-mini (lightweight), gpt-5.4-nano (ultra-lightweight)

What sets GPT-5.4 apart from its predecessors is that it's the first general-purpose model with native computer-use capabilities. It can write Playwright code, read screenshots, and issue keyboard and mouse actions to interact with software directly. On WebArena-Verified, it achieves a 67.3% success rate for browser use, and on Online-Mind2Web, it hits 92.8% using screenshot-based observations alone.

Mastering Reasoning Effort: The 5-Level System

The reasoning.effort parameter is arguably GPT-5.4's most practically important feature for developers. It controls how many hidden "reasoning tokens" the model generates before producing its visible response—essentially how deeply it thinks before answering.

Understanding Each Level

none (default) — Zero reasoning tokens, minimum latency. This is the right choice for deterministic, lightweight tasks: data extraction, formatting, simple classification, short rewrites. Since GPT-5.2, none has been the default, prioritizing speed.

low — A touch of reasoning for straightforward queries that need quick turnaround. Think customer support triage or simple Q&A where context awareness matters but deep analysis doesn't.

medium — The jack-of-all-trades setting. It delivers a strong balance of performance and speed, handling everything from writing to coding to moderate analytical tasks. This is the sweet spot for most business applications.

high — For multi-document review, complex debugging, conflict resolution in data, and strategy writing. When the task involves synthesizing information from multiple sources, high delivers noticeably better results.

xhigh — Maximum reasoning power. Reserve this for mathematical proofs, intricate logic puzzles, large codebase analysis, and problems where correctness is paramount and latency is acceptable. Costs scale significantly at this level.

Code Examples

Python — Basic reasoning control:

from openai import OpenAI
client = OpenAI()

# Fast extraction (no reasoning)
response = client.responses.create(
    model="gpt-5.4",
    input="Extract all email addresses from this text: ...",
    reasoning={"effort": "none"}
)

# Deep analysis (high reasoning)
response = client.responses.create(
    model="gpt-5.4",
    input="Compare these 3 contracts and identify key differences.",
    reasoning={"effort": "high"}
)

JavaScript — With verbosity control:

const response = await openai.responses.create({
  model: "gpt-5.4",
  input: "Summarize this quarterly report.",
  reasoning: { effort: "medium" },
  text: { verbosity: "low" }
});

Critical insight: Treat reasoning.effort as a tuning knob, not a quality recovery mechanism. If output quality is lacking, improve your prompt first—add output contracts, verification loops, and explicit completion criteria—before reaching for higher reasoning levels.

Parameter Compatibility Note

When reasoning effort is none, you can use temperature, top_p, and logprobs. At all other reasoning levels, these parameters are rejected by the API.

Thinking Mode in ChatGPT

Not an API developer? GPT-5.4 Thinking is available directly in ChatGPT through the model picker for Plus, Pro, and Business subscribers.

Thinking mode excels in several key scenarios:

Multi-step reasoning tasks — Drafting decision briefs with options, trade-offs, and recommendations. Extracting themes from large volumes of customer feedback. Building scenario narratives from structured inputs. Where Fast mode might occasionally oversimplify or miscalculate, Thinking mode works through problems step by step.

Long-context analysis — Combined with the 1M token context window, you can analyze entire codebases, process large document collections, transform policy documents into navigable knowledge bases, or auto-generate SOPs.

Persistent research — GPT-5.4 Thinking is significantly stronger at "needle-in-a-haystack" research questions, persistently searching across multiple rounds to find the most relevant sources.

The 1 Million Token Context Window

One crucial detail: the 1M context window is not enabled by default. In the API, you need to explicitly configure model_context_window and model_auto_compact_token_limit. Without these parameters, you're working with the standard 272K window.

The cost implications matter too. Once your prompt exceeds 272K input tokens, pricing jumps to 2x for input and 1.5x for output across the entire session. Use the extended context strategically—full codebase analysis, large-scale legal document review, extended agent trajectories, or multi-paper research synthesis—rather than as a default setting.

To put 1M tokens in perspective: that's roughly 750,000 words, enough to hold seven Harry Potter novels, an entire medium-sized codebase, or hundreds of research papers simultaneously.

Computer Use: AI That Operates Software

GPT-5.4 introduces native computer use—the ability to inspect screenshots and return structured actions (mouse clicks, keyboard input, Playwright code) that your harness executes. This isn't a plugin; it's a built-in capability.

Practical applications include browser automation for web scraping and testing, automated form filling across enterprise applications, financial workflow plugins for Excel and Google Sheets, and desktop workflow automation for repetitive tasks.

OpenAI recommends using computer use in isolated environments with human oversight. It's a powerful capability, but the responsible deployment pattern involves sandboxed execution with approval gates for irreversible actions.

Pricing Breakdown

Here's the complete pricing picture as of March 2026:

GPT-5.4 (Standard)

• Input: $2.50 per 1M tokens
• Cached Input: $0.25 per 1M tokens (90% savings on repeated context)
• Output: $15.00 per 1M tokens
• Long context surcharge (>272K): 2x input, 1.5x output

GPT-5.4 Pro — $30.00 per 1M input tokens. Maximum performance for the most demanding tasks.

GPT-5.4 Mini — $0.75/$4.50 per 1M input/output tokens. Ideal for high-volume workloads where cost efficiency matters.

Cost optimization tip: Structure prompts with static content (system prompts, fixed instructions) at the beginning and dynamic content (user input, variable context) at the end. This maximizes cache hit rates, and cached tokens cost 90% less than standard input tokens.

ChatGPT Plan Access

Plus ($20/month): Access to GPT-5.4 Thinking via model picker, up to 3,000 messages per week.

Pro ($200/month): Unlimited access to all GPT-5 models including GPT-5.4 Pro.

Business: Full model access including GPT-5.4 Pro with team management features.

Enterprise/Edu: Full access with admin-configurable early access settings.

Prompt Engineering Best Practices for GPT-5.4

The prompt engineering landscape has shifted significantly with GPT-5.4. Here are the highest-leverage strategies:

The CTCO Framework

Context → Task → Constraints → Output format. This pattern is the most reliable way to prevent hallucinations and generic outputs. The era of "You are a helpful assistant" is over.

Output Contracts

Explicitly define sections, ordering, length limits, and required formats. Combine with the text.verbosity parameter (low, medium, high) for precise control over response length:

response = client.responses.create(
    model="gpt-5.4",
    input="Summarize this report.",
    text={"verbosity": "low"},
    reasoning={"effort": "medium"}
)

Tool Persistence for Agents

When building agentic workflows, always include: "Don't stop early on tool calls. Keep calling tools until the task completes and verification passes. Add dependency checks before actions."

Verification Loops

Instruct the model to check correctness against requirements, ground claims in provided context, verify formatting matches the schema, and gate irreversible actions—all before finalizing output.

Caching-Friendly Layouts

The biggest operational shift is toward prompt structures that maximize cache hits. Static system prompts go first; variable user content goes last. This directly reduces costs through the cached input discount.

Migration Guide

From GPT-5.2: GPT-5.4 works as a drop-in replacement in most cases. Keep your existing reasoning effort levels, run evals, then iterate.

From GPT-4o: Start with reasoning.effort: none and work upward. The model architecture is significantly different, so existing prompts may need adaptation.

For long-running agents: Preserve the phase field in assistant messages ("commentary" for intermediate updates, "final_answer" for completed responses). Dropping phase causes preambles to be misinterpreted as final answers. Use previous_response_id when possible for cleaner state recovery.

What's Next

GPT-5.4 isn't just a smarter model—it's a fundamentally different kind of tool. Fine-grained reasoning effort control lets you optimize the cost-performance tradeoff per request. The 1M token context window enables tasks that were previously impossible in a single pass. Native computer use turns AI from an advisor into an operator. The best way to start is small: use reasoning.effort: none for your existing lightweight tasks, confirm you're getting the same or better results at lower cost, then gradually explore medium and high effort for your most complex workflows. The model rewards precise, well-structured prompts with output contracts and verification loops—invest in your prompt engineering, and GPT-5.4 will meet you more than halfway.