An AI That Can Read an Entire Book in One Sitting

On March 5, 2026, OpenAI released GPT-5.4 — and it's not just another incremental update. For the first time, a single general-purpose model combines a 1,050,000-token context window, native computer control, and full-resolution vision. One million tokens translates to roughly 750,000 words: an entire codebase, a year's worth of financial reports, a complete legal discovery package, or multiple academic research papers — all processable in a single conversation.

This matters because the biggest practical limitation of AI models has always been memory. You had to chunk documents, lose context between sessions, and constantly remind the model what you were working on. GPT-5.4 effectively demolishes that wall.

The GPT-5 Evolution: How We Got Here

GPT-5.4 is the fourth major iteration in the GPT-5 series. Starting with GPT-5.0 in mid-2025, each version has pushed boundaries: GPT-5.2 dramatically improved reasoning, GPT-5.3-Codex became the industry's top coding model, and now GPT-5.4 consolidates all of these advances into one unified package.

What makes this release particularly significant is that GPT-5.4 inherits GPT-5.3-Codex's industry-leading coding capabilities while expanding into professional workflows — spreadsheets, presentations, document analysis, and financial modeling. OpenAI describes it as "the most capable and efficient frontier model for professional work."

Efficiency gains are equally notable. GPT-5.4 uses significantly fewer tokens to solve problems compared to GPT-5.2, translating to both faster responses and lower costs for equivalent tasks. Individual claim error rates dropped by 33%.

Core Capabilities Deep Dive

The 1 Million Token Context Window

GPT-5.4 supports exactly 1,050,000 input tokens and 128,000 output tokens. But there's a critical pricing threshold developers need to understand: once input exceeds 272,000 tokens, OpenAI charges 2x for input and 1.5x for output across the entire session.

Here's the complete pricing breakdown:

• Standard input (≤272K): $2.50 per million tokens
• Extended input (>272K): $5.00 per million tokens
• Standard output: $15.00 per million tokens
• Extended output (>272K): $22.50 per million tokens
• Cached input: $0.25 per million tokens (90% automatic discount)

In Codex, the 1M context window is available experimentally, enabling agents to plan, execute, and verify tasks across long horizons — think entire repository analysis or multi-file refactoring sessions.

Native Computer Use

This is where GPT-5.4 makes history. It achieves 75.0% on the OSWorld benchmark, surpassing the human performance baseline of 72.4% and crushing GPT-5.2's 47.3%. The model can interpret screenshots, execute mouse and keyboard commands, and write automation code using libraries like Playwright.

Practical example: an agent that reads emails, extracts assignment attachments, uploads and grades them, then records results in a spreadsheet — all autonomously across multiple applications.

from openai import OpenAI
client = OpenAI()

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

On BrowseComp, which measures persistent web browsing for hard-to-find information, GPT-5.4 Pro reached 89.3% — a new state of the art, improving 17 percentage points over GPT-5.2.

Tool Search

GPT-5.4 introduces Tool Search, a feature that helps agents efficiently find and use the right tools across large ecosystems of APIs and connectors. Instead of stuffing every tool description into the prompt, Tool Search lets the model dynamically discover what it needs, reducing token usage by 47% without sacrificing intelligence. This is particularly valuable in enterprise environments with hundreds of integrated services.

Reasoning Effort Control

Developers can fine-tune the model's thinking depth via the reasoning.effort parameter across five levels:

• none: Simple formatting and extraction (fastest, cheapest)
• low: Straightforward Q&A and classification
• medium: General coding and analysis (default)
• high: Complex debugging and architecture decisions
• xhigh: Hard math, security audits, research (most accurate)

Each step up costs roughly 3-5x more due to internal chain-of-thought token generation. The recommendation: start at medium and escalate only when needed.

Benchmark Performance: GPT-5.4 by the Numbers

SWE-Bench Verified (software engineering): 80.0%, just 0.8 points behind Claude Opus 4.6's 80.8%. But on SWE-Bench Pro — the harder variant that strips away memorizable patterns — GPT-5.4 scored 57.7% versus Opus's ~45%, a roughly 28% advantage that suggests stronger generalization on novel engineering challenges.

GDPval (knowledge work): GPT-5.4 matched or outperformed industry professionals in 83% of real-world comparisons across 44 professional occupations.

Financial modeling: 87.3% on spreadsheet modeling tasks typical of junior investment banking analysts, up from GPT-5.2's 68.4%.

OSWorld (computer use): 75.0%, above the 72.4% human baseline — a first for any AI model.

GPT-5.4 vs Claude Opus 4.6: The 2026 Showdown

As of March 2026, GPT-5.4 and Claude Opus 4.6 are the two titans. Both support million-token contexts, but their approaches differ significantly.

Context window: GPT-5.4 offers 1.05M tokens natively, while Claude Opus 4.6 defaults to 200K with 1M available in beta via a special header. Claude scores 76% on MRCR v2 (a needle-in-a-haystack retrieval test), demonstrating strong long-context retrieval accuracy.

Pricing: The gap is substantial. Claude Opus 4.6 charges $5/million input and $25/million output flat. GPT-5.4's base rate of $2.50 input is cheaper for shorter contexts, but the 272K surcharge makes extended contexts more expensive — $5 input and $22.50 output. The GPT-5.4 Pro tier at $30/$180 per million tokens is dramatically more expensive than any Claude tier.

Strengths: GPT-5.4 dominates in computer use (75% OSWorld), general knowledge work (83% GDPval), and SWE-Bench Pro (57.7%). Claude Opus 4.6 excels at code-heavy agentic engineering and SWE-Bench Verified precision (80.8%). If you want one model for everything, GPT-5.4 is the current best bet. If your workflow is primarily coding-focused agentic tasks, Claude Opus 4.6 remains the stronger specialist.

Getting Started: API Integration

You need a paid OpenAI API account with at least $5 in prior usage (Tier 1). GPT-5.4 is a drop-in replacement if you're already using the Chat Completions API:

from openai import OpenAI
client = OpenAI()

response = client.chat.completions.create(
    model="gpt-5.4",
    messages=[{"role": "user", "content": "Analyze this document..."}]
)

For advanced features, OpenAI recommends the Responses API, which natively supports reasoning parameters, tool registration, and large context sizes. Model IDs are gpt-5.4 for standard and gpt-5.4-pro for maximum accuracy on high-stakes tasks.

Practical Tips for Cost Optimization

Respect the 272K threshold. Beyond this point, your entire session's pricing doubles for input. Most tasks are perfectly handleable within 272K tokens — reserve the full million for genuinely massive document analysis or whole-codebase operations.

Leverage caching aggressively. Cached input costs just $0.25 per million tokens — a 90% discount that's applied automatically to repeating context. If you're asking multiple questions against the same document base, this adds up to massive savings.

Match reasoning effort to the task. Don't use xhigh for simple data extraction. Start at medium, evaluate quality, and escalate only when results warrant it.

Use Tool Search for complex agent systems. With a 47% reduction in token usage, Tool Search pays for itself in environments with many integrated tools. Let the model discover capabilities dynamically rather than front-loading every tool description.

Looking Ahead

GPT-5.4 represents a genuine inflection point. The combination of million-token context, superhuman computer use, and professional knowledge work capabilities signals that AI models are evolving from text generators into digital coworkers that can understand entire projects and operate across applications autonomously. While independent benchmark verification is still catching up, the breadth of GPT-5.4's capabilities — coding, reasoning, computer operation, financial modeling, and document analysis in a single model — is unprecedented. For developers and enterprises, the time to explore what's possible with this new class of AI is now.