An AI That Can Read Seven Harry Potter Books in One Sitting

On March 5, 2026, OpenAI released GPT-5.4 — what they're calling their "most capable and efficient frontier model for professional work." The headline feature is a 1 million token context window, roughly equivalent to 750,000 words or the entire Harry Potter series loaded into a single prompt. But GPT-5.4 isn't just about reading more text. It's the first general-purpose AI model that can directly operate your computer, autonomously find and use tools from massive ecosystems, and reason across codebases that previously required multiple sessions to analyze.

Before you rush to max out that context window, though, there's a catch — accuracy drops to 36% at the extreme end, and costs double past a critical threshold. This guide breaks down exactly how GPT-5.4 works, what the benchmarks actually show, and how to get optimal performance without burning through your API budget.

Why GPT-5.4, Why Now

GPT-5.4 represents a consolidation moment for OpenAI. It merges the coding prowess of GPT-5.3-Codex with the reasoning capabilities of GPT-5.2 into a single unified model. Previously, developers had to choose between specialized models for different tasks — coding, reasoning, general conversation. GPT-5.4 handles all three, and adds computer use on top.

The timing isn't coincidental. Google's Gemini has offered million-token context for over a year, and the AI agent paradigm demands models that can hold entire projects in memory. When an AI agent needs to plan, execute, and verify a multi-step workflow — say, refactoring a large codebase or auditing a 500-page regulatory document — a 128K window simply doesn't cut it. OpenAI needed to match the competition while pushing ahead on the agentic capabilities that make long context actually useful.

The 1M Context Window: Specs, Reality, and the 272K Cliff

What You Get

GPT-5.4's total context window is 1,050,000 tokens — 922K for input and 128K for output. However, the default configuration runs at 272K tokens. The full million-token capacity is an experimental opt-in feature, primarily available through Codex and the API.

The Performance Reality

Here's what the benchmarks actually show — and it's a story of diminishing returns:

• 16K–32K tokens: ~97% retrieval accuracy
• 127K–272K tokens: ~97% accuracy maintained — the sweet spot
• 256K–512K tokens: Noticeable degradation begins
• 512K–1M tokens: Accuracy drops to approximately 36%

On OpenAI's own MRCR v2 8-needle benchmark, GPT-5.4 scores 79.3% at 128K–256K and just 36.6% at the 512K–1M range. On the Graphwalks BFS benchmark, accuracy at the 256K–1M range is 21.4%. In practical terms, if you fill the full million-token window, the model will reliably find only about one-third of the information you put in.

The Pricing Cliff

Cost compounds the accuracy problem. GPT-5.4 introduces a hard pricing threshold at 272K tokens:

| Token Range | Input Cost (per 1M) | Output Cost (per 1M) | |-------------|---------------------|---------------------| | ≤272K | $2.50 | $15.00 | | >272K | $5.00 | $22.50 |

Critically, once your prompt exceeds 272K tokens, the higher rate applies to all tokens in the session — not just the overflow. Moving from 272K to 400K roughly triples your per-call cost. Cached inputs at $0.25/M (a 90% discount) can offset this, but you need to architect your prompts around caching to benefit.

How to Enable 1M Context in Codex

For cases where you genuinely need the extended window — analyzing an entire codebase, comparing massive document sets — here's the setup:

Step 1: Update Codex CLI

npm install -g @openai/codex@latest

Step 2: Configure config.toml

model_context_window = 1000000
model_auto_compact_token_limit = 900000

Step 3: Select the model Use the /model command within Codex to switch to GPT-5.4.

Without these explicit settings, you stay on the standard 272K window.

Native Computer Use: The First AI to Beat Humans at Desktop Tasks

Perhaps more consequential than the context window is GPT-5.4's native computer use capability. This is the first general-purpose model that can look at your screen, understand what it sees, and take action — clicking buttons, filling forms, navigating between applications, writing and executing code.

It works through two mechanisms:

• Screenshot-based direct control: The model captures screenshots, interprets the visual state, and issues mouse and keyboard commands to interact with the UI.
• Playwright code generation: For web and application automation, it writes and executes Playwright scripts for programmatic control.

The benchmark results are striking. On OSWorld-Verified, GPT-5.4 achieves a 75.0% success rate on autonomous desktop tasks, surpassing the human expert baseline of 72.4%. This is the first time any frontier model has beaten humans at general-purpose computer operation. Additional scores include 67.3% on browser-specific tasks and 92.8% on screenshot interpretation.

For comparison, GPT-5.2 scored just 47.3% on the same benchmark — GPT-5.4 represents a 58% improvement in a single generation.

Practical applications include automated email drafting and sending, spreadsheet data manipulation, web form completion, cross-application workflow automation, and software testing through UI interaction.

Tool Search: Smarter Tool Use at 47% Lower Cost

If you've built AI agents that connect to multiple services via MCP (Model Context Protocol) servers, you've likely hit the tool definition bloat problem. With 36 MCP servers enabled, the tool descriptions alone can consume thousands of tokens before the model even starts working.

GPT-5.4's Tool Search feature fundamentally restructures this. Instead of loading every tool definition into the prompt, the model receives a lightweight index of available tools and a search capability. It retrieves full definitions only for tools it actually needs.

The results are significant: on Scale's MCP Atlas benchmark with 250 tasks across 36 MCP servers, tool search reduced total token usage by 47% while maintaining identical accuracy to the full-context approach. For agent-heavy production workloads, this translates directly to cost savings.

To enable tool search, you'll need to configure it explicitly through the API — it's not on by default. The feature works with the Responses API framework and supports MCP server integration.

Benchmark Scorecard: Where GPT-5.4 Excels

Beyond computer use, GPT-5.4 posts strong numbers across professional benchmarks:

Knowledge Work (GDPval): 83.0% overall (up from GPT-5.2's 70.9%), with spreadsheet modeling at 87.3% and presentation quality preferred by human raters 68% of the time. Across 44 professional occupations, the model matches human expert performance 83% of the time.

Coding: SWE-Bench Pro at 57.7%, terminal operations at 75.1%. It's the first mainline model to incorporate frontier-level coding from the Codex line.

Advanced Reasoning: ARC-AGI-2 jumps to 73.3% (from 52.9%), GPQA Diamond hits 92.8%, and FrontierMath's hardest tier reaches 27.1%.

Choosing the Right GPT-5.4 Variant

GPT-5.4 ships in three configurations:

• GPT-5.4 (standard): $2.50 input / $15.00 output per million tokens. Best for most workflows.
• GPT-5.4 Thinking: Adds configurable reasoning effort (none, low, medium, high, xhigh). Use when complex multi-step reasoning justifies the extra computation.
• GPT-5.4 Pro: $30 input / $180 output per million tokens. Enterprise-grade maximum performance.

The reasoning effort parameter is particularly useful. Benchmarks are measured at xhigh, but production workloads often perform well at medium or high, saving significant compute.

Practical Optimization Strategies

Keep context under 272K whenever possible. Most multi-turn conversations, document analyses, and code reviews fit comfortably in this range. You get ~97% accuracy at standard pricing — the best cost-performance ratio available.

Use the 1M window strategically. When you do need it — full codebase analysis, large document comparison — front-load critical information in the prompt. The model retrieves earlier content more reliably than content buried deep in a million-token context.

Enable Tool Search for agent workflows. If you're connecting to multiple MCP servers, the 47% token reduction isn't optional — it's essential for keeping costs manageable at scale.

Leverage input caching aggressively. At $0.25 per million tokens (compared to $2.50 standard), cached inputs offer 90% savings. Structure your prompts with static system instructions and tool definitions in cacheable positions.

Right-size your reasoning effort. Don't default to xhigh. Start at medium, evaluate quality, and only increase when the task genuinely requires deeper reasoning.

For computer use, match image detail to the task. The original setting supports up to 10.24M pixels, while high supports 2.56M. Higher resolution costs more tokens — use it only when the task demands pixel-level precision.

The Bigger Picture

GPT-5.4 marks a meaningful shift in what a single AI model can do. It's not just smarter — it can act. It can see your screen and click buttons. It can find the right tool from hundreds of options without you specifying which one. It can hold an entire project in memory while planning multi-step workflows.

But the practical reality is more nuanced than the marketing. The million-token context degrades significantly past 272K tokens. Computer use, while groundbreaking at 75% success, still fails one in four tasks. And the pricing structure penalizes careless usage. The developers and teams who will get the most from GPT-5.4 are those who understand these boundaries and optimize within them — using the right context size for the job, enabling tool search to cut waste, and matching reasoning effort to task complexity. The model is genuinely powerful. Using it well requires understanding where that power has limits.