GPT-5.4 Launches: 1M Token Window, Native Computer Use, 75% on OSWorld
April 1, 2026 · 7 min read
TL;DR
OpenAI launched GPT-5.4 on March 5, 2026. Key numbers: 1M token context window (API), 75% on OSWorld-Verified (beats average human at 72.4%), 47% reduction in token consumption via Tool Search, 33% fewer hallucinations vs GPT-5.2. Available to ChatGPT Plus, Team, Pro, and Enterprise users. Use Happycapy for complex workflows that pair AI planning with execution.
GPT-5.4 is the most significant OpenAI model release since GPT-5.0. It ships three capabilities that didn't exist together in a single model before: a 1 million token context window, the ability to directly control a computer, and a merged reasoning-coding pipeline. Here is what changed, what it means, and who should upgrade.
The 1 Million Token Context Window
GPT-5.4's API supports 1,050,000 tokens — the largest context window OpenAI has shipped. In practice, this means the model can process in a single prompt:
- An entire large codebase (500K+ lines)
- A full legal contract archive for due diligence
- All earnings call transcripts from a company's history
- Multiple research papers simultaneously for synthesis
The standard ChatGPT interface provides 272,000 tokens — still larger than most competing models. API developers get the full 1M limit. This eliminates the need for manual chunking, retrieval-augmented generation (RAG) pipelines, or summarization workarounds for most use cases.
| Model | Context Window |
|---|---|
| GPT-5.4 (API) | 1,050,000 tokens |
| Claude Code (Anthropic) | 1,000,000 tokens |
| GPT-5.4 (ChatGPT) | 272,000 tokens |
| Cursor (AI editor) | 256,000 tokens |
| GPT-5.3 (previous) | 400,000 tokens |
| GitHub Copilot | 32K–128K tokens |
Native Computer Use: 75% on OSWorld
GPT-5.4 is the first general-purpose model to include native computer use — the ability to see your screen and move your mouse and keyboard. On the OSWorld-Verified benchmark, which measures a model's ability to complete real computer tasks:
- GPT-5.4: 75%
- Average human: 72.4%
- Previous models: under 40%
The model can navigate applications, edit spreadsheets, fill in web forms, manage files, and execute multi-step workflows on your actual computer. This is not a separate API (unlike Anthropic's Computer Use beta) — it is built into the model itself.
Practical applications: automated data entry across systems, end-to-end report generation, browser-based research workflows, and managing software tools that lack APIs.
Merged Reasoning + Coding
Previous GPT-5 releases required users to choose between the reasoning-optimized model and the coding-optimized model (GPT-5.3 Codex). GPT-5.4 merges both into a single model. You no longer need to context-switch between models for different task types.
On GDPval — OpenAI's internal benchmark for knowledge work — GPT-5.4 scored 83%. On Mercor's APEX-Agents benchmark for professional skills in law and finance, it ranked first among all tested models.
Tool Search: 47% Fewer Tokens, Same Accuracy
GPT-5.4 introduces Tool Search, a new routing mechanism that reduces token consumption by up to 47% on agentic workflows by intelligently selecting which tools to invoke. For developers building on the API, this directly reduces cost without sacrificing output quality.
Combined with a 33% reduction in hallucination rate vs GPT-5.2, these efficiency improvements make GPT-5.4 the most cost-effective GPT-5 model to run at scale.
Availability and Pricing
| Version | Who Gets It | Context |
|---|---|---|
| GPT-5.4 Thinking | Plus, Team, Pro users | 272K tokens |
| GPT-5.4 Pro | Pro and Enterprise | 1M tokens |
| GPT-5.4 API | Developers | 1,050,000 tokens |
GPT-5.2 Thinking will remain available in Legacy Models until June 5, 2026, then retire.
Plan Your AI Workflows
Use Happycapy to design complex AI automations, evaluate which model fits your task, and get expert guidance on the AI tools landscape.
Try Happycapy →Frequently Asked Questions
What is GPT-5.4 and when did it launch?
GPT-5.4 is OpenAI's latest flagship model, launched March 5, 2026. It features a 1 million token context window (API version), native computer use that scored 75% on OSWorld-Verified (beating human average of 72.4%), and merged reasoning + coding in one model.
How big is GPT-5.4's context window?
GPT-5.4 supports up to 1,050,000 tokens in the API version and Pro plan, and 272,000 tokens in the standard ChatGPT interface. This matches Claude Code's 1M token window as the largest available in a mainstream model.
What is native computer use in GPT-5.4?
Native computer use means GPT-5.4 can interpret your computer screen and directly control mouse and keyboard actions to navigate applications, edit documents, fill forms, and manage workflows—without separate automation software. It scored 75% on OSWorld-Verified, beating average human performance of 72.4%.
How does GPT-5.4 compare to Claude and Gemini?
GPT-5.4 matches Claude Code's 1M token context window and leads on OSWorld computer use benchmarks. Gemini 3.1 Pro has a 256K context window. For computer use and autonomous workflow automation, GPT-5.4 is the current leader among generally available models.
Sources
- • OpenAI GPT-5.4 launch announcement and documentation, March 5, 2026
- • OSWorld-Verified benchmark results, March 2026
- • OpenAI GDPval and Mercor APEX-Agents benchmark reports, 2026
- • OpenAI Developer documentation: context window and pricing, 2026