On March 5, 2026, OpenAI released GPT-5.4 in ChatGPT, the API and Codex, calling it its most capable and efficient frontier model for professional work. GPT-5.4 was OpenAI's first general-purpose model with native computer-use capabilities, it supports up to one million tokens of context in the API, and it introduced tool search, a way for agents to work with very large tool catalogs without paying for every tool definition on every request. It also came with a price increase over GPT-5.2, which OpenAI justified with better results and fewer tokens per task.

Key Facts#

  • Release date: March 5, 2026. API models gpt-5.4 and gpt-5.4-pro, with gpt-5.4-mini and gpt-5.4-nano following on March 17, 2026.
  • Pricing per million tokens: gpt-5.4 at $2.50 input ($0.25 cached) and $15 output, up from $1.75 and $14 for GPT-5.2. gpt-5.4-pro costs $30 and $180. Batch and Flex run at half price, and Priority at double.
  • Context: up to one million tokens in the API. Microsoft's Azure documentation lists 1,050,000 tokens in total, with up to 922,000 input and 128,000 output tokens.
  • Computer use: OpenAI reports 75.0% on OSWorld-Verified, above GPT-5.2's 47.3% and the 72.4% human baseline cited by OpenAI.
  • Other vendor-reported results: 83.0% on the GDPval knowledge-work evaluation versus 70.9% for GPT-5.2, 57.7% on SWE-Bench Pro, and 82.7% on BrowseComp, with GPT-5.4 Pro at 89.3%.
  • Tool search: in a test with 36 MCP servers enabled, OpenAI says tool search cut total token usage by 47% at the same accuracy.
  • Safety: OpenAI treats GPT-5.4 as having "High" cyber capability under its Preparedness Framework and deployed it with corresponding protections.

What Happened#

GPT-5.4 merged two lines of work. OpenAI said it brought the coding strength of GPT-5.3-Codex into a mainline reasoning model and improved how that model handles tools, software environments and office documents. In ChatGPT it appeared as GPT-5.4 Thinking, replacing GPT-5.2 Thinking for paid users, with GPT-5.2 Thinking scheduled for retirement on June 5, 2026.

Computer use was the headline. OpenAI said GPT-5.4 can operate computers both by writing automation code with libraries such as Playwright and by issuing mouse and keyboard commands in response to screenshots. Developers can steer that behavior through developer messages, including custom confirmation policies that decide when the agent must ask before acting. OpenAI's SDK release the same day promoted the computer tool to a generally available API feature.

Tool search addressed a scaling problem that every agent builder hits. Normally, every tool definition is sent with every request, which can add thousands or tens of thousands of tokens when an agent has access to many tools. With tool search, the model receives a lightweight list and looks up a tool's full definition only when it needs it. OpenAI tested this on 250 tasks from Scale's MCP Atlas benchmark with all 36 MCP servers enabled and reported a 47% reduction in total tokens with no loss of accuracy.

OpenAI also emphasized efficiency and accuracy. It called GPT-5.4 its most token-efficient reasoning model, using significantly fewer tokens than GPT-5.2 on the same problems, and its most factual model, with individual claims 33% less likely to be false than GPT-5.2's on prompts where users had flagged errors. On safety, OpenAI expanded the cyber safeguards introduced with GPT-5.3-Codex and published a new open-source evaluation of chain-of-thought controllability, reporting that GPT-5.4 has little ability to disguise its reasoning, which keeps reasoning monitoring useful as a safety tool.

Background#

OpenAI's GPT-5 family had evolved quickly since GPT-5 launched in August 2025, through GPT-5.1 in November and GPT-5.2 in December. In parallel, the Codex line produced specialized coding models, most recently GPT-5.3-Codex. GPT-5.4 reunited those tracks so developers no longer had to choose between the best coding model and the best general model.

The competitive backdrop explained the emphasis on agents. Anthropic's Claude Opus 4.6 and Google's Gemini 3.1 Pro had both launched in February 2026 with strong agentic and long-context results. GPT-5.4's one-million-token context and native computer use put OpenAI on equal footing with the capabilities rivals had been promoting.

Why It Matters for Developers#

Tool search is the most important feature for teams building agents on the Model Context Protocol. If you expose internal systems through servers built with the MCP C# SDK, the number of tools can grow quickly, and so does the per-request token bill. Tool search lets you keep a broad catalog without sending it every time. It is an OpenAI API feature, so check support in whatever client you use; in the meantime, designing smaller, well-described tool sets remains good practice, as covered in our guide to AI agent architecture patterns.

Computer use changes what agents can automate, especially legacy line-of-business applications without APIs. It also raises the stakes. An agent that clicks buttons in a real UI can make real mistakes, and content on screen can carry prompt injection. Use OpenAI's confirmation policies, run agents in isolated environments with least-privilege accounts, and log every action; our guide to responsible AI and LLM security covers these controls. For browser automation, Playwright's official .NET library is a natural fit with the model's code-driven approach.

Pricing needs fresh modeling. The per-token price rose, while OpenAI says total tokens per task fell. Whether your bill goes up or down depends on your workload, so measure cost per completed task rather than per token, using the approach in our guide to observability and cost control for LLM apps. Watch long prompts too: Codex counts requests above the standard 272,000-token window at twice the normal usage rate, and API list pricing is higher for prompts above that size.

What's Next#

OpenAI followed with GPT-5.4 mini and nano on March 17, 2026, then GPT-5.5 on April 23, and a GPT-5.6 family with Sol, Terra and Luna models on July 9. The GPT-6 generation began in September 2026. gpt-5.4 remains in OpenAI's model list and in Azure's catalog, where Microsoft lists computer use among its supported features.

The larger trend is clear: model releases increasingly arrive with platform features, such as tool search, computer tools and compaction, that change application architecture. Keep provider-specific features behind your own interfaces so you can adopt them without locking your codebase to one vendor.

Sources#