For the past two years, OpenAI has shipped reasoning models and standard models as separate products — GPT-5 for general tasks, o3 and o4 for deep thinking problems. That separation ends today. GPT-5.4 is a single unified model that does both, and it ships with the largest context window OpenAI has ever released: one million tokens.

The announcement, made on March 5, 2026, positions GPT-5.4 as OpenAI's most capable and efficient frontier model to date. It is available immediately in two versions — GPT-5.4 Pro and GPT-5.4 Thinking — with the API context window available to enterprise customers first.

What's Actually New in GPT-5.4

The headline number is accuracy. According to OpenAI's internal evaluations, individual claims made by GPT-5.4 are 33% less likely to be false than those made by GPT-5.2. Full responses are 18% less likely to contain any factual errors. For a model being used in legal research, financial analysis, and medical workflows, that gap is not cosmetic — it's the difference between a useful tool and a liability.

The second major improvement is token efficiency. GPT-5.4 uses substantially fewer tokens to solve the same reasoning problems as previous models. This is not just a convenience feature. Token usage maps directly to API cost, and for companies running GPT at scale, a 20-30% efficiency gain translates to meaningful savings in monthly infrastructure spend.

The third change is agentic capability. GPT-5.4 can now plan multi-step tasks upfront, revise its plan mid-execution based on new information, and handle significantly longer task chains without losing context. OpenAI describes this as "extended agency" — the model can be given a complex objective and pursue it over dozens of steps without requiring constant human re-prompting.

The 1 Million Token Context Window

This is the most practically significant technical spec in the release. One million tokens is roughly 750,000 words — enough to load an entire legal case file, a company's quarterly earnings history, or a software codebase with hundreds of files, and reason across all of it at once.

Previous OpenAI models topped out at 128,000 tokens in standard configurations. The jump to one million puts GPT-5.4 in direct competition with Google's Gemini 1.5 Pro, which also offers a 1M token context, but on a model that is now a full generation behind Gemini 2.0's architecture.

The API version of the 1M context window is rolling out to select enterprise customers first, with broader availability expected within 30 days. Consumer ChatGPT users on Pro plans will receive an expanded context window, though OpenAI has not confirmed whether the full 1M limit will be available at the consumer tier.

ChatGPT for Excel — Finance Workflows Enter the Chat

Alongside the model release, OpenAI announced a beta for ChatGPT integrated directly into Microsoft Excel. The add-in builds, updates, and analyzes financial models inside live workbooks. It uses GPT-5.4 Thinking specifically — the reasoning-optimized version — to handle multi-step formula construction, scenario modeling, and data interpretation.

The rollout is beginning for select business plan customers in the US, Canada, and Australia. Enterprise controls are included, meaning IT administrators can restrict the add-in from accessing data outside approved scopes. This matters: Excel is used for financial modeling that often includes sensitive, non-public information, and any AI integration that can read that data requires careful permission management.

Key Takeaway
GPT-5.4 ends the era of OpenAI maintaining separate reasoning and standard model lines. With a 1M token context window, 33% fewer factual errors, and built-in agentic planning, it is the most complete frontier model OpenAI has shipped — and the one most directly aimed at enterprise workflows that require long context, high accuracy, and multi-step task execution.

What This Means for the Market

OpenAI is not the only company shipping powerful models in March 2026. Google's Gemini 2.0 Ultra is in enterprise preview. Anthropic released Claude Sonnet 4.6 with its own 1M token context window in beta. DeepSeek V4 just matched GPT-5 benchmarks at a fraction of the inference cost.

The frontier model market is more competitive than at any point in AI's short history. GPT-5.4's advantage is not raw capability alone — it is the depth of OpenAI's distribution through ChatGPT's 500+ million users, Microsoft's enterprise agreements, and the API ecosystem built on GPT as a foundation. A 33% accuracy improvement matters more when it is delivered through channels that already touch a significant fraction of the world's knowledge workers.

What This Means For You

If you use ChatGPT on a Pro subscription, GPT-5.4 is the model you will be using for complex tasks starting this week. The accuracy improvements are real and noticeable — fewer hallucinated citations, fewer confident wrong answers, better handling of nuanced questions with many moving parts.

If you build on the OpenAI API, the token efficiency gains reduce your costs without any code changes on your end. The 1M context window opens up use cases that were previously impractical — document analysis, codebase Q&A, long-form research workflows.

And if you are watching the AI industry as a whole: the convergence of reasoning and standard capabilities into a single model is a structural shift. The next generation of AI products will not ask users to choose between fast and smart. They will simply be both.