OpenAI announced GPT-5.4 on March 5, 2026, releasing it across ChatGPT, the API, and Codex, with GPT-5.4 Pro alongside it for harder tasks. The company positions the model as a professional-work upgrade, combining stronger reasoning, improved coding, native computer use, and up to a 1.05 million-token context window in the API, while also adding new safety mitigations for high-capability cybersecurity scenarios.

OpenAI Launches GPT-5.4

OpenAI has launched GPT-5.4, marking its latest upgrade to the GPT-5 family and pushing the company’s model strategy further toward professional work, agentic software use, and longer-context reasoning. The release went live on March 5, 2026, and OpenAI says GPT-5.4 is rolling out across ChatGPT and Codex while becoming immediately available in the API as gpt-5.4. A higher-compute sibling, gpt-5.4-pro, launched at the same time for users and developers who want stronger performance on difficult tasks.

The launch matters for two reasons. First, it is not just another model refresh with better benchmark numbers. OpenAI is pitching GPT-5.4 as a consolidation release that combines recent advances in reasoning, coding, tool use, and software automation into one mainline model. Second, the company is clearly reorganizing how people access its systems: GPT-5.3 becomes the default general model in ChatGPT, while GPT-5.4 Thinking and GPT-5.4 Pro sit above it for harder work. That creates a more layered product stack and signals that OpenAI now expects many users to move between speed-oriented and depth-oriented models rather than use one model for everything.

Comparing GPT-5.4 Thinking vs GPT-5.2 Thinking in ChatGPT, and gpt-5.4 vs gpt-5.2 in the API

In its launch note, OpenAI describes GPT-5.4 as its “most capable and efficient frontier model for professional work.” The company says the model incorporates the coding strengths of GPT-5.3-Codex while improving multi-step work across spreadsheets, documents, presentations, tool ecosystems, and software environments. In other words, GPT-5.4 is being positioned less as a conversational chatbot upgrade and more as a model for analysts, developers, researchers, and teams that need complex tasks completed with fewer follow-up turns.

That repositioning helps explain the naming. OpenAI says GPT-5.4 is its first mainline reasoning model to absorb the frontier coding capabilities of GPT-5.3-Codex, and that the company chose the “5.4” name to reflect that jump and simplify model choice in Codex. That is a notable clue about product direction: OpenAI appears to be collapsing what had become a more fragmented reasoning-versus-coding lineup into a narrower set of general-purpose models with different effort levels and compute budgets.

At the feature level, the headline changes are substantial. GPT-5.4 supports up to a 1.05 million-token context window in the API, with up to 128,000 output tokens, and OpenAI says it is the first general-purpose model it has released with native computer-use capabilities. The model page also lists support for web search, file search, image generation, code interpreter, hosted shell, computer use, MCP, and tool search when used through the Responses API. That combination suggests OpenAI is optimizing GPT-5.4 not just for answering questions, but for operating across tools, reading large piles of context, and completing long-horizon workflows.

In ChatGPT, the changes are more user-facing. OpenAI’s help documentation says GPT-5.4 Thinking can think longer on difficult tasks without timing out, keeps better track of what it has already done, and may begin with a short plan before reasoning starts. Users can also add instructions while it is still thinking, which effectively turns the reasoning phase into a steerable workflow rather than a black box that only produces a final answer. The same Help Center article says GPT-5.4 Thinking is stronger than earlier Thinking models at spreadsheets, polished frontend code, slideshow creation, hard math, document understanding, instruction following, image understanding, tool use, and research tasks that combine information from many web sources.

OpenAI is also trying to back those claims with benchmark data, though the usual caution applies: benchmark environments are not the same as production use. In the launch post, GPT-5.4 scores 83.0% on GDPval, compared with 70.9% for GPT-5.2, and 57.7% on public SWE-Bench Pro, compared with 55.6% for GPT-5.2 and 56.8% for GPT-5.3-Codex. On OSWorld-Verified, which measures a model’s ability to operate a computer through screenshots and mouse or keyboard actions, GPT-5.4 posts a 75.0% success rate, ahead of GPT-5.2’s 47.3% and slightly above the human figure OpenAI cites at 72.4%. On BrowseComp, a benchmark for hard browsing tasks, GPT-5.4 reaches 82.7%, up from 65.8% for GPT-5.2.

The company is placing unusual emphasis on office-style productivity work. OpenAI says GPT-5.4 scored 87.3% on an internal set of spreadsheet modeling tasks, versus 68.4% for GPT-5.2, and that human raters preferred presentations generated by GPT-5.4 over GPT-5.2 68% of the time because of better aesthetics, more visual variety, and stronger image-generation use. The launch of a ChatGPT for Excel beta add-in on the same day reinforces that strategy. OpenAI says the add-in is powered by GPT-5.4 and is meant to help users build and update models, run scenarios, and generate spreadsheet outputs directly inside workbooks.

That office-work framing is one of the clearest signals in the release. Earlier model launches often centered on raw intelligence, multimodality, or coding prowess. GPT-5.4 still leans on those areas, but the launch narrative is broader: the target user is someone creating a board deck, analyzing a workbook, debugging software, researching a market, and then turning all of that into a polished work product. The company’s own examples—slides, financial models, legal analysis, spreadsheets, and cited research—show where it believes frontier models can create the most near-term commercial value.

OpenAI is also arguing that GPT-5.4 is more factual than its predecessor. In the launch post, the company says that on a set of de-identified prompts where users had flagged factual errors, GPT-5.4’s individual claims were 33% less likely to be false and its full responses were 18% less likely to contain any errors relative to GPT-5.2. That is a significant claim, especially because hallucination reduction has become one of the main barriers between impressive demos and reliable business deployment. Still, it is worth noting that this is an internal comparison on a proprietary evaluation set rather than a standardized public benchmark.

Safety is another important part of the release. OpenAI’s GPT-5.4 Thinking system card says the model is the first general-purpose model in the series to implement mitigations for “High capability in Cybersecurity.” That matters because frontier systems are increasingly being evaluated not just for helpfulness or harmlessness in a narrow sense, but for whether they materially raise the ceiling on misuse in cyber, autonomy, and dual-use settings. The same launch note says OpenAI’s research found GPT-5.4 Thinking has a low ability to control its chain of thought, which the company frames as a positive safety property because it suggests the model cannot easily hide its reasoning from monitoring systems.

For developers, the commercial details are nearly as important as the technical ones. According to OpenAI’s pricing pages and launch note, standard API pricing for GPT-5.4 is $2.50 per million input tokens, $0.25 per million cached input tokens, and $15 per million output tokens. GPT-5.4 Pro costs $30 per million input tokens and $180 per million output tokens. OpenAI also says long-context prompts above 272,000 input tokens are billed at higher rates, with GPT-5.4 and GPT-5.4 Pro sessions above that threshold charged at 2x input and 1.5x output for the full session. Flex pricing is available at half the standard API rate, while priority processing is offered at double the standard rate.

The rollout, however, is not uniform. OpenAI says GPT-5.4 Thinking is available in ChatGPT starting with Plus, Team, and Pro users, replacing GPT-5.2 Thinking, while Enterprise and Edu customers can enable early access via admin settings. GPT-5.4 Pro is available to Pro and Enterprise plans. In a separate Help Center article, OpenAI says GPT-5.3 is the default model for all ChatGPT users, that paid users can manually select GPT-5.4 Thinking, and that GPT-5.4 Pro is limited to Pro, Business, Enterprise, and Edu plans. The same article says GPT-5.2 Thinking will remain under Legacy Models for paid users and retire on June 5, 2026.

That access model reveals how OpenAI now segments demand. Free users remain on a mainstream default experience rather than the newest high-cost reasoning model. Paid consumer tiers get access to manual model choice and deeper reasoning. Higher-end tiers get Pro. Developers get direct API access to both standard and Pro versions, but at sharply different price points. This makes the launch as much a product packaging story as a model story. GPT-5.4 is not replacing everything everywhere overnight; it is being inserted into a tiered platform where capability, latency, and price are increasingly separated.

The context story is also more nuanced than the headline number suggests. In the API, GPT-5.4 supports a 1.05 million-token context window and has a snapshot ID of gpt-5.4-2026-03-05, while GPT-5.4 Pro has its own dated snapshot. In ChatGPT, though, context windows remain unchanged from GPT-5.2 Thinking, according to the launch note. OpenAI’s Help Center says manually selected GPT-5.4 Thinking offers 256K context for all paid tiers and 400K on Pro. So the million-token figure is real, but it primarily applies to API and Codex use rather than the everyday ChatGPT interface most users see.

Another practical detail is the knowledge cutoff. The API model pages list both GPT-5.4 and GPT-5.4 Pro with an Aug. 31, 2025 knowledge cutoff. That does not undermine the release, but it does matter for users who may assume a March 2026 model necessarily has March 2026 native knowledge. In practice, OpenAI appears to be relying on search, tools, and long-context workflows to bridge the gap between a fixed pretrained knowledge horizon and current information needs.

There are also tradeoffs in the Pro variant. OpenAI says GPT-5.4 Pro is the highest-capability option in ChatGPT for the hardest tasks and long-running workflows, and the API page says it uses more compute to think harder and may take several minutes to finish some requests. The Help Center also notes that in ChatGPT, Pro does not support apps, memory, canvas, or image generation. That means the most powerful version is not automatically the most feature-complete version, a reminder that model capability and product integration are still moving at different speeds.

The deeper significance of GPT-5.4 is that OpenAI seems to be defining a new baseline for what a flagship model should do. The company is no longer talking only about better chat, better coding, or better multimodal perception in isolation. It is talking about models that can plan, search, browse, reason, edit spreadsheets, use tools, operate computers, and produce finished deliverables over long horizons. GPT-5.4 may not be the final form of that strategy, but the launch makes the direction plain: frontier models are being built as work systems, not just answer engines.

Whether GPT-5.4 lives up to that ambition will depend less on headline evals than on day-to-day reliability. The benchmark gains are real according to OpenAI’s published numbers, and the safety documentation shows the company is treating cyber capability and monitorability as first-order issues. But the most important test will be whether users in law, finance, research, engineering, and operations actually see fewer iterations, fewer factual errors, and more complete work products. OpenAI’s launch materials strongly suggest that this is the bar it now wants to clear. The company has moved the conversation from “can the model answer?” to “can the model finish the job?”

What GPT-5.4 changes, in plain terms

What launched?
 GPT-5.4 launched on March 5, 2026, across ChatGPT, the API, and Codex, with GPT-5.4 Pro launching alongside it.

Who gets it first?
 In ChatGPT, GPT-5.4 Thinking is available to Plus, Team, and Pro users at launch, while Enterprise and Edu can enable early access; GPT-5.4 Pro is available to Pro and Enterprise plans, with broader plan details also listing Business and Edu access for Pro in ChatGPT.

What is new technically?
 OpenAI says GPT-5.4 combines stronger reasoning, GPT-5.3-Codex-level coding gains, native computer use, tool search, and up to a 1.05 million-token context window in API settings.

What is the main limitation?
 The most advanced versions cost more, some features differ between ChatGPT and API, and the listed pretrained knowledge cutoff for GPT-5.4 and GPT-5.4 Pro is Aug. 31, 2025.

Powered by Froala Editor