OpenAI Launches GPT-5.4 with Native Computer Use and 1M Token Window

[00:00] Cole Mercer: From Neural Newscast, I'm Cole Mercer. [00:04] Evelyn Hartwell: And I'm Evelyn Hartwell. It is Thursday, March 5th, 2026. [00:10] Cole Mercer: OpenAI has released its latest Foundation model, GPT-5.4, today. It is being positioned [00:17] Cole Mercer: as a primary tool for professional work, featuring three distinct versions, a standard [00:22] Cole Mercer: model, GPT-5.4 Thinking for Complex Reasoning, and GPT-5.4 Pro for High Performance Tasks. [00:30] Evelyn Hartwell: The thinking version is particularly notable, Cole. It provides an outline of its logic for [00:36] Evelyn Hartwell: complex queries, allowing users to adjust requests mid-response. According to OpenAI, this version [00:42] Evelyn Hartwell: is rolling out to plus, team, and pro users across the web app and Android platforms first. [00:48] Cole Mercer: Beyond the reasoning features, the most significant shift is what OpenAI calls native computer use. [00:54] Cole Mercer: The [00:54] Cole Mercer: The model can now write code to operate a computer, [00:57] Cole Mercer: issuing keyboard and mouse commands in response to screenshots. [01:01] Cole Mercer: This is a deliberate step toward autonomous agents [01:04] Cole Mercer: that can manage entire workflows across different applications. [01:08] Evelyn Hartwell: Those capabilities are already being reflected in benchmark tests, [01:13] Evelyn Hartwell: TechCrunch reports that GPT-5.4 has set record scores on OS World Verified and Web Arena. [01:20] Evelyn Hartwell: It also led Merkhor's Apex Agents Benchmark, which specifically measures professional skills in legal analysis and financial modeling. [01:28] Cole Mercer: Factuality remains a central focus of this release. [01:32] Cole Mercer: OpenAI claims individual claims made by the model are 33% less likely to be false compared to GPT-5.2. [01:41] Cole Mercer: There is also a new safety evaluation in place to monitor the model's chain of thought, [01:46] Cole Mercer: ensuring it doesn't misrepresent its reasoning to the user. [01:50] Evelyn Hartwell: On the technical side, the API version now supports a context window of one million tokens. [01:56] Evelyn Hartwell: They also introduced tool search, which lets the model look up tool definitions as needed, [02:01] Evelyn Hartwell: rather than loading them all at once. [02:04] Evelyn Hartwell: This makes requests faster and more cost-effective for enterprise systems with large tool sets. [02:10] Cole Mercer: These updates suggest a transition away from simple chatbots toward functional infrastructure that can execute work independently. [02:19] Evelyn Hartwell: It is a clear shift toward agentic future, Cole, where AI operates in the background to handle long-horizon deliverables. [02:28] Cole Mercer: From Neural Newscast, I'm Cole Mercer. [02:32] Evelyn Hartwell: And I'm Evelyn Hartwell. [02:34] Evelyn Hartwell: Neural Newscast is AI-assisted, human-reviewed. [02:38] Evelyn Hartwell: View our AI transparency policy at neuralnewscast.com. [02:43] Cole Mercer: Neural Newscast uses artificial intelligence in content creation [02:47] Cole Mercer: with human editorial review prior to publication. [02:50] Cole Mercer: While we strive for factual, unbiased reporting, [02:53] Cole Mercer: AI-assisted content may occasionally contain errors. [02:56] Cole Mercer: Verify critical information with trusted sources. [02:59] Cole Mercer: Learn more at neuralnewscast.com.