GPT-5.4 arrives as its rival fights the Pentagon — and a war for the future of AI intensifies
The timing was not accidental. With Anthropic still reeling from a Pentagon designation that triggered a lawsuit threat just days earlier, OpenAI unveiled its most capable model yet — a calculated strike into the markets its embattled rival has been courting most aggressively.
GPT-5.4, released Thursday across ChatGPT, the API, and Codex, is engineered for the professional workplace. It writes investment memos, builds financial models inside Excel and Google Sheets, browses the web with greater precision, and — for the first time in a general-purpose OpenAI model — operates computers natively, executing multi-step tasks across applications without human hand-holding. On OSWorld-Verified, an autonomous computer-use benchmark, it scored 75.0%, clearing the human baseline and leaving its predecessor's 47.3% in the dust.
The model is faster, cheaper to run, and more honest. OpenAI says it produces 33% fewer false claims than GPT-5.2 and consumes 47% fewer tokens in complex agentic workflows. On GDPval, a benchmark simulating professional knowledge work across 44 occupations, GPT-5.4 matched or outperformed industry professionals in 83% of tasks. On a simulation of junior investment banking analyst work, it scored 87.5% — up nearly 20 percentage points from GPT-5.2's 68.4%. Walleye Capital's head of Intelligence Solutions confirmed that internal finance and Excel assessments showed a 30 percentage point accuracy improvement over prior models.
For the financial sector — the exact territory Anthropic's Claude Opus 4.6 had been targeting since February — the message from OpenAI was unmistakable.
The other story unfolding Thursday was darker, and far more consequential.
Anthropic's dispute with the Pentagon had been simmering for weeks. The Department of Defense, under Defense Secretary Pete Hegseth, demanded unrestricted access to Anthropic's AI systems for government use. Anthropic refused, drawing firm lines against two applications: mass surveillance of American citizens and autonomous weapons operating without human oversight. Negotiations collapsed. Hegseth responded by designating Anthropic a "supply chain risk" — a classification ordinarily reserved for foreign adversaries and hostile state actors — and terminated the government relationship entirely.
Anthropic announced it will sue the Pentagon over the designation.
While those talks were deteriorating, OpenAI was negotiating its own agreement with the DoD — and succeeded. The result is a decisive asymmetry: one AI company now has a framework agreement with the federal government; the other has been branded a national security threat.
Public reaction cleaved sharply along that fault line. Evaluators tracking user sentiment found a 295% spike in ChatGPT uninstalls following the disclosure of OpenAI's Pentagon deal, with significant numbers of users migrating to Claude in protest. The optics of one AI company securing a military contract while another was sued for refusing one proved polarizing in ways that benchmark scores could not smooth over.
Users who stayed had substantive complaints of their own. Developers reported that despite OpenAI advertising a one-million-token context window, the Codex CLI continues to display a 258,000-token limit — an inconsistency the company has not publicly addressed. Others noted that GPT-5.4 shipped just two days after GPT-5.3, compounding residual frustration from users who felt that model was inferior to GPT-5.2. A pattern is emerging: releases arriving faster than trust can be rebuilt.
"Wait and see" has become the prevailing posture among skeptical developers, some of whom raised a more fundamental concern — that the competitive pressure driving this pace of releases may be crowding out the careful alignment and safety work the moment demands.
The two stories of March 5, 2026, are not separate. They are the same story told from two angles: what happens when the most consequential technology ever built becomes a prize that governments want to control and companies race to deploy. OpenAI signed with the Pentagon. Anthropic drew a line and got labeled an enemy for it.
GPT-5.4 is impressive. But the more durable question — who gets to decide how these systems are used, and against whom — was not settled by any benchmark on Thursday. It may be settled in court.
not investment advice