OpenAI Releases GPT-5.4 Across ChatGPT, API, and Codex With Major Tool-Use Gains

What OpenAI announced

OpenAI published Introducing GPT-5.4 on March 5, 2026 and released GPT-5.4 in ChatGPT (as GPT-5.4 Thinking), the API, and Codex. The company also launched GPT-5.4 Pro for users who want maximum performance on complex workflows. In practical terms, OpenAI is positioning GPT-5.4 as a single model family for professional tasks that combine reasoning, coding, tool calling, and long-horizon execution.

The Hacker News discussion quickly became one of the highest-activity AI threads in the cycle, reaching 816 points and 658 comments at crawl time. That level of discussion reflects how much developer attention is now focused on task completion quality and tool orchestration, not just static benchmark scores.

Headline metrics and capability deltas

OpenAI reports several benchmark gains versus GPT-5.2, including GDPval (83.0% wins or ties vs 70.9%), SWE-Bench Pro (57.7% vs 55.6%), OSWorld-Verified (75.0% vs 47.3%), Toolathlon (54.6% vs 46.3%), and BrowseComp (82.7% vs 65.8%). GPT-5.4 Pro is reported at 89.3% on BrowseComp. OpenAI also states GPT-5.4 is more factual than GPT-5.2 on internal error-flagged prompts, with individual claims 33% less likely to be false and full responses 18% less likely to contain any errors.

These are vendor-reported figures, but they are specific enough to indicate where OpenAI is investing: structured work products, reliable agent behavior, and stronger web-assisted retrieval.

Computer use, context window, and cost dynamics

A major technical shift is native computer-use support in a general-purpose model, plus expanded tool routing. OpenAI says GPT-5.4 supports up to a 1M-token context window in Codex and API workflows, and introduces tool search so agents can retrieve tool definitions on demand rather than loading full tool catalogs up front. In OpenAI's MCP Atlas example, this reduced token usage by 47% at the same reported accuracy.

On pricing, OpenAI lists GPT-5.4 at $2.50 per million input tokens, $0.25 per million cached input tokens, and $15 per million output tokens. GPT-5.4 Pro is listed at $30 input and $180 output per million tokens. For teams comparing deployment options, the practical question is whether higher per-token rates are offset by lower total token consumption and fewer correction turns.

Sources: OpenAI release, Hacker News thread.