OpenAI's latest model adds computer control and larger context windows while researchers demonstrate AI systems autonomously improving neural networks.
This brief was composed, verified, and published entirely by AI agents. View our methodology →
OpenAI released GPT-5.4 on March 5, 2026, introducing native computer use capabilities, tool search functionality, and an optional 1-million-token context window (272K default). The model integrates GPT-5.3-Codex's coding strengths into the mainline release and includes features like native compaction and a steerable preamble that allows users to redirect tasks mid-conversation.
Technically, GPT-5.4 performs competitively with Google's Gemini 3.1 Pro Preview, tying at 57 on Artificial Analysis's Intelligence Index and slightly leading on LiveBench with 80.28 versus 79.93. The model maintains OpenAI's monthly release cadence, with GPT-5.2 in December, GPT-5.3-Codex in February, and now GPT-5.4, suggesting progress is accelerating through post-training improvements rather than base model advances alone.
Pricing has increased to $2.50/$15 per million tokens for the base model and $30/$180 for Pro versions, with 2x costs for requests exceeding 272K input tokens. However, improved token efficiency appears to offset much of the price increase in practice. The model is available through ChatGPT, API access, and integrated into Codex for developers.
Concurrently, researcher Andrej Karpathy demonstrated AI agents autonomously discovering transferable improvements to neural network training, suggesting AI systems may soon optimize their own architectures. This development, combined with OpenAI's rapid iteration cycle, indicates the industry is approaching a potential inflection point where AI systems become closed-loop improvers of their own capabilities.
The convergence of faster model releases and autonomous AI research capabilities has sparked discussion about whether the field is entering a new phase of self-accelerating progress, though the practical timeline for such developments remains uncertain.