GPT-5.4 Brings Steerable Reasoning and 1M Token Context
OpenAI has released GPT-5.4, its most capable frontier model for professional work, featuring three variants designed for different use cases. The standard GPT-5.4 offers improved efficiency and accuracy, while GPT-5.4 Thinking introduces steerable chain-of-thought reasoning, and GPT-5.4 Pro delivers maximum performance for complex tasks.
The model's most notable technical achievement is its 1 million token context window in the API version—the largest from OpenAI to date. This enables handling of extensive documents, large codebases, and multi-step research tasks without fragmentation.
Benchmark Performance
GPT-5.4 demonstrates significant improvements across professional and agentic benchmarks:
- WebArena-Verified: 67.3% success rate (up from 65.4% with GPT-5.2)
- Online-Mind2Web: 92.8% on screenshot-based browser tasks (up from 70.9%)
- GDPval: 83% on knowledge work tasks
- OSWorld-Verified: Record score for computer use
Beyond raw performance, the model is more efficient—producing 47% fewer tokens on benchmarks while maintaining accuracy, and reducing errors by 33% in individual claims and 18% overall.
Chain-of-Thought Controllability Findings
Alongside the model release, OpenAI published research on reasoning model transparency. The CoT-Control evaluation found that GPT-5.4 Thinking has extremely low ability to deliberately hide or obfuscate its reasoning—a feature the company frames as a safety benefit.
"Reasoning models struggle to control their chains of thought, and that's good," OpenAI noted. The low controllability reinforces monitorability as an AI safety safeguard, making it difficult for models to evade oversight through reasoning manipulation.
The GPT-5.4 Thinking variant allows users to view an upfront reasoning plan and intervene mid-response, adding instructions or redirecting before final output—reducing back-and-forth interactions.
Availability
GPT-5.4 is available now via ChatGPT Plus, Team, and Pro tiers, as well as the API (gpt-5.4 and gpt-5.4-pro). The GPT-5.4 Thinking variant replaces GPT-5.2 Thinking in the model picker, with the older model remaining available until June 5, 2026.