GPT-5.4
Latest GPT-5.4 flagship chat model with stronger reasoning and accuracy.
About GPT-5.4
GPT-5.4 was built for the actual work that happens inside offices — financial modeling, legal analysis, complex codebases, and multi-step document workflows — rather than for chasing narrow benchmarks. That strategic shift shows in the numbers: it matched or outperformed human professionals in 83% of head-to-head comparisons, and developers have called its coding output "flawless," with some declaring it the definitive choice for complex software engineering work. Native computer-use capabilities let it operate browsers and desktop apps directly, and it scored above the human baseline on UI interaction tasks. The 1.05 million token context window handles large codebases and lengthy legal documents in a single pass, though you need to configure it explicitly — the default is 272K. Where GPT-5.4 falls short is nuance: it tends to interpret requests too literally, missing the intent behind ambiguous prompts in ways that Claude handles more naturally. Writing personality is another common frustration, with verbose follow-up suggestions that can feel mechanical. For structured professional tasks where thoroughness and tool integration matter more than prose feel, it is the strongest model in the GPT-5 line prior to the release of GPT-5.5.
Best for
- Complex software engineering — refactoring, multi-file codebases, and extended coding workflows
- Financial modeling and legal analysis — professional knowledge work where thoroughness outweighs speed
- Multi-step agentic workflows across documents, spreadsheets, web apps, and tool chains with native computer-use support
- Long-context analysis — processing large codebases or multi-document research using the 1M token window
- Writing and editing for clarity — fact assembly, logical organization, and content editing while preserving author voice
Specs & capabilities
How GPT-5.4 stacks up — intelligence, speed, context, and modalities.
Intelligence
High
Speed
Fast
Context window
1,050,000 tokens
Knowledge cutoff
August 31, 2025
Frequently asked questions
What does GPT-5.4 cost?
Input is $2.50 per million tokens, output is $15.00 per million tokens. Cached input drops to $0.25 per million. Note that prompts over 272K tokens trigger 2x input and 1.5x output charges for the entire session, and the model tends to generate more output tokens than average, so real-world costs can be higher than the per-token rate implies.
How large is the context window?
The maximum is 1.05 million tokens with up to 128K output tokens. However, the default context window is 272K tokens — you must explicitly configure the model to use extended context. Sessions exceeding 272K are billed at a premium rate.
What is GPT-5.4 best at?
Coding, structured professional work (financial modeling, legal analysis, spreadsheet tasks), multi-step agentic workflows, and long-context document analysis. It scored 83% on professional comparison benchmarks and 87.3% on investment banking modeling tasks.
What are its main weaknesses?
It takes prompts too literally and can miss the underlying intent of ambiguous requests — an area where Claude Opus tends to do better. Its writing personality is frequently described as verbose and mechanical, with unnecessary follow-up suggestions. It can also stop short in very long agentic workflows.
How does it compare to GPT-5.2?
GPT-5.4 is a meaningful upgrade: professional benchmark scores improved from 70.9% to 83%, investment banking modeling from 68.4% to 87.3%, and the model gained native computer-use capabilities entirely absent in GPT-5.2. It is more expensive, however.
Has GPT-5.4 been superseded?
GPT-5.5 was released in June 2026 with better token efficiency, improved long-context reasoning, and stronger agentic autonomy. GPT-5.4 remains available and is a practical option for cost-conscious users who do not need those specific improvements.