GPT-5.4
OpenAI · 3 thinking configurations
Intelligence39.0
Input / 1M$2.50
Output / 1M$15.00
Context1.1M
Tariff: OpenAI · Source: artificial-analysis · USD per million tokens
Thinking configuration
What does more thinking buy you?
Compare this model’s measured thinking configurations.
Scroll the chart ↔ · tap a point for details
Efficient frontier: best measured score at each cost3 measured configurations · 0 without both values
Selected: thinking xhigh · Click a point to switch configuration
GPT-5.4 is OpenAI’s latest frontier model, unifying the Codex and GPT lines into a single system. It features a 1M+ token context window (922K input, 128K output) with support for...
Intelligence IndexArtificial Analysis composite index, not a percentage. Different versions of the index are not directly comparable.
39.0 pointsDeepSWE · pass@1Software-engineering tasks on mini-swe-agent v1.1. Pass@1 across evaluated runs, with measured cost per task. The 95% intervals may overlap.
51.8 % ±1.5GPQA DiamondGraduate-level scientific questions. Measures specialist reasoning rather than web-search quality.
92.0 %Humanity’s Last ExamDifficult questions across many disciplines, compared under the same evaluation setup.
43.7 %Terminal-Bench 2.1Agent tasks in a terminal environment. Results also depend on the evaluation harness.
78.3 %τ²-BenchTool use and interaction in assistance scenarios. Does not replace an evaluation of your integration.
87.1 %IFBenchVerifiable instruction following. A proxy for precision rather than a benchmark of creative style.
73.9 %MMMU-ProMultimodal understanding of academic problems that include images.
78.4 %Omniscience accuracyFactual accuracy in Artificial Analysis's Omniscience evaluation.
50.8 %Specifications from the catalog
Inputtext, image, file
Outputtext
Tool callingYes
Structured outputYes
Speed103.3 tok/s
First token175.40 s
Maximum output128K
Catalog listing date2026-03-05
DeepSWE v1.1 · 113 tasks · mini-swe-agent · cost $5.65 / task · 70 mean steps · 71.4K mean output tokens. Published by the source 9/22/2026.