GPT-5.4 nano

Leaderboards

Where GPT-5.4 nano sits among the top 20 in each category.

By task

Scored from votes on each kind of prompt alone, on the same 1,000-point scale.

Test results

Every test GPT-5.4 nano has a published result on, out of 100%, with its rank among the models that took it.

• GPQA

  #84 / 256

  Graduate-level science questions written to be hard to look up.

  82.8%

  Verified

• Humanity's Last Exam

  #64 / 108

  Humanity's Last Exam (HLE) is a multi-modal academic benchmark with 2,500 questions across mathematics, humanities, and natural sciences, designed to test LLM capabilities at the frontier of human knowledge with unambiguous, verifiable solutions

  24.3%

  Self-reported · In the text score· In the vision score

• MMMU-Pro

  #47 / 73

  College-level questions that need both an image and text to answer.

  66.1%

  Self-reported · In the vision score

• SWE-Bench Pro

  #50 / 61

  Harder, longer software engineering tasks from real codebases.

  52.4%

  Self-reported · In the coding score

• Terminal-Bench 2.0

  #42 / 54

  Getting real work done in a command-line terminal.

  46.3%

  Self-reported

• Toolathlon

  #37 / 42

  Tool Decathlon is a comprehensive benchmark for evaluating AI agents' ability to use multiple tools across diverse task categories. It measures proficiency in tool selection, sequencing, and execution across ten different tool-use scenarios.

  35.5%

  Self-reported · In the agents score

• Multi-Challenge

  #40 / 41

  MultiChallenge is a realistic multi-turn conversation evaluation benchmark that challenges frontier LLMs across four key categories: instruction retention (maintaining instructions throughout conversations), inference memory (recalling and connecting details from previous turns), reliable versioned editing (adapting to evolving instructions during collaborative editing), and self-coherence (avoiding contradictions in responses). The benchmark evaluates models on sustained, contextually complex dialogues across diverse topics including travel planning, technical documentation, and professional communication.

  18.8%

  Verified

• LiveBench

  #30 / 38

  LiveBench is a challenging, contamination-limited LLM benchmark that addresses test set contamination by releasing new questions monthly based on recently-released datasets, arXiv papers, news articles, and IMDb movie synopses. It comprises tasks across math, coding, reasoning, language, instruction following, and data analysis with verifiable, objective ground-truth answers.

  70.1%

  Self-reported · In the text score

• MCP Atlas

  #36 / 38

  MCP Atlas is a benchmark for evaluating AI models on scaled tool use capabilities, measuring how well models can coordinate and utilize multiple tools across complex multi-step tasks.

  56.1%

  Self-reported

• Tau2 Telecom

  #15 / 36

  τ²-Bench telecom domain evaluates conversational agents in a dual-control environment modeled as a Dec-POMDP, where both agent and user use tools in shared telecommunications troubleshooting scenarios that test coordination and communication capabilities.

  92.5%

  Self-reported

• MRCR v2 (8-needle)

  #18 / 32

  MRCR v2 (8-needle) is a variant of the Multi-Round Coreference Resolution benchmark that includes 8 needle items to retrieve from long contexts. This tests models' ability to simultaneously track and reason about multiple pieces of information across extended conversations.

  33.1%

  Verified · In the text score

• ERQA

  #1 / 29

  Embodied Reasoning Question Answering benchmark consisting of 400 multiple-choice visual questions across spatial reasoning, trajectory reasoning, action reasoning, state estimation, and multi-view reasoning for evaluating AI capabilities in physical world interactions

  3300.0%

  Verified

• Finance Agent v2

  #19 / 28

  Finance Agent v2 is an agentic financial-analysis benchmark from Vals that evaluates models on real-world finance workflows, measuring their ability to retrieve and reason over financial documents, perform multi-step calculations, and produce accurate analyses.

  38.2%

  Self-reported

• OSWorld-Verified

  #28 / 28

  OSWorld-Verified is a verified subset of OSWorld, a scalable real computer environment for multimodal agents supporting task setup, execution-based evaluation, and interactive learning across Ubuntu, Windows, and macOS.

  39.0%

  Self-reported

• ScreenSpot Pro

  #1 / 27

  ScreenSpot-Pro is a novel GUI grounding benchmark designed to rigorously evaluate the grounding capabilities of multimodal large language models (MLLMs) in professional high-resolution computing environments. The benchmark comprises 1,581 instructions across 23 applications spanning 5 industries and 3 operating systems, featuring authentic high-resolution images from professional domains with expert annotations. Unlike previous benchmarks that focus on cropped screenshots in consumer applications, ScreenSpot-Pro addresses the complexity and diversity of real-world professional software scenarios, revealing significant performance gaps in current MLLM GUI perception capabilities.

  114.0%

  Verified

• HealthBench Hard

  #7 / 24

  A challenging variation of HealthBench that evaluates large language models' performance and safety in healthcare through 5,000 multi-turn conversations with particularly rigorous evaluation criteria validated by 262 physicians from 60 countries

  27.5%

  Verified

• FACTS Grounding

  #4 / 20

  A benchmark evaluating language models' ability to generate factually accurate and well-grounded responses based on long-form input context, comprising 1,719 examples with documents up to 32k tokens requiring detailed responses that are fully grounded in provided documents

  90.6%

  Verified

• OmniDocBench 1.5

  #13 / 20

  OmniDocBench 1.5 is a comprehensive benchmark for evaluating multimodal large language models on document understanding tasks, including OCR, document parsing, information extraction, and visual question answering across diverse document types. Lower Overall Edit Distance scores are better.

  75.8%

  Self-reported

• Graphwalks BFS <128k

  #6 / 13

  A graph reasoning benchmark that evaluates language models' ability to perform breadth-first search (BFS) operations on graphs with context length under 128k tokens, returning nodes reachable at specified depths.

  73.4%

  Self-reported

• HealthBench (hard)

  #1 / 13

  27.5%

  Unverified

• Graphwalks parents <128k

  #10 / 12

  A graph reasoning benchmark that evaluates language models' ability to find parent nodes in graphs with context length under 128k tokens, requiring understanding of graph structure and edge relationships.

  50.8%

  Self-reported

• GPQA (diamond)

  #9 / 10

  Graduate-level Google-Proof Q&A benchmark with paper-aligned full and diamond subsets.

  60.6%

  Unverified

• mrcr-v2 (8needle_32768_65536)

  #8 / 10

  16.2%

  Unverified

• mrcr-v2 (8needle_4096_8192)

  #10 / 10

  21.6%

  Unverified

• MultiChallenge (default)

  #9 / 9

  18.8%

  Unverified

• mrcr-v2 (8needle_upto_128k)

  #9 / 9

  15.4%

  Unverified

• facts-grounding (default)

  #4 / 8

  FACTS Grounding 1.0 — 860 public examples from Google DeepMind (CC-BY-4.0)

  90.6%

  Unverified

• nolima (hard_16k)

  #3 / 6

  20.4%

  Independent

• nolima (hard_4k)

  #2 / 5

  55.8%

  Independent

• ERQA (default)

  #2 / 3

  ERQA (Embodied Reasoning Question Answering) is a 400-question multimodal multiple-choice benchmark introduced with Google DeepMind's Gemini Robotics report. Questions span spatial, trajectory, action, state, multi-view, pointing, task, and other reasoning categories aimed at physical-world understanding.

  3425.0%

  Unverified

• independence-bench (default)

  #2 / 3

  67.7%

  Unverified

• ScreenSpot-Pro (default)

  #1 / 1

  ScreenSpot-Pro is a GUI grounding benchmark with 1,581 expert-annotated, high-resolution screenshots from 23 professional applications across 5 industries (Development, Creative, CAD, Scientific, Office) and 3 operating systems (Windows, macOS, Linux). Each task pairs a natural-language instruction with a screenshot and a ground-truth bounding box for the target UI element. Scoring follows the paper's center-point-in-bbox rule (Li et al. 2025, Section 5).

  114.0%

  Unverified