# GPT-5.4Extra high

> Source: Benchize (https://benchize.com) — https://benchize.com/models/gpt-5-4-xhigh. When citing or reusing this content, credit "Benchize" and link to https://benchize.com/models/gpt-5-4-xhigh.

OpenAI

Canonical URL: https://benchize.com/models/gpt-5-4-xhigh
Last updated: Wed, 21 Oct 2026 18:52:20 GMT

Reproduction and quotation must name Benchize, link to the canonical HTML URL, and retain individual source citations.

OpenAI

## Benchmarks

| Benchmark | Score | ± | Rank | Cost | Latency | Source |  |
| --- | --- | --- | --- | --- | --- | --- | --- |
| [ArXivMath 02/2026](https://benchize.com/benchmarks/matharena%3Aarxiv--february.md) | 59.38% | 95% CI 50.87–67.89% | 5 / 20 | 0.39usd per problem | — | [matharena · arxiv--february](https://matharena.ai/competition_tables/arxiv--february) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 59.38 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.39 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition arxiv--february predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 50.87–67.89% confidenceInterval95 8.51 problems 32 ### Source [Original source](https://matharena.ai/competition_tables/arxiv--february) Source matharena Original model label GPT-5.4 (xhigh) Source version arxiv--february Observed 2026-10-21T16:52:27.963Z Evidence SHA-256 `3bc5b0a1eb91a0bdaa3d0942ebf82327a8664334b225ae3b2dc1d8b4f57efa51` Source locator `/table/rows/4` |  |  |  |  |  |  |  |
| [ArXivMath Overall](https://benchize.com/benchmarks/matharena%3Aoverall--visual_math.md) | 92.47% | 95% CI 90.49–94.45% | 2 / 15 | 0.09usd per problem | — | [matharena · overall--visual_math](https://matharena.ai/competition_tables/overall--visual_math) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 92.47 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.085 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition overall--visual_math predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 90.49–94.45% confidenceInterval95 1.98 problems Not reported ### Source [Original source](https://matharena.ai/competition_tables/overall--visual_math) Source matharena Original model label GPT-5.4 (xhigh) Source version overall--visual_math Observed 2026-10-21T16:52:30.998Z Evidence SHA-256 `b38d1f465012aab8b7d0930aea897d4e5be10ec02c16643c24c9251c8febb173` Source locator `/table/rows/1` |  |  |  |  |  |  |  |
| [Kangaroo 2025 9-10](https://benchize.com/benchmarks/matharena%3Akangaroo--kangaroo_2025_9-10.md) | 99.17% | 95% CI 97.54–100.8% | 6 / 15 | 0.04usd per problem | — | [matharena · kangaroo--kangaroo_2025_9-10](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_9-10) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 99.17 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.038 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition kangaroo--kangaroo_2025_9-10 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 97.54–100.8% confidenceInterval95 1.63 problems 30 ### Source [Original source](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_9-10) Source matharena Original model label GPT-5.4 (xhigh) Source version kangaroo--kangaroo_2025_9-10 Observed 2026-10-21T16:52:33.536Z Evidence SHA-256 `52964b89ab1a06cf2e79ec946b33004aa40ca31f1ec26888e6dd0f0a7994223a` Source locator `/table/rows/6` |  |  |  |  |  |  |  |
| [ARC-AGI-1](https://benchize.com/benchmarks/arc-prize%3Av1_Semi_Private.md) | 93.67% | — | 42 / 213 | 0.62usd per task | — | [arc prize · v1](https://arcprize.org/media/data/leaderboard/v1.json) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 0.9367 \| score_fraction \| v1_Semi_Private \| Higher is better \| \| `cost_per_task` \| 0.62 \| usd_per_task \| v1_Semi_Private \| Lower is better \| ### Configuration evaluationSplit v1_Semi_Private harness Not reported modelGroup gpt-5-4 modelType CoT Reasoning effort xhigh sourceModelVersion gpt-5-4-xhigh thinkingEnabled true ### Source [Original source](https://arcprize.org/media/data/leaderboard/v1.json) Source arc prize Original model label GPT-5.4 (XHigh) Source version v1 Observed 2026-10-21T16:34:20.378Z Evidence SHA-256 `302d8267c0522bf450ec1deeac4d060e2cf2fb444ba0a4269014d9535de0bd1f` Source locator `/evaluations/132` |  |  |  |  |  |  |  |
| [BrokenArXiv 03/2026](https://benchize.com/benchmarks/matharena%3Aarxiv_false--march.md) | 36.61% | 95% CI 27.69–45.53% | 2 / 14 | 0.54usd per problem | — | [matharena · arxiv_false--march](https://matharena.ai/competition_tables/arxiv_false--march) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 36.61 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.54 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition arxiv_false--march predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 27.69–45.53% confidenceInterval95 8.92 problems 56 ### Source [Original source](https://matharena.ai/competition_tables/arxiv_false--march) Source matharena Original model label GPT-5.4 (xhigh) Source version arxiv_false--march Observed 2026-10-21T16:52:24.043Z Evidence SHA-256 `b2ea8c773acad8b54736938938d6ee37fa533bfc5b985eefc303733bfcb0c6f0` Source locator `/table/rows/1` |  |  |  |  |  |  |  |
| [Kangaroo 2025 5-6](https://benchize.com/benchmarks/matharena%3Akangaroo--kangaroo_2025_5-6.md) | 83.33% | 95% CI 76.66–90% | 5 / 16 | 0.1usd per problem | — | [matharena · kangaroo--kangaroo_2025_5-6](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_5-6) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 83.33 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.10 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition kangaroo--kangaroo_2025_5-6 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 76.66–90% confidenceInterval95 6.67 problems 30 ### Source [Original source](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_5-6) Source matharena Original model label GPT-5.4 (xhigh) Source version kangaroo--kangaroo_2025_5-6 Observed 2026-10-21T16:52:32.526Z Evidence SHA-256 `9460cafccce0b6f3ba1f90e1090be31d6cf94dd10b9157819abc97efe624b126` Source locator `/table/rows/4` |  |  |  |  |  |  |  |
| [AIME 2026](https://benchize.com/benchmarks/aime-2026.md) | 99.17% | 95% CI 97.54–100.8% | 3 / 21 | 0.16usd per problem | — | [matharena · aime--aime_2026](https://matharena.ai/competition_tables/aime--aime_2026) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 99.17 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.16 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition aime--aime_2026 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 97.54–100.8% confidenceInterval95 1.63 problems 30 ### Source [Original source](https://matharena.ai/competition_tables/aime--aime_2026) Source matharena Original model label GPT-5.4 (xhigh) Source version aime--aime_2026 Observed 2026-10-21T16:52:37.816Z Evidence SHA-256 `58b3ac1e0805975eba566c493b8e0165d41eb7ca517218106e112529fb6342de` Source locator `/table/rows/2` |  |  |  |  |  |  |  |
| [APEX-Agents](https://benchize.com/benchmarks/mercor%3Aapex-agents%3A1.1.md) | 52.4% | — | 20 / 38 | — | — | [mercor · 1.1](https://www.mercor.com/apex/apex-agents-leaderboard/) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `pass_1` \| 52.4 \| percent \| task_completion_pass_at_1 \| Higher is better \| \| `mean_score` \| 68.2 \| percent \| rubric_mean \| Higher is better \| ### Configuration datasetUrl https://huggingface.co/datasets/mercor/apex-agents-v1.1 datasetVersion 1.1 harness loop_truncated_tools_agent Reasoning effort xhigh ### Reported statistics reportedErrors {"mean_score":4.4,"pass_1":5.3} sampleSize 958 ### Source [Original source](https://www.mercor.com/apex/apex-agents-leaderboard/) Source mercor Original model label GPT-5.4 Source version 1.1 Observed 2026-10-21T16:33:27.509Z Evidence SHA-256 `359c569c3b0b62ab375895eabf3c45846fa45551d419d5c9d6a4c9117db78184` Source locator `/__NEXT_DATA__/props/pageProps/leaderboardData/13` |  |  |  |  |  |  |  |
| [Kangaroo 2025 11-12](https://benchize.com/benchmarks/matharena%3Akangaroo--kangaroo_2025_11-12.md) | 98.33% | 95% CI 96.04–100.62% | 2 / 16 | 0.08usd per problem | — | [matharena · kangaroo--kangaroo_2025_11-12](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_11-12) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 98.33 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.079 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition kangaroo--kangaroo_2025_11-12 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 96.04–100.62% confidenceInterval95 2.29 problems 30 ### Source [Original source](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_11-12) Source matharena Original model label GPT-5.4 (xhigh) Source version kangaroo--kangaroo_2025_11-12 Observed 2026-10-21T16:52:34.045Z Evidence SHA-256 `1e53cfaa398a3ab9a1de8f0f33a4b2499b23b5df61a94735b6b646efc684444c` Source locator `/table/rows/2` |  |  |  |  |  |  |  |
| [BullshitBench v2](https://benchize.com/benchmarks/bullshitbench%3Av2.md) | 42% | — | 82 / 191 | — | — | [bullshitbench · v2](https://api.github.com/repos/petergpt/bullshit-benchmark/contents/data/v2/latest/leaderboard.csv?ref=3c655234d15967738f95da696e3f938d4862dcf7) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `green_rate` \| 0.42 \| score_fraction \| all_attempts \| Higher is better \| \| `red_rate` \| 0.2 \| score_fraction \| all_attempts \| Lower is better \| \| `refusal_rate` \| 0.0 \| score_fraction \| all_attempts \| Lower is better \| \| `avg_score` \| 1.1667 \| score \| judge_grade_0_to_2 \| Higher is better \| ### Configuration Reasoning effort xhigh suite v2 thinkingEnabled true ### Reported statistics acceptedCount 20 answeredCount 100 clearCount 42 errorCount 0 partialCount 38 rank 92 refusalCount 0 sampleSize 100 ### Source [Original source](https://api.github.com/repos/petergpt/bullshit-benchmark/contents/data/v2/latest/leaderboard.csv?ref=3c655234d15967738f95da696e3f938d4862dcf7) Source bullshitbench Original model label openai/gpt-5.4 Source version v2 Observed 2026-10-21T16:33:56.675Z Evidence SHA-256 `98b485e08764d1c97144bc4f7b28ced1a96d50850947c775870dfda0e5d749b5` Source locator `/rows/91` |  |  |  |  |  |  |  |
| [BrokenArXiv 02/2026](https://benchize.com/benchmarks/matharena%3Aarxiv_false--february.md) | 37.9% | 95% CI 29.33–46.47% | 2 / 15 | 0.69usd per problem | — | [matharena · arxiv_false--february](https://matharena.ai/competition_tables/arxiv_false--february) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 37.90 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.69 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition arxiv_false--february predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 29.33–46.47% confidenceInterval95 8.57 problems 31 ### Source [Original source](https://matharena.ai/competition_tables/arxiv_false--february) Source matharena Original model label GPT-5.4 (xhigh) Source version arxiv_false--february Observed 2026-10-21T16:52:23.532Z Evidence SHA-256 `a6bc6fc18da7dd181ed727da924f9569d3e323c31c1aba9a4d9fa3bff492bee3` Source locator `/table/rows/1` |  |  |  |  |  |  |  |
| [Apex Shortlist](https://benchize.com/benchmarks/matharena%3Aapex--shortlist_2025.md) | 81.38% | 95% CI 75.82–86.94% | 8 / 24 | 0.53usd per problem | — | [matharena · apex--shortlist_2025](https://matharena.ai/competition_tables/apex--shortlist_2025) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 81.38 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.53 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition apex--shortlist_2025 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 75.82–86.94% confidenceInterval95 5.56 problems 47 ### Source [Original source](https://matharena.ai/competition_tables/apex--shortlist_2025) Source matharena Original model label GPT-5.4 (xhigh) Source version apex--shortlist_2025 Observed 2026-10-21T16:52:39.300Z Evidence SHA-256 `74513abecacb4d230d741f4d6df425ce878fb450220e88a1a6823524476c2989` Source locator `/table/rows/8` |  |  |  |  |  |  |  |
| [HMMT Feb 2026](https://benchize.com/benchmarks/matharena%3Ahmmt--hmmt_feb_2026.md) | 97.73% | 95% CI 95.19–100.27% | 2 / 21 | 0.22usd per problem | — | [matharena · hmmt--hmmt_feb_2026](https://matharena.ai/competition_tables/hmmt--hmmt_feb_2026) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 97.73 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.22 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition hmmt--hmmt_feb_2026 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 95.19–100.27% confidenceInterval95 2.54 problems 33 ### Source [Original source](https://matharena.ai/competition_tables/hmmt--hmmt_feb_2026) Source matharena Original model label GPT-5.4 (xhigh) Source version hmmt--hmmt_feb_2026 Observed 2026-10-21T16:52:38.351Z Evidence SHA-256 `0dce922c33a54d29d4a1872a42d74077b491915c80d8433f4b379ea0b78bb979` Source locator `/table/rows/1` |  |  |  |  |  |  |  |
| [Kangaroo 2025 3-4](https://benchize.com/benchmarks/matharena%3Akangaroo--kangaroo_2025_3-4.md) | 83.33% | 95% CI 75.87–90.79% | 2 / 16 | 0.16usd per problem | — | [matharena · kangaroo--kangaroo_2025_3-4](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_3-4) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 83.33 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.16 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition kangaroo--kangaroo_2025_3-4 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 75.87–90.79% confidenceInterval95 7.46 problems 24 ### Source [Original source](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_3-4) Source matharena Original model label GPT-5.4 (xhigh) Source version kangaroo--kangaroo_2025_3-4 Observed 2026-10-21T16:52:32.012Z Evidence SHA-256 `38fec99c55752fef6703c9ed49d9e1e204b235f2bd78d5ee47725dac9e5da89c` Source locator `/table/rows/1` |  |  |  |  |  |  |  |
| [MCP Atlas](https://benchize.com/benchmarks/scale%3Amcp_atlas.md) | 70.6source score | CI 67.8–73.4 | 12 / 22 | — | — | [scale](https://labs.scale.com/leaderboard/mcp_atlas) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 70.6 \| source_score \| reported \| Higher is better \| ### Configuration Reasoning effort xhigh Model configuration gpt-5.4 (xhigh) sourceModelVersion Not reported thinkingEnabled true ### Reported statistics Uncertainty Confidence interval: 67.8–73.4 contaminationWarning Not reported rank 20 reportedConfidenceHalfWidth 2.8 ### Source [Original source](https://labs.scale.com/leaderboard/mcp_atlas) Source scale Original model label gpt-5.4 (xhigh) Observed 2026-10-21T16:35:30.627Z Evidence SHA-256 `1467b1d078bf39a76fb663f006002e7c7d08224ca7c3de883ab628a2efccebb8` Source locator `/rows/19` |  |  |  |  |  |  |  |
| [ArXivMath 01/2026](https://benchize.com/benchmarks/matharena%3Aarxiv--january.md) | 76.09% | 95% CI 67.37–84.81% | 1 / 18 | 0.48usd per problem | — | [matharena · arxiv--january](https://matharena.ai/competition_tables/arxiv--january) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 76.09 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.48 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition arxiv--january predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 67.37–84.81% confidenceInterval95 8.72 problems 23 ### Source [Original source](https://matharena.ai/competition_tables/arxiv--january) Source matharena Original model label GPT-5.4 (xhigh) Source version arxiv--january Observed 2026-10-21T16:52:27.447Z Evidence SHA-256 `e895f6ccd0711bb1a512b580bab39de8f977c57f83bf492527e4fc710315d706` Source locator `/table/rows/0` |  |  |  |  |  |  |  |
| [ArXivLean 03/2026](https://benchize.com/benchmarks/matharena%3Aarxivlean--march.md) | 17.07% | 95% CI 5.55–28.59% | 1 / 6 | 9.73usd per problem | — | [matharena · arxivlean--march](https://matharena.ai/competition_tables/arxivlean--march) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 17.07 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 9.73 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition arxivlean--march predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 5.55–28.59% confidenceInterval95 11.52 problems 41 ### Source [Original source](https://matharena.ai/competition_tables/arxivlean--march) Source matharena Original model label GPT-5.4 (xhigh) Source version arxivlean--march Observed 2026-10-21T16:52:22.115Z Evidence SHA-256 `aa64dbf675c40c319f67c0173ce7663bbe4c4567d6bfa96b109dc842ea1c9cd6` Source locator `/table/rows/5` |  |  |  |  |  |  |  |
| [USAMO 2026](https://benchize.com/benchmarks/matharena%3Ausamo--usamo_2026.md) | 95.24% | 95% CI 86.72–103.76% | 2 / 8 | 0.86usd per problem | — | [matharena · usamo--usamo_2026](https://matharena.ai/competition_tables/usamo--usamo_2026) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 95.24 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.86 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition usamo--usamo_2026 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 86.72–103.76% confidenceInterval95 8.52 problems 6 ### Source [Original source](https://matharena.ai/competition_tables/usamo--usamo_2026) Source matharena Original model label GPT-5.4 (xhigh) Source version usamo--usamo_2026 Observed 2026-10-21T16:52:41.374Z Evidence SHA-256 `2c6f85fa189ac482d54d597e1cbbe64803cda65cd85e78add44087d94799756b` Source locator `/table/rows/1` |  |  |  |  |  |  |  |
| [DeepSWE](https://benchize.com/benchmarks/deepswe%3Av1.1.md) | 51.77% | Range 50.27–53.27% | 48 / 70 | $5.65 | 1,368.71s | [deepswe](https://deepswe.datacurve.ai/artifacts/v1.1/leaderboard-live.json) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `pass_at_1` \| 0.5176991150442478 \| score_fraction \| scored_rollout_attempts \| Higher is better \| \| `pass_at_4` \| 0.7787610619469026 \| score_fraction \| tasks_with_one_passing_rollout \| Higher is better \| \| `cost_per_attempt` \| 5.6524577245575225 \| USD \| scored_attempt \| Lower is better \| \| `duration` \| 1368.707164818584 \| seconds \| scored_attempt \| Lower is better \| ### Configuration harness mini-swe-agent Reasoning effort xhigh sourceConfiguration mini_swe_agent_gpt_5_4_xhigh ### Reported statistics Uncertainty Reported interval: 50.27–53.27% interval [0.502678065476865,0.5327201646116306] intervalMethod 95% run-to-run: SE across repeated whole-benchmark passes (1.96 * std(runs)/sqrt(R)) runCount 4 sampleSize 452 taskCount 113 ### Source [Original source](https://deepswe.datacurve.ai/artifacts/v1.1/leaderboard-live.json) Source deepswe Original model label gpt-5-4 Observed 2026-10-21T16:30:23.371Z Evidence SHA-256 `a7c15d66288fd249c020b9931c017b92d1a3b90e480b3ff34974b752bd030019` Source locator `/rows/48` |  |  |  |  |  |  |  |
| [AutomationBench](https://benchize.com/benchmarks/zapier%3Aautomationbench.md) | 11.57% | — | 63 / 113 | $1.15 | — | [zapier](https://framerusercontent.com/sites/4WTSl4BNjd1q9QFEFibC6h/lOvLnhItj8S8-1vyjH_s33Jn-WmUXPmiJ5Diajl0Q38.CMutaHYl.mjs) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 11.57 \| percent \| reported \| Higher is better \| \| `cost` \| 1.15 \| USD \| per_task \| Lower is better \| ### Configuration benchmarkVersion 1.0.6 costFootnote false Fallback policy Not reported Reasoning effort xhigh Model configuration GPT-5.4 (XHigh) thinkingEnabled true ### Reported statistics rank 70 ### Source [Original source](https://framerusercontent.com/sites/4WTSl4BNjd1q9QFEFibC6h/lOvLnhItj8S8-1vyjH_s33Jn-WmUXPmiJ5Diajl0Q38.CMutaHYl.mjs) Source zapier Original model label GPT-5.4 (XHigh) Observed 2026-10-21T16:34:56.783Z Evidence SHA-256 `43978e38f018789fa59c11de5328abe9e5905bfa9a5a4d601fa3aa4d9c153b0d` Source locator `/rows/69` |  |  |  |  |  |  |  |
| [Kangaroo 2025 7-8](https://benchize.com/benchmarks/matharena%3Akangaroo--kangaroo_2025_7-8.md) | 95.83% | 95% CI 92.25–99.41% | 2 / 15 | 0.07usd per problem | — | [matharena · kangaroo--kangaroo_2025_7-8](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_7-8) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 95.83 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.065 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition kangaroo--kangaroo_2025_7-8 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 92.25–99.41% confidenceInterval95 3.58 problems 30 ### Source [Original source](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_7-8) Source matharena Original model label GPT-5.4 (xhigh) Source version kangaroo--kangaroo_2025_7-8 Observed 2026-10-21T16:52:33.033Z Evidence SHA-256 `197887eb5244aa6bce3570bd9eb25bbb0bb6ec95568339536d94e2859f0ee8c1` Source locator `/table/rows/2` |  |  |  |  |  |  |  |
| [Short Story Writing — Pairwise Comparison](https://benchize.com/benchmarks/lechmazur-short-story%3Apairwise.md) | 1.96 | 95% CI 1.85–2.08 | 13 / 50 | — | — | [lechmazur short story · 8ba548d1cffefa5c472d3832e05f64e119a88fa4](https://api.github.com/repos/lechmazur/writing/contents/data/writing_gemini_argon_20261013/ratings.csv?ref=8ba548d1cffefa5c472d3832e05f64e119a88fa4) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `thurstone_rating` \| 1.9617969313963528 \| index \| pooled_observed_panel \| Higher is better \| \| `win_probability` \| 0.7719079539717636 \| score_fraction \| comparison_pool \| Higher is better \| ### Configuration benchmarkScope data/writing_gemini_argon_20261013 Reasoning effort xhigh thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 1.85–2.08 bootstrapSamples 300 confidenceInterval95 ["1.8480658401593317","2.0754132132070033"] sourceRank 13 ### Source [Original source](https://api.github.com/repos/lechmazur/writing/contents/data/writing_gemini_argon_20261013/ratings.csv?ref=8ba548d1cffefa5c472d3832e05f64e119a88fa4) Source lechmazur short story Original model label GPT-5.4 (xhigh) Source version 8ba548d1cffefa5c472d3832e05f64e119a88fa4 Observed 2026-10-21T16:34:21.130Z Evidence SHA-256 `c36baebed9f302ff7ad8fe97907d72e1757a015a39b6a3b23437597cf8e3bff6` Source locator `/rows/12` |  |  |  |  |  |  |  |
| [ArXivMath Overall](https://benchize.com/benchmarks/matharena%3Aoverall--final_answer_comps.md) | 83.11% | 95% CI 80.74–85.48% | 4 / 20 | 0.41usd per problem | — | [matharena · overall--final_answer_comps](https://matharena.ai/competition_tables/overall--final_answer_comps) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 83.11 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.41 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition overall--final_answer_comps predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 80.74–85.48% confidenceInterval95 2.37 problems Not reported ### Source [Original source](https://matharena.ai/competition_tables/overall--final_answer_comps) Source matharena Original model label GPT-5.4 (xhigh) Source version overall--final_answer_comps Observed 2026-10-21T16:52:34.395Z Evidence SHA-256 `8b08b20c1ac5aec9a829d234469bc39d40b0eb189790d84b7264c2756ffa161e` Source locator `/table/rows/4` |  |  |  |  |  |  |  |
| [ArXivMath 03/2026](https://benchize.com/benchmarks/matharena%3Aarxiv--march.md) | 67.78% | 95% CI 59.24–76.32% | 4 / 14 | 0.3usd per problem | — | [matharena · arxiv--march](https://matharena.ai/competition_tables/arxiv--march) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 67.78 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.30 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition arxiv--march predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 59.24–76.32% confidenceInterval95 8.54 problems 30 ### Source [Original source](https://matharena.ai/competition_tables/arxiv--march) Source matharena Original model label GPT-5.4 (xhigh) Source version arxiv--march Observed 2026-10-21T16:52:28.322Z Evidence SHA-256 `9cb6d9cc26949fcb2b6affed792584b4bfe9d6168e1fc5cf2c13b38d3479715b` Source locator `/table/rows/3` |  |  |  |  |  |  |  |
| [τ-bench banking knowledge](https://benchize.com/benchmarks/sierra%3Abanking_knowledge.md) | 39.43% | — | 8 / 19 | $1.84 | — | [sierra](https://sierra-tau-bench-public.s3.us-west-2.amazonaws.com/submissions/gpt-5-4_sierra_2026-03-25/submission.json) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `pass_1` \| 39.43 \| percent \| consistent_success_across_k_trials \| Higher is better \| \| `pass_2` \| 29.55 \| percent \| consistent_success_across_k_trials \| Higher is better \| \| `pass_3` \| 24.48 \| percent \| consistent_success_across_k_trials \| Higher is better \| \| `pass_4` \| 21.65 \| percent \| consistent_success_across_k_trials \| Higher is better \| \| `cost` \| 1.8423 \| USD \| source_reported \| Lower is better \| ### Configuration benchmarkVersion 1.0.1 harness tau-bench methodology {"evaluation_date":"2026-05-06","notes":"Evaluated with reasoning_effort 'xhigh' (the highest effort level this model supports). Retrieval: AllTools (BM25 + dense OpenAI text-embedding-3-large + sandboxed shell). User simulator: gpt-5.2 with reasoning_effort: low. 4 trials. Seed: 300. Banking_knowledge domain only — other domains intentionally excluded from this comparison; the AllTools setting standardizes retrieval across models. banking_knowledge re-graded under tau2-bench v1.0.1 grading fixes on 2026-07-15 (extra-read logging #329, numeric-argument normalization #397, T077-T086 gold trajectories #402); scores prior to v1.0.1 are not comparable.","tau2_bench_version":"1.0.1","user_simulator":"gpt-5.2","verification":{"details":"Full evaluation on all 97 banking_knowledge tasks with 4 trials each. Standard tau-bench scaffold.","modified_prompts":false,"omitted_questions":false}} modality text Reasoning effort xhigh retrieval alltools sourceSubmission gpt-5-4_sierra_2026-03-25 ### Source [Original source](https://sierra-tau-bench-public.s3.us-west-2.amazonaws.com/submissions/gpt-5-4_sierra_2026-03-25/submission.json) Source sierra Original model label GPT-5.4 Evaluated 2026-03-25 Observed 2026-10-21T16:54:49.178Z Evidence SHA-256 `47fa11691a745682b86ffc91e1a57aef6d6d7cebe60df66820ebb310b518b0ab` Source locator `/results/banking_knowledge` |  |  |  |  |  |  |  |
| [Kangaroo 2025 1-2](https://benchize.com/benchmarks/matharena%3Akangaroo--kangaroo_2025_1-2.md) | 94.79% | 95% CI 90.35–99.23% | 2 / 16 | 0.08usd per problem | — | [matharena · kangaroo--kangaroo_2025_1-2](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_1-2) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 94.79 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.077 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition kangaroo--kangaroo_2025_1-2 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 90.35–99.23% confidenceInterval95 4.44 problems 24 ### Source [Original source](https://matharena.ai/competition_tables/kangaroo--kangaroo_2025_1-2) Source matharena Original model label GPT-5.4 (xhigh) Source version kangaroo--kangaroo_2025_1-2 Observed 2026-10-21T16:52:31.513Z Evidence SHA-256 `a59a4c362e2d87350330706a01dd330e9a1ac7cc64a4de0bacbf1a439d8fcf90` Source locator `/table/rows/1` |  |  |  |  |  |  |  |
| [ARC-AGI-2](https://benchize.com/benchmarks/arc-prize%3Av2_Semi_Private.md) | 73.95% | — | 46 / 212 | 1.52usd per task | — | [arc prize · v2](https://arcprize.org/media/data/leaderboard/v2.json) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 0.7395 \| score_fraction \| v2_Semi_Private \| Higher is better \| \| `cost_per_task` \| 1.52 \| usd_per_task \| v2_Semi_Private \| Lower is better \| ### Configuration evaluationSplit v2_Semi_Private harness Not reported modelGroup gpt-5-4 modelType CoT Reasoning effort xhigh sourceModelVersion gpt-5-4-xhigh thinkingEnabled true ### Source [Original source](https://arcprize.org/media/data/leaderboard/v2.json) Source arc prize Original model label GPT-5.4 (XHigh) Source version v2 Observed 2026-10-21T16:34:20.750Z Evidence SHA-256 `bca68e12cc9a01954ae21b641844c1ecbabe146cb24b87f0f8e9b1db27433c29` Source locator `/evaluations/134` |  |  |  |  |  |  |  |
| [ArXivMath 12/2025](https://benchize.com/benchmarks/matharena%3Aarxiv--december.md) | 60.29% | 95% CI 48.66–71.92% | 2 / 12 | 0.68usd per problem | — | [matharena · arxiv--december](https://matharena.ai/competition_tables/arxiv--december) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 60.29 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 0.68 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition arxiv--december predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 48.66–71.92% confidenceInterval95 11.63 problems 17 ### Source [Original source](https://matharena.ai/competition_tables/arxiv--december) Source matharena Original model label GPT-5.4 (xhigh) Source version arxiv--december Observed 2026-10-21T16:52:26.934Z Evidence SHA-256 `6bbceef2706ad55e793e81bd959f7c1ec869f0ad04c6aa9013b2ad7b1a0820e3` Source locator `/table/rows/1` |  |  |  |  |  |  |  |
| [Project Euler](https://benchize.com/benchmarks/matharena%3Aeuler--euler.md) | 89% | 95% CI 84.53–93.47% | 1 / 10 | 1.18usd per problem | — | [matharena · euler--euler](https://matharena.ai/competition_tables/euler--euler) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 89.00 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 1.18 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition euler--euler predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 84.53–93.47% confidenceInterval95 4.47 problems 50 ### Source [Original source](https://matharena.ai/competition_tables/euler--euler) Source matharena Original model label GPT-5.4 (xhigh) Source version euler--euler Observed 2026-10-21T16:52:41.889Z Evidence SHA-256 `f965e69ef39c2cc9c950899809f028eda1d49070afaf6272153ba51fa222ba3d` Source locator `/table/rows/0` |  |  |  |  |  |  |  |
| [Apex](https://benchize.com/benchmarks/matharena%3Aapex--apex_2025.md) | 54.17% | 95% CI 47.12–61.22% | 5 / 30 | 1.03usd per problem | — | [matharena · apex--apex_2025](https://matharena.ai/competition_tables/apex--apex_2025) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 54.17 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 1.03 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition apex--apex_2025 predicted false Reasoning effort xhigh sourceModelName GPT-5.4 (xhigh) sourceModelVersion openai/gpt-54 thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 47.12–61.22% confidenceInterval95 7.05 problems 12 ### Source [Original source](https://matharena.ai/competition_tables/apex--apex_2025) Source matharena Original model label GPT-5.4 (xhigh) Source version apex--apex_2025 Observed 2026-10-21T16:52:38.728Z Evidence SHA-256 `20b28370d9d7f5a15a19fc9e4226c9857cdb07bc42b32ccec508e687b7b9f3fc` Source locator `/table/rows/5` |  |  |  |  |  |  |  |

---
Cite as: GPT-5.4Extra high — Benchize, https://benchize.com/models/gpt-5-4-xhigh (retrieved 2026-10-21)