# GPT-5.4 ProExtra high

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

OpenAI

Canonical URL: https://benchize.com/models/gpt-5-4-pro-xhigh
Last updated: Thu, 22 Oct 2026 06:30:09 GMT

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

OpenAI

## Benchmarks

| Benchmark | Score | ± | Rank | Cost | Source |  |
| --- | --- | --- | --- | --- | --- | --- |
| [ARC-AGI-1](https://benchize.com/benchmarks/arc-prize%3Av1_Semi_Private.md) | 94.5% | — | 38 / 245 | 7.21usd per task | [arc prize · v1](https://arcprize.org/media/data/leaderboard/v1.json) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 0.945 \| score_fraction \| v1_Semi_Private \| Higher is better \| \| `cost_per_task` \| 7.21 \| usd_per_task \| v1_Semi_Private \| Lower is better \| ### Configuration evaluationSplit v1_Semi_Private harness Not reported modelGroup gpt-5-4-pro modelType CoT Reasoning effort xhigh sourceModelVersion gpt-5-4-pro-xhigh thinkingEnabled true ### Source [Original source](https://arcprize.org/media/data/leaderboard/v1.json) Source arc prize Original model label GPT-5.4 Pro (XHigh) Source version v1 Observed 2026-10-22T05:20:04.066Z Evidence SHA-256 `2bf68d22dafba37047d07ad833ebc0e8199c9aa2f9eeaed50c75c1b704462836` Source locator `/evaluations/133` |  |  |  |  |  |  |
| [ARC-AGI-2](https://benchize.com/benchmarks/arc-prize%3Av2_Semi_Private.md) | 83.33% | — | 36 / 243 | 16.41usd per task | [arc prize · v2](https://arcprize.org/media/data/leaderboard/v2.json) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `score` \| 0.8333 \| score_fraction \| v2_Semi_Private \| Higher is better \| \| `cost_per_task` \| 16.41 \| usd_per_task \| v2_Semi_Private \| Lower is better \| ### Configuration evaluationSplit v2_Semi_Private harness Not reported modelGroup gpt-5-4-pro modelType CoT Reasoning effort xhigh sourceModelVersion gpt-5-4-pro-xhigh thinkingEnabled true ### Source [Original source](https://arcprize.org/media/data/leaderboard/v2.json) Source arc prize Original model label GPT-5.4 Pro (XHigh) Source version v2 Observed 2026-10-22T05:20:04.306Z Evidence SHA-256 `4351eb696e78113b0459e9bb5f0258318b905efef34f473dcea51ce641b0d94d` Source locator `/evaluations/134` |  |  |  |  |  |  |
| [Apex](https://benchize.com/benchmarks/matharena%3Aapex--apex_2025.md) | 69.79% | 95% CI 60.6–78.98% | 3 / 31 | 8.65usd per problem | [matharena · apex--apex_2025](https://matharena.ai/competition_tables/apex--apex_2025) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 69.79 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 8.65 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition apex--apex_2025 predicted false Reasoning effort xhigh sourceModelName GPT-5.4-Pro (xhigh) sourceModelVersion openai/gpt-54-pro thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 60.6–78.98% confidenceInterval95 9.19 problems 12 ### Source [Original source](https://matharena.ai/competition_tables/apex--apex_2025) Source matharena Original model label GPT-5.4-Pro (xhigh) Source version apex--apex_2025 Observed 2026-10-22T05:30:13.849Z Evidence SHA-256 `20b28370d9d7f5a15a19fc9e4226c9857cdb07bc42b32ccec508e687b7b9f3fc` Source locator `/table/rows/2` |  |  |  |  |  |  |
| [ArXivMath 02/2026](https://benchize.com/benchmarks/matharena%3Aarxiv--february.md) | 75.78% | 95% CI 68.36–83.2% | 1 / 20 | 6.14usd per problem | [matharena · arxiv--february](https://matharena.ai/competition_tables/arxiv--february) |  |
| ### Original source values \| Metric \| Value \| Unit \| Scope \| Direction \| \| --- \| --- \| --- \| --- \| --- \| \| `accuracy` \| 75.78 \| percent \| competition \| Higher is better \| \| `cost_per_problem` \| 6.14 \| usd_per_problem \| competition \| Lower is better \| ### Configuration competition arxiv--february predicted false Reasoning effort xhigh sourceModelName GPT-5.4-Pro (xhigh) sourceModelVersion openai/gpt-54-pro thinkingEnabled true ### Reported statistics Uncertainty 95% confidence interval: 68.36–83.2% confidenceInterval95 7.42 problems 32 ### Source [Original source](https://matharena.ai/competition_tables/arxiv--february) Source matharena Original model label GPT-5.4-Pro (xhigh) Source version arxiv--february Observed 2026-10-22T05:30:08.096Z Evidence SHA-256 `3bc5b0a1eb91a0bdaa3d0942ebf82327a8664334b225ae3b2dc1d8b4f57efa51` Source locator `/table/rows/0` |  |  |  |  |  |  |

---
Cite as: GPT-5.4 ProExtra high — Benchize, https://benchize.com/models/gpt-5-4-pro-xhigh (retrieved 2026-10-22)