GPT-5.4 mini (xhigh): benchmark scores, price and ranking
GPT-5.4 mini (xhigh) from OpenAI ranks #531 of 743 LLMs on the LLMs Tiger aggregate benchmark score (46.3), based on 19 benchmark results captured up to 2026-10-13.
- Organisation
- OpenAI
- Type
- Fast / low-cost (flash) model
- Open weights
- No
- Released
- 2026-03-17 (source: epoch)
- Cheapest API price
- $0.750 in / $4.50 out per million tokens (via aihubmix, observed 2026-10-13T19:47Z)
- Aggregate rank
- #531 of 743 (score 46.3)
Relative to other models, GPT-5.4 mini (xhigh) is strongest on GPQA Diamond (#63 of 275) and weakest on LiveBench Reasoning (#16 of 17).
Across the 19 benchmarks where at least ten models are scored, GPT-5.4 mini (xhigh) is in the top 10% on 0 and in the top half on 7.
Its cheapest listed input price, $0.750 per million tokens, is lower than that of 40% of the 427 models with a listed price.
GPT-5.4 mini (xhigh) benchmark results
About GPT-5.4 mini (xhigh)
GPT-5.4 mini (xhigh) is a smaller, lower-cost model from OpenAI. It was released on 2026-03-17, according to Epoch AI. The cheapest API listing we track charges $0.750 per million input tokens and $4.50 per million output tokens, through aihubmix. We hold 19 benchmark results for it across 4 areas, and none is in the top tenth of its board.
Where GPT-5.4 mini (xhigh) stands
On our aggregate score, GPT-5.4 mini (xhigh) is #531 of 743 ranked models, in the lower half. Its score is 46.3; the scale is centred on 50, so it sits 3.7 points below the typical model. That beats 29% of the ranked field. Among the 88 ranked OpenAI models it is #69. The next one up is GPT-4.5 Preview (Feb 2025) at 46.6. The next one down is o3-mini-2025-01-31 (low) at 46.0.
GPT-5.4 mini (xhigh), benchmark by benchmark
GPQA Diamond puts GPT-5.4 mini (xhigh) at 86.9%, #63 of 275. That beats 77% of the models tested, but 8.9 points short of GPT-6 Astra (max) (95.8%). The results either side are Darwin-27B-Opus (86.9%) and Qwen3.5-122B-A10B (86.6%). It covers expert-level science questions. The board median is 71.7%, so GPT-5.4 mini (xhigh) is 15.2 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
On OTIS Mock AIME, GPT-5.4 mini (xhigh) reaches 88.9%. It ranks #45 of 189, ahead of 76% of the field, and 11.1 points behind Claude Fable 5.1 (max) at 100.0%. The results either side are Muse Spark (88.9%) and gpt-oss-120b (high) (88.9%). It covers competition maths (AIME 2024-25 style). The board median is 57.8%, so GPT-5.4 mini (xhigh) is 31.1 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
GPT-5.4 mini (xhigh) manages 40.8% on LMCA, good for #28 of 108. Claude Opus 5.5 (max) leads at 68.2%, 27.4 points further up, and 74% of models sit below. The results either side are DeepSeek v4 Pro (max) (41.2%) and GPT-5 (high) (40.0%). It covers language-model capability assessment. The board median is 29.2%, so GPT-5.4 mini (xhigh) is 11.6 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
Chess Puzzles puts GPT-5.4 mini (xhigh) at 24.0%, #34 of 122. That beats 72% of the models tested, but 48.0 points short of GPT-6 Astra (max) (72.0%). The results either side are grok-4.3 (high) (25.0%) and grok-4.20-0309-reasoning (24.0%). It covers chess tactics puzzles. The board median is 12.5%, so GPT-5.4 mini (xhigh) is 11.5 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
On CritPt, GPT-5.4 mini (xhigh) reaches 10.0%. It ranks #38 of 115, ahead of 67% of the field, and 22.3 points behind GPT-5.6 Sol (max) at 32.3%. The results either side are Muse Spark (11.3%) and Kimi K2.7 Code (10.0%). It covers unpublished research-level physics problems. The board median is 2.9%, so GPT-5.4 mini (xhigh) is 7.1 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
GPT-5.4 mini (xhigh) manages 49.9% on SciCode, good for #38 of 106. Claude Opus 5.5 (max) leads at 66.9%, 17.0 points further up, and 64% of models sit below. The results either side are DeepSeek V4 Flash 0731 (max) (49.9%) and Kimi K2.5 (49.0%). It covers scientific research code. The board median is 44.3%, so GPT-5.4 mini (xhigh) is 5.6 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
LiveBench Language is a weak spot. GPT-5.4 mini (xhigh) scores 71.0%, #15 of 17, behind 82% of the models measured. The leader, GPT-5.5 (xhigh), reaches 87.4%. It covers LiveBench release 2026_06_25. The board median is 77.9%, so GPT-5.4 mini (xhigh) is 7.0 points below it. Source: LiveBench 2026_06_25 official CSV, captured 2026-10-13.
LiveBench Mathematics is a weak spot. GPT-5.4 mini (xhigh) scores 78.5%, #15 of 17, behind 82% of the models measured. The leader, GPT-5.5 (xhigh), reaches 95.9%. It covers LiveBench release 2026_06_25. The board median is 88.2%, so GPT-5.4 mini (xhigh) is 9.8 points below it. Source: LiveBench 2026_06_25 official CSV, captured 2026-10-13.
MCP Atlas is a weak spot. GPT-5.4 mini (xhigh) scores 56.7%, #13 of 14, behind 86% of the models measured. The leader, Muse Spark, reaches 82.2%. It covers tool use across 1,000 MCP tasks. The board median is 70.4%, so GPT-5.4 mini (xhigh) is 13.7 points below it. Source: Scale Labs MCP Atlas leaderboard, captured 2026-10-13.
LiveBench Reasoning is a weak spot. GPT-5.4 mini (xhigh) scores 71.3%, #16 of 17, behind 88% of the models measured. The leader, claude-opus-4-8-xhigh-effort, reaches 89.7%. It covers LiveBench release 2026_06_25. The board median is 81.1%, so GPT-5.4 mini (xhigh) is 9.8 points below it. Source: LiveBench 2026_06_25 official CSV, captured 2026-10-13.
The full table above also lists the other 9 results, including boards with fewer than ten models, where a rank says little.
Strengths and weaknesses
Software engineering: uneven, ahead of 35% of the field on average across 3 benchmarks. Its best board here is SciCode (#38 of 106) and its weakest is LiveBench Coding (#14 of 17). Reasoning, maths and science: uneven, ahead of 44% of the field on average across 9 benchmarks. Its best board here is GPQA Diamond (#63 of 275) and its weakest is LiveBench Reasoning (#16 of 17). Agents and tool use: uneven, ahead of 35% of the field on average across 2 benchmarks. Its best board here is DTBench (#55 of 146) and its weakest is MCP Atlas (#13 of 14). General knowledge and chat: uneven, ahead of 43% of the field on average across 5 benchmarks. Its best board here is LMCA (#28 of 108) and its weakest is LiveBench Language (#15 of 17).
Price and value
The median input price across the 427 priced models is $0.500 per million tokens, so GPT-5.4 mini (xhigh) costs more than the typical listing. Models that score within 1.5 points or better and cost less: Qwen2.5-Coder-7B-Instruct ($0.010, 49.5), DeepSeek V4.1 Flash (high) ($0.010, 48.9), Qwen2.5-Coder-3B-Instruct ($0.010, 46.5). Similar scores cost more on amazon.nova-pro-v1:0 ($0.800) and Claude Haiku 4.5 (Thinking 32K) ($1.00).
Release history
The OpenAI model released before it was GPT-5.4 (xhigh) on 2026-03-05, which scores 56.9. The next one was GPT-5.4 nano (xhigh) on 2026-03-17, at 47.1.
Reading these numbers
The 19 scores on this page come from 4 sources: Epoch AI Benchmarking Hub (10), LiveBench 2026_06_25 official CSV (7), ARC Prize official leaderboard (1), Scale Labs MCP Atlas leaderboard (1). 9 of them come from an official leaderboard and 10 from datasets or third-party runs. The leaderboard also lists 6 other reasoning-effort settings of this model, which redirect here. This page shows the highest.
GPT-5.4 mini (xhigh) is also listed at 6 other reasoning-effort settings; this page shows the highest-effort row, which is the one the leaderboard keeps by default.
Models ranked nearby
- #528 DeepSeek Chat V2.5 (46.3)
- #529 qwen1.5-110b-chat (46.3)
- #530 Qwen2.5-3B (46.3)
- #532 mistral-small-2409 (46.3)
- #533 MAmmoTH2-8B-Plus (46.2)
- #534 Yi-34B (46.2)
More from OpenAI
- GPT-6 Astra (max) (#2, 64.0)
- GPT-6.1 Sol (max) (#5, 60.6)
- GPT-5.6 Sol (max) (#6, 60.1)
- GPT-6 Sol (max) (#7, 60.0)
- GPT-5.5 (xhigh) (#11, 58.0)
- GPT-5.6 Terra (max) (#12, 57.8)
Compare it on the live leaderboard → · How the aggregate score works