o3 (high): benchmark scores, price and ranking

o3 (high) from OpenAI ranks #233 of 743 LLMs on the LLMs Tiger aggregate benchmark score (49.8), based on 17 benchmark results captured up to 2026-10-13.

Organisation

OpenAI

Type

Frontier model

Open weights

No

Released

2025-04-16 (source: curated)

Cheapest API price

$2.00 in / $8.00 out per million tokens (via azure, observed 2026-10-13T19:47Z)

Aggregate rank

#233 of 743 (score 49.8)

Relative to other models, o3 (high) is strongest on EnigmaEval (#3 of 24) and weakest on FrontierMath T1-3 (#50 of 68).

Across the 17 benchmarks where at least ten models are scored, o3 (high) is in the top 10% on 0 and in the top half on 14.

Its cheapest listed input price, $2.00 per million tokens, is lower than that of 18% of the 427 models with a listed price.

o3 (high) benchmark results

About o3 (high)

o3 (high) is a frontier model from OpenAI. It was released on 2025-04-16, according to the developer's announcement. The cheapest API listing we track charges $2.00 per million input tokens and $8.00 per million output tokens, through azure. We hold 17 benchmark results for it across 4 areas, and none is in the top tenth of its board.

Where o3 (high) stands

On our aggregate score, o3 (high) is #233 of 743 ranked models, around the middle of the table. Its score is 49.8; the scale is centred on 50, so it sits 0.2 points below the typical model. That beats 69% of the ranked field. Among the 88 ranked OpenAI models it is #33. The next one up is gpt-5.6-sol-xhigh (codex-harness) at 49.8. The next one down is gpt-4o-2024-11-20 at 49.6.

o3 (high), benchmark by benchmark

EnigmaEval puts o3 (high) at 11.9%, #3 of 24. That beats 88% of the models tested, but 6.3 points short of Gemini 3 Pro Preview (18.2%). The results either side are GPT-5.4 (xhigh) (16.0%) and o4-mini (high) (9.2%). It covers multi-step puzzle hunts. The board median is 2.9%, so o3 (high) is 9.0 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

On Chess Puzzles, o3 (high) reaches 34.0%. It ranks #20 of 122, ahead of 84% of the field, and 38.0 points behind GPT-6 Astra (max) at 72.0%. The results either side are Grok 4.5 (high) (36.0%) and DeepSeek V4 Flash 0731 (max) (33.0%). It covers chess tactics puzzles. The board median is 12.5%, so o3 (high) is 21.5 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

o3 (high) manages 52.4% on WeirdML, good for #27 of 109. GPT-6 Astra (max) leads at 93.3%, 40.9 points further up, and 75% of models sit below. The results either side are o4-mini (high) (52.6%) and Gemma 4 31B IT (52.3%). It covers unusual machine-learning tasks. The board median is 41.2%, so o3 (high) is 11.2 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

DTBench puts o3 (high) at 84.8%, #38 of 146. That beats 74% of the models tested, but 14.1 points short of Claude Opus 5.5 (max) (98.9%). The results either side are DeepSeek-V3.2 (Thinking; Novita) (85.6%) and mimo-v2.5-pro (84.5%). It covers dynamic tool-use benchmark. The board median is 68.8%, so o3 (high) is 16.0 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

On GSO-Bench, o3 (high) reaches 8.8%. It ranks #5 of 19, ahead of 74% of the field, and 31.4 points behind GPT-5.5 (xhigh) at 40.2%. The results either side are GPT-5.1 (high) (13.7%) and Claude Opus 4 (6.9%). It covers software performance optimisation. The board median is 4.9%, so o3 (high) is 3.9 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

o3 (high) manages 49.4% on SimpleQA Verified, good for #19 of 69. gemini-4-argon-high leads at 76.0%, 26.6 points further up, and 72% of models sit below. The results either side are GPT-5 (high) (50.1%) and Grok 4.6 (xhigh) (48.9%). It covers short factual questions. The board median is 41.0%, so o3 (high) is 8.4 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

CL-bench puts o3 (high) at 17.8%, #8 of 16. That beats 50% of the models tested, but 10.1 points short of GPT-5.4 (xhigh) (27.9%). The results either side are GLM-5 (18.7%) and Kimi K2 Thinking (Together) (17.6%). It covers learning new rules from context. The board median is 17.7%, so o3 (high) is 0.1 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

o3 (high) lands below the middle on CritPt: 1.4%, #65 of 115, with 56% of the field scoring higher. GPT-5.6 Sol (max) leads at 32.3%; the closest result above is gpt-oss-20b (high) at 1.4%. It covers unpublished research-level physics problems. The board median is 2.9%, so o3 (high) is 1.5 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

o3 (high) lands below the middle on ARC-AGI-2: 6.5%, #53 of 87, with 60% of the field scoring higher. GPT-6 Astra (max) leads at 95.0%; the closest result above is GPT-5.1 (Thinking, Medium) at 6.5%. It covers abstract reasoning, ARC Prize public eval set. The board median is 22.8%, so o3 (high) is 16.2 points below it. Source: ARC Prize official leaderboard, captured 2026-10-13.

o3 (high) lands below the middle on FrontierMath T1-3: 33.3%, #50 of 68, with 72% of the field scoring higher. GPT-6 Astra (max) leads at 93.7%; the closest result above is Inkling (xhigh) at 33.3%. It covers research-level maths, tiers 1-3. The board median is 55.6%, so o3 (high) is 22.3 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.

The full table above also lists the other 7 results, including boards with fewer than ten models, where a rank says little.

Strengths and weaknesses

Software engineering: solid, ahead of 71% of the field on average across 2 benchmarks. Its best board here is GSO-Bench (#5 of 19) and its weakest is ALE-Bench (#25 of 79). Reasoning, maths and science: solid, ahead of 59% of the field on average across 8 benchmarks. Its best board here is EnigmaEval (#3 of 24) and its weakest is FrontierMath T1-3 (#50 of 68). General knowledge and chat: solid, ahead of 69% of the field on average across 6 benchmarks. Its best board here is Chess Puzzles (#20 of 122) and its weakest is CL-bench (#8 of 16).

Price and value

The median input price across the 427 priced models is $0.500 per million tokens, so o3 (high) 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), Qwen3.7 Flash ($0.030, 50.5). Similar scores cost more on gpt-4o-2024-11-20 ($2.50) and gpt-4o-2024-08-06 ($2.50).

Release history

The OpenAI model released before it was GPT-4.1 nano on 2025-04-14, which scores 38.7. The next one was o4-mini (high) on 2025-04-16, at 46.8.

Reading these numbers

The 17 scores on this page come from 2 sources: Epoch AI Benchmarking Hub (16), ARC Prize official leaderboard (1). 1 of them come from an official leaderboard and 16 from datasets or third-party runs. The leaderboard also lists 4 other reasoning-effort settings of this model, which redirect here. This page shows the highest.

o3 (high) is also listed at 4 other reasoning-effort settings; this page shows the highest-effort row, which is the one the leaderboard keeps by default.

Models ranked nearby

• #230 Claude Opus 4.5 (128k thinking) (49.9)
• #231 Mistral Large 2 (2407) (49.9)
• #232 gpt-5.6-sol-xhigh (codex-harness) (49.8)
• #234 ERNIE-4.5-300B-A47B-Base-PT (49.8)
• #235 claude-haiku-4-5-20251001_1K (49.8)
• #236 Qwen2.5-32B (49.7)

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