o3 vs o4-mini
Wins 6 of 8 areas
Coding · Reasoning · Facts · Images and charts · Long documents · Following instructions
Wins 2 of 8 areas
Agents · Math
o3 is the stronger all-rounder.o4-mini is cheaper and better at agents.
Scores updated · 58 tests both models report · How we compare
Where each one wins
Tests won in each of the eight areas we test. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- Images and chartsUnderstanding pictures, charts and video40o34 of 5 tests · 1 tie
- ReasoningHard problems that need careful thinking20o32 of 4 tests · 2 ties
- Following instructionsDoing exactly what it is asked20o32 of 2 tests
- CodingWriting and fixing software21o32 of 3 tests
- FactsGetting facts right instead of making them up21o32 of 3 tests
- Long documentsFinding answers in very long texts10o31 of 1 test
- AgentsCarrying out multi-step tasks on its own02o4-mini2 of 2 tests
- MathCompetition and research-level math01o4-mini1 of 1 test
Math and long documents rest on a single test each.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
o4-mini costs 45% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where o3 pulls ahead
- Short factual questions, answered correctlySimpleQA Verified+30.6points ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+21.9points ahead
- Keeps track of context across a multi-turn chatMulti-Challenge+17.4points ahead
Where o4-mini pulls ahead
- Real work tasks from 44 professionsGDPVal+12.6points ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+7.6points ahead
- Code for real scientific research problemsSciCode+5.5points ahead
Every test, side by side
All 58 tests both models report. The winning score is in its model's colour; marks a score checked independently.
Codingo3
- Terminal-Bench Hardo3 by 21.937.115.2+21.9
- SciCodeo4-mini by 5.54146.5+5.5
- SWE-bench Verifiedo3 by 169.168.1+1
Agentso4-mini
- GDPValo4-mini by 12.612.825.4+12.6
- BrowseCompo4-mini by 1.849.751.5+1.8
Reasoningo3
- GPQA Diamondo3 by 4.382.778.4+4.3
- Humanity's Last Examo3 by 3.620.116.5+3.6
- CritPttie1.10.6tie
- ARC-AGI-2tie6.56.1tie
Factso3
- SimpleQA Verifiedo3 by 30.649.418.8+30.6
- AA-Omniscience · Accuracyo3 by 13.838.624.8+13.8
- AA-Omniscience · Non-hallucinationo4-mini by 7.611.919.5+7.6
Images and chartso3
- CharXiv (RQ)o3 by 6.678.672+6.6
- MathVistao3 by 2.586.884.3+2.5
- LMArena · Visiono3 by 20 rating points12141194+20 rating
- MMMUo3 by 1.382.981.6+1.3
- MMMU-Protie70.169.2tie
Following instructionso3
- Multi-Challengeo3 by 17.460.443+17.4
- IFBencho3 by 2.771.468.7+2.7
Other results37 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- τ²-Bench Telecom (AA run)o3 by 25.180.755.6+25.1
- AA-Omniscienceo3 by 20.2-15.6-35.7+20.2
- livecodebench_hardo3 by 17.16648.9+17.1
- Artificial Analysis Coding Indexo3 by 12.838.425.6+12.8
- Aider-Polygloto3 by 12.481.368.9+12.4
- LiveCodeBencho3 by 7.384.777.4+7.3
- AIME 2025o4-mini by 6.386.492.7+6.3
- AIR-Bench 2024o3 by 684.578.5+6
- MMLU Language (0-shot) - Arabico3 by 4.390.486.1+4.3
- frontiermath_tier_4_v1o4-mini by 4.22.16.3+4.2
- bbqo3 by 3.997.994+3.9
- MMLU Language (0-shot) - Hindio3 by 3.989.885.9+3.9
- MMLU Language (0-shot) - Germano3 by 3.890.586.7+3.8
- livecodebench_mediumo3 by 3.689.886.2+3.6
- MMLU Language (0-shot) - Averageo3 by 3.688.885.2+3.6
- AA Intelligenceo3 by 3.520.216.7+3.5
- MMLU Language (0-shot) - Italiano3 by 3.591.287.7+3.5
- MMLU Language (0-shot) - Frencho3 by 3.290.687.4+3.2
- MMLU Language (0-shot) - Indonesiano3 by 2.989.886.9+2.9
- TAU-bench (airline)o3 by 2.85249.2+2.8
- MMLU Language (0-shot) - Koreano3 by 2.689.386.7+2.6
- imo_2025o3 by 2.416.714.3+2.4
- MMLU Language (0-shot) - Chinese (Simplified)o3 by 2.489.386.9+2.4
- ARC-AGI-1o3 by 2.160.858.7+2.1
- TAU-bench (retail)o3 by 2.173.971.8+2.1
- HarmBencho3 by 1.498.497+1.4
- simple_safety_testso4-mini by 199100+1
- StrongREJECTo3 by 19796+1
- livecodebench_easytie99.198.8tie
- SimpleQA hallucination ratelower is bettertie0.50.8tie
- PersonQA hallucination ratelower is bettertie0.30.5tie
- anthropic_red_teamtie98.398.2tie
- BBQ Accuracy on Ambiguous Questionstie0.90.8tie
- BBQ Accuracy on Unambiguous Questionstie0.90.9tie
- XSTesttie97.397.4tie
- AA Agentic Indextie36.136.1tie
- BBQ P(not stereotyping | ambiguous question, not unknown)tie0.30.3tie
Questions people ask
Which is better, o3 or o4-mini?
o3 wins six of the eight areas we test: coding, reasoning, facts, images and charts, long documents and following instructions. o4-mini wins agents and math, and costs 45% less.
Which is better for coding?
o3. It wins 2 of the 3 coding tests both models report; o4-mini wins 1.
Which is cheaper?
o3 costs $2.00 per million input tokens and $8.00 per million output tokens; o4-mini costs $1.10 and $4.40. That makes o4-mini about 45% cheaper for the same work.
How do you compare the two?
We use the 58 benchmark tests both models have published scores on. The verdict counts the 21 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 37 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.