GPT-5.2 vs GPT-5.2 Codex
Wins 3 of 8 areas
Agents · Reasoning · Math
Wins 2 of 8 areas
Images and charts · Following instructions
GPT-5.2 wins more areas, narrowly.GPT-5.2 Codex is better at following instructions.
Scores updated · 28 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.
- ReasoningHard problems that need careful thinking30GPT-5.23 of 4 tests · 1 tie
- AgentsCarrying out multi-step tasks on its own10GPT-5.21 of 1 test
- MathCompetition and research-level math10GPT-5.21 of 1 test
- Following instructionsDoing exactly what it is asked02GPT-5.2 Codex2 of 2 tests
- Images and chartsUnderstanding pictures, charts and video01GPT-5.2 Codex1 of 1 test
- CodingWriting and fixing software33Even3 each · 1 tie
- FactsGetting facts right instead of making them up11Even1 each
- Long documentsFinding answers in very long texts00Even0 each · 1 tie
Agents, images and charts, 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.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where GPT-5.2 pulls ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+9.9points ahead
- Real work tasks from 44 professionsGDPVal+8.9points ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+7.2points ahead
Where GPT-5.2 Codex pulls ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+26.5points ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+7.8points ahead
- Programming problems, refreshed regularlyLiveBench · Coding+7.5points ahead
Every test, side by side
All 28 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingEven
- SWE-bench ProGPT-5.2 Codex by 26.529.956.4+26.5
- Terminal-Bench HardGPT-5.2 by 9.94737.1+9.9
- LMArena · WebDevGPT-5.2 by 76 rating points14151339+76 rating
- LiveBench · CodingGPT-5.2 Codex by 7.576.183.6+7.5
- SWE-bench VerifiedGPT-5.2 by 7.28072.8+7.2
- SciCodeGPT-5.2 Codex by 2.552.154.6+2.5
- LiveBench · Agentic Codingtie50.349.4tie
ReasoningGPT-5.2
- LiveBench · ReasoningGPT-5.2 by 5.583.277.7+5.5
- CritPtGPT-5.2 by 2.911.68.7+2.9
- Humanity's Last ExamGPT-5.2 by 237.735.7+2
- GPQA Diamondtie90.389.9tie
FactsEven
- AA-Omniscience · Non-hallucinationGPT-5.2 Codex by 7.818.826.6+7.8
- AA-Omniscience · AccuracyGPT-5.2 by 3.244.341.1+3.2
Images and chartsGPT-5.2 Codex
- MMMU-ProGPT-5.2 Codex by 1.774.676.3+1.7
Following instructionsGPT-5.2 Codex
- LiveBench · Instruction FollowingGPT-5.2 Codex by 4.761.866.5+4.7
- IFBenchGPT-5.2 Codex by 2.275.477.6+2.2
Other results9 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.
- Terminal-Bench 2.0GPT-5.2 Codex by 105464+10
- τ²-Bench Telecom (AA run)GPT-5.2 Codex by 7.384.892.1+7.3
- livebench_languageGPT-5.2 by 6.179.873.7+6.1
- Artificial Analysis Coding IndexGPT-5.2 by 5.748.743+5.7
- AA Agentic IndexGPT-5.2 by 3.760.256.5+3.7
- AA IntelligenceGPT-5.2 by 1.930.428.5+1.9
- AA-OmniscienceGPT-5.2 by 1.3-0.9-2.2+1.3
- LiveBenchtie74.874.3tie
- livebench_data_analysistie78.278.2tie
Questions people ask
Which is better, GPT-5.2 or GPT-5.2 Codex?
GPT-5.2 wins three of the eight areas we test: agents, reasoning and math. GPT-5.2 Codex wins images and charts and following instructions. They are level on coding, facts and long documents.
Which is better for coding?
Neither. They win 3 coding tests each of the 7 both models report, and 1 is a tie.
Which is cheaper?
GPT-5.2 costs $1.75 per million input tokens and $14.00 per million output tokens; GPT-5.2 Codex costs $1.75 and $14.00. For a million tokens read plus a million written, they cost about the same.
How do you compare the two?
We use the 28 benchmark tests both models have published scores on. The verdict counts the 19 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 9 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.