GPT-5.3 Codex vs GPT-5.4
Wins 3 of 7 areas
Facts · Long documents · Following instructions
Wins 2 of 7 areas
Coding · Reasoning
GPT-5.3 Codex wins more areas, narrowly.GPT-5.4 is better at coding.
Scores updated · 22 tests both models report · How we compare
Where each one wins
Tests won in each of the seven areas where both have results. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- FactsGetting facts right instead of making them up20GPT-5.3 Codex2 of 2 tests
- Long documentsFinding answers in very long texts10GPT-5.3 Codex1 of 1 test
- Following instructionsDoing exactly what it is asked10GPT-5.3 Codex1 of 1 test
- CodingWriting and fixing software03GPT-5.43 of 4 tests · 1 tie
- ReasoningHard problems that need careful thinking02GPT-5.42 of 3 tests · 1 tie
- AgentsCarrying out multi-step tasks on its own11Even1 each
- Images and chartsUnderstanding pictures, charts and video00Even0 each · 1 tie
Images and charts, long documents and following instructions 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.
GPT-5.3 Codex costs 10% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where GPT-5.3 Codex pulls ahead
- Real work tasks from 44 professionsGDPVal+11.6points ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+2.5points ahead
- Answers hard knowledge questions correctlyAA-Omniscience · Accuracy+2points ahead
Where GPT-5.4 pulls ahead
- Completes tasks by operating a computer desktopOSWorld-Verified+10.3points ahead
- Research-level physics problemsCritPt+6.5points ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+4.6points ahead
Every test, side by side
All 22 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingGPT-5.4
- LMArena · WebDevGPT-5.4 by 56 rating points14091465+56 rating
- Terminal-Bench HardGPT-5.4 by 4.65357.6+4.6
- SciCodeGPT-5.4 by 3.453.256.6+3.4
- SWE-bench Protie56.857.7tie
AgentsEven
- GDPValGPT-5.3 Codex by 11.64937.4+11.6
- OSWorld-VerifiedGPT-5.4 by 10.364.775+10.3
ReasoningGPT-5.4
- CritPtGPT-5.4 by 6.516.923.4+6.5
- Humanity's Last ExamGPT-5.4 by 1.242.543.7+1.2
- GPQA Diamondtie91.592tie
FactsGPT-5.3 Codex
- AA-Omniscience · Non-hallucinationGPT-5.3 Codex by 2.510.88.3+2.5
- AA-Omniscience · AccuracyGPT-5.3 Codex by 252.950.9+2
Following instructionsGPT-5.3 Codex
- IFBenchGPT-5.3 Codex by 1.575.473.9+1.5
Other results8 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.
- Artificial Analysis Coding IndexGPT-5.4 by 1853.171.1+18
- AA Agentic IndexGPT-5.3 Codex by 16.360.544.2+16.3
- LiveBenchGPT-5.4 by 7.572.880.3+7.5
- AA IntelligenceGPT-5.4 by 6.532.539+6.5
- HiL-Bench (Tools-allowed)GPT-5.4 by 5.44.39.7+5.4
- AA-OmniscienceGPT-5.3 Codex by 5.110.95.8+5.1
- Terminal-Bench 2.0GPT-5.3 Codex by 2.277.375.1+2.2
- τ²-Bench Telecom (AA run)GPT-5.4 by 1.18687.1+1.1
Questions people ask
Which is better, GPT-5.3 Codex or GPT-5.4?
GPT-5.3 Codex wins three of the seven areas where both have results: facts, long documents and following instructions. GPT-5.4 wins coding and reasoning. They are level on agents and images and charts.
Which is better for coding?
GPT-5.4. It wins 3 of the 4 coding tests both models report; GPT-5.3 Codex wins none, and 1 is a tie.
Which is cheaper?
GPT-5.3 Codex costs $1.75 per million input tokens and $14.00 per million output tokens; GPT-5.4 costs $2.50 and $15.00. That makes GPT-5.3 Codex about 10% cheaper for the same work.
How do you compare the two?
We use the 22 benchmark tests both models have published scores on. The verdict counts the 14 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 8 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.