Gemma 4 31B vs GPT-5.2
Wins 0 of 8 areas
—
Wins 7 of 8 areas
Coding · Agents · Reasoning · Facts · Images and charts · Math · Long documents
GPT-5.2 is the stronger all-rounder.Gemma 4 31B is cheaper.
Scores updated · 45 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.
- CodingWriting and fixing software15GPT-5.25 of 6 tests
- ReasoningHard problems that need careful thinking03GPT-5.23 of 3 tests
- MathCompetition and research-level math03GPT-5.23 of 3 tests
- FactsGetting facts right instead of making them up13GPT-5.23 of 4 tests
- Images and chartsUnderstanding pictures, charts and video13GPT-5.23 of 4 tests
- Long documentsFinding answers in very long texts02GPT-5.22 of 2 tests
- AgentsCarrying out multi-step tasks on its own12GPT-5.22 of 3 tests
- Following instructionsDoing exactly what it is asked00Even0 each · 1 tie
Following instructions rests on a single test.
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.
Gemma 4 31B costs 96% less for the same work.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where GPT-5.2 pulls ahead
- Real work tasks from 44 professionsGDPVal+42.2points ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+28points ahead
- Short factual questions, answered correctlySimpleQA Verified+26.7points ahead
Where Gemma 4 31B pulls ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+5.8points ahead
- Customer-service tasks in a simulated bankτ-Bench V3 · Banking+3.7points ahead
Every test, side by side
All 45 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingGPT-5.2
- SWE-bench VerifiedGPT-5.2 by 285280+28
- SWE-bench MultilingualGPT-5.2 by 1551.766.7+15
- Terminal-Bench HardGPT-5.2 by 10.636.447+10.6
- SciCodeGPT-5.2 by 6.645.552.1+6.6
- SWE-bench ProGemma 4 31B by 5.835.729.9+5.8
- LMArena · WebDevGPT-5.2 by 51 rating points13651416+51 rating
AgentsGPT-5.2
- GDPValGPT-5.2 by 42.26.148.3+42.2
- τ-Bench V3 · BankingGemma 4 31B by 3.714.811.1+3.7
- MCP AtlasGPT-5.2 by 3.457.260.6+3.4
ReasoningGPT-5.2
- Humanity's Last ExamGPT-5.2 by 14.123.637.7+14.1
- CritPtGPT-5.2 by 10.21.411.6+10.2
- GPQA DiamondGPT-5.2 by 4.685.790.3+4.6
FactsGPT-5.2
- SimpleQA VerifiedGPT-5.2 by 26.710.437.1+26.7
- AA-Omniscience · AccuracyGPT-5.2 by 24.32044.3+24.3
- AA-Omniscience · Non-hallucinationGPT-5.2 by 3.81518.8+3.8
- Vectara HHEM hallucination ratelower is betterGemma 4 31B by 3.47.410.8+3.4
Images and chartsGPT-5.2
- CharXiv (RQ)GPT-5.2 by 14.267.982.1+14.2
- MathVistaGPT-5.2 by 3.579.382.8+3.5
- LMArena · VisionGemma 4 31B by 31 rating points12771246+31 rating
- MMMU-ProGPT-5.2 by 1.273.474.6+1.2
MathGPT-5.2
- HMMT Feb. 2026GPT-5.2 by 19.877.297+19.8
- IMOAnswerBenchGPT-5.2 by 11.874.586.3+11.8
- AIME 2026GPT-5.2 by 9.189.298.3+9.1
Long documentsGPT-5.2
- MRCR v2 8 needle 128k (average)GPT-5.2 by 17.466.483.8+17.4
- AA-LCRGPT-5.2 by 1369.782.7+13
Other results19 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.
- AA Agentic IndexGPT-5.2 by 53.56.760.2+53.5
- AA-OmniscienceGPT-5.2 by 47-47.9-0.9+47
- τ²-Bench Telecom (AA run)GPT-5.2 by 24.959.984.8+24.9
- HLE (with tools)GPT-5.2 by 1926.545.5+19
- AA IntelligenceGPT-5.2 by 15.714.730.4+15.7
- HMMT Feb. 2025GPT-5.2 by 11.388.7100+11.3
- Terminal-Bench 2.0GPT-5.2 by 11.142.954+11.1
- vectara_avg_summary_lengthGPT-5.2 by 110.5 rating points75.8186.3+110.5 rating
- τ²-BenchGPT-5.2 by 8.676.985.5+8.6
- HMMT Nov. 2025GPT-5.2 by 8.387.595.8+8.3
- GPQA (unspecified)GPT-5.2 by 8.184.392.4+8.1
- OmniDocBench 1.5GPT-5.2 by 5.680.185.7+5.6
- Artificial Analysis Coding IndexGPT-5.2 by 5.343.448.7+5.3
- vectara_factual_consistencyGemma 4 31B by 3.492.689.2+3.4
- SimpleVQAGPT-5.2 by 2.952.955.8+2.9
- MATH-VisionGemma 4 31B by 2.685.683+2.6
- MMLU-ProGPT-5.2 by 1.585.286.7+1.5
- MMMLUGPT-5.2 by 1.288.489.6+1.2
- vectara_answer_ratetie100100tie
Questions people ask
Which is better, Gemma 4 31B or GPT-5.2?
GPT-5.2 wins seven of the eight areas we test: coding, agents, reasoning, facts, images and charts, math and long documents. Gemma 4 31B wins none, but costs 96% less. They are level on following instructions.
Which is better for coding?
GPT-5.2. It wins 5 of the 6 coding tests both models report; Gemma 4 31B wins 1.
Which is cheaper?
Gemma 4 31B costs $0.17 per million input tokens and $0.40 per million output tokens; GPT-5.2 costs $1.75 and $14.00. That makes Gemma 4 31B about 96% cheaper for the same work.
How do you compare the two?
We use the 45 benchmark tests both models have published scores on. The verdict counts the 26 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 19 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.