Gemma 3 12B vs Gemma 4 31B
Gemma 4 31B leads Gemma 3 12B on the Artificial Analysis Intelligence Index, 14.7 to 3.8, and Gemma 3 12B costs 2.5x less per output token ($0.1625 vs $0.40 per million). Updated October 22, 2026; 37 benchmarks have a result for both.
Gemma 3 12B or Gemma 4 31B: which to pick
Only Gemma 4 31B has a coding score, 39.0 on the Vibe Coding Index; Gemma 3 12B has none, so the two cannot be ranked on coding.
Pick Gemma 3 12B for output tokens 2.5x cheaper ($0.1625 vs $0.40 per million).
- Shared results
- 37
- Gemma 3 12B only
- 6
- Gemma 4 31B only
- 34
- Like-for-like categories
- 7 of 9
Only benchmarks that tested both models count toward a pick; a result only one model has is shown but not compared.
Which one for your work
Each card is decided by one number both models carry. Where only one has it, the card says so and makes no call.
Coding
Vibe Coding Index, six public coding benchmarks
No call
Only Gemma 4 31B has a coding score, so no call is made.
One side not measured
Intelligence
AA Intelligence Index
Gemma 4 31B
Gemma 4 31B leads, 14.7 against 3.8.
Same board, both measured
Agentic work
AA Agentic Index
Gemma 4 31B
Gemma 4 31B leads, 4.2 against 0.1.
Same board, both measured
Long documents
Listed context window
Gemma 4 31B
Gemma 4 31B takes 262K tokens against 131K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
Gemma 3 12B
Gemma 3 12B $0.0002, Gemma 4 31B $0.0003 per turn.
List prices
At a glance
- Coding index
- Gemma 3 12B not benchmarked, Gemma 4 31B 39.0 from 2 benchmarks.
- Widest category
- Math, 34.7 points apart on average over 3 shared benchmarks.
- Widest single gap
- Long context reasoning (AA-LCR): Gemma 3 12B 8.3% against Gemma 4 31B 69.7%, 61.3 points apart on a 0 to 100 scale.
- Output price
- Gemma 3 12B $0.1625 against Gemma 4 31B $0.40 per million output tokens, 2.5x apart.
- Context window
- Gemma 3 12B takes 131K tokens, Gemma 4 31B 262K.
Benchmark by benchmark
First the score: boards each model leads, per category, where both have a result. Then every board, a category to a fold; bars are each result on a 0 to 100 scale, so compare the two in a row, not one row with another.
| Category | Gemma 3 12B | Bars | Gemma 4 31B | Ties |
|---|---|---|---|---|
| Coding | 0 | 6 | ||
| Intelligence | 0 | 2 | ||
| Agentic work | 0 | 2 | ||
| Reasoning | 0 | 6 | ||
| Math | 0 | 3 | ||
| Knowledge | 0 | 4 | ||
| General | 0 | 14 | ||
| All shared boards | 0 | 37 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index06
| Benchmark | Gemma 3 12B | Gemma 4 31B | Leader |
|---|---|---|---|
| LiveBench coding | Not measured | 49.0% | Only Gemma 4 31B measured |
| LMArena WebDev | Not measured | 1367 | Only Gemma 4 31B measured |
| AA Coding Index | 5.8 | 43.4 | Gemma 4 31B |
| LiveCodeBench (Artificial Analysis run) | 13.7% | Not measured | Only Gemma 3 12B measured |
| SciCode | 16.4% | 45.5% | Gemma 4 31B |
| Terminal-Bench Hard (Artificial Analysis run) | 0.8% | 36.4% | Gemma 4 31B |
| WeirdML | Not measured | 52.3% | Only Gemma 4 31B measured |
| ALE-Bench | Not measured | 925.5 | Only Gemma 4 31B measured |
| LMArena Text: coding | 1281 | 1459 | Gemma 4 31B |
| LMArena Text: software and IT | 1313 | 1463 | Gemma 4 31B |
| LMArena WebDev: React | Not measured | 1359 | Only Gemma 4 31B measured |
| LMArena WebDev: HTML | Not measured | 1380 | Only Gemma 4 31B measured |
| SciCode | 0.0% | 8.9% | Gemma 4 31B |
| Neither model has a result on the other 26 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | Gemma 3 12B | Gemma 4 31B | Leader |
|---|---|---|---|
| AA Intelligence Index | 3.8 | 14.7 | Gemma 4 31B |
| Epoch Capabilities Index | 123.5 | 143.0 | Gemma 4 31B |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | Gemma 3 12B | Gemma 4 31B | Leader |
|---|---|---|---|
| AA Agentic Index | 0.1 | 4.2 | Gemma 4 31B |
| LiveBench research | Not measured | 44.4% | Only Gemma 4 31B measured |
| LiveBench browser | Not measured | 4.0% | Only Gemma 4 31B measured |
| LiveBench long agentic | Not measured | 13.3% | Only Gemma 4 31B measured |
| τ²-Bench Telecom | 10.8% | 65.5% | Gemma 4 31B |
| Berkeley Function Calling Leaderboard (BFCL) | 30.4% | Not measured | Only Gemma 3 12B measured |
| Neither model has a result on the other 35 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index06
| Benchmark | Gemma 3 12B | Gemma 4 31B | Leader |
|---|---|---|---|
| LiveBench reasoning | Not measured | 68.3% | Only Gemma 4 31B measured |
| GPQA Diamond (Artificial Analysis run) | 34.9% | 85.7% | Gemma 4 31B |
| Humanity's Last Exam (Artificial Analysis run) | 4.2% | 23.6% | Gemma 4 31B |
| GPQA Diamond (Epoch AI run) | 39.5% | 75.8% | Gemma 4 31B |
| LMArena Text: hard prompts | 1309 | 1448 | Gemma 4 31B |
| LMArena Text: expert | 1248 | 1464 | Gemma 4 31B |
| LMArena Text: hard prompts in English | 1318 | 1456 | Gemma 4 31B |
| LMArena Vision: diagrams | Not measured | 1283 | Only Gemma 4 31B measured |
| LMArena Vision: homework | Not measured | 1310 | Only Gemma 4 31B measured |
| Kagi LLM Benchmark | Not measured | 63.5% | Only Gemma 4 31B measured |
| Neither model has a result on the other 12 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index03
| Benchmark | Gemma 3 12B | Gemma 4 31B | Leader |
|---|---|---|---|
| AA Math Index | 18.3 | Not measured | Only Gemma 3 12B measured |
| AIME 2025 (Artificial Analysis run) | 18.3% | Not measured | Only Gemma 3 12B measured |
| OTIS Mock AIME | 16.7% | 73.3% | Gemma 4 31B |
| LMArena Text: math | 1307 | 1463 | Gemma 4 31B |
| LMArena Text: mathematical work | 1336 | 1465 | Gemma 4 31B |
| Neither model has a result on the other 8 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index04
| Benchmark | Gemma 3 12B | Gemma 4 31B | Leader |
|---|---|---|---|
| CaseLaw (Vals AI) | Not measured | 52.6% | Only Gemma 4 31B measured |
| MortgageTax (Vals AI) | Not measured | 61.4% | Only Gemma 4 31B measured |
| MMLU Pro (Artificial Analysis run) | 59.5% | Not measured | Only Gemma 3 12B measured |
| SimpleQA Verified | Not measured | 10.4% | Only Gemma 4 31B measured |
| LMArena Text: science | 1366 | 1457 | Gemma 4 31B |
| LMArena Text: medicine and healthcare | 1334 | 1451 | Gemma 4 31B |
| LMArena Text: legal and government | 1381 | 1438 | Gemma 4 31B |
| LMArena Text: business and finance | 1376 | 1430 | Gemma 4 31B |
| FACTS Benchmark Suite | Not measured | 45.3% | Only Gemma 4 31B measured |
| Neither model has a result on the other 9 boards here | |||
Web designDirectional onlyPublic boards, shown as published, kept out of the index1 board, none shared
| Benchmark | Gemma 3 12B | Gemma 4 31B | Leader |
|---|---|---|---|
| LiveBench design | Not measured | 58.4% | Only Gemma 4 31B measured |
| Neither model has a result on the other 8 boards here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index014
| Benchmark | Gemma 3 12B | Gemma 4 31B | Leader |
|---|---|---|---|
| LiveBench data analysis | Not measured | 53.3% | Only Gemma 4 31B measured |
| SAGE (Vals AI) | Not measured | 55.0% | Only Gemma 4 31B measured |
| IFBench | 36.7% | 75.6% | Gemma 4 31B |
| Long context reasoning (AA-LCR) | 8.3% | 69.7% | Gemma 4 31B |
| DTBench | 48.8% | 82.7% | Gemma 4 31B |
| LMCA | 4.5% | 39.3% | Gemma 4 31B |
| Surface Evolver Bench | Not measured | 30.6% | Only Gemma 4 31B measured |
| LMArena Text | 1334 | 1443 | Gemma 4 31B |
| LMArena Text: instruction following | 1299 | 1433 | Gemma 4 31B |
| LMArena Text: creative writing | 1331 | 1417 | Gemma 4 31B |
| LMArena Vision | Not measured | 1276 | Only Gemma 4 31B measured |
| LMArena Text: English | 1349 | 1451 | Gemma 4 31B |
| LMArena Text: non-English | 1318 | 1430 | Gemma 4 31B |
| LMArena Text: Chinese | Not measured | 1475 | Only Gemma 4 31B measured |
| LMArena Text: German | 1370 | Not measured | Only Gemma 3 12B measured |
| LMArena Text: French | Not measured | 1436 | Only Gemma 4 31B measured |
| LMArena Text: Spanish | Not measured | 1442 | Only Gemma 4 31B measured |
| LMArena Text: Russian | 1335 | 1460 | Gemma 4 31B |
| LMArena Text: Polish | Not measured | 1433 | Only Gemma 4 31B measured |
| LMArena Text: multi-turn | 1334 | 1452 | Gemma 4 31B |
| LMArena Text: longer queries | 1317 | 1446 | Gemma 4 31B |
| LMArena Text: writing and language | 1307 | 1432 | Gemma 4 31B |
| LMArena Text: entertainment and media | 1297 | 1405 | Gemma 4 31B |
| LMArena Vision: English | Not measured | 1271 | Only Gemma 4 31B measured |
| LMArena Vision: Chinese | Not measured | 1326 | Only Gemma 4 31B measured |
| LMArena Vision: reading text | Not measured | 1282 | Only Gemma 4 31B measured |
| LMArena Vision: humor | Not measured | 1280 | Only Gemma 4 31B measured |
| LMArena Vision: captioning | Not measured | 1218 | Only Gemma 4 31B measured |
| LMArena Vision: recognizing things | Not measured | 1226 | Only Gemma 4 31B measured |
| LMArena Vision: creative writing | Not measured | 1285 | Only Gemma 4 31B measured |
| LMArena Document | Not measured | 1425 | Only Gemma 4 31B measured |
| Neither model has a result on the other 15 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | Gemma 3 12B | Gemma 4 31B | Leader |
|---|---|---|---|
| Neither model has a result on 5 boards here | |||
What each workload costs
Three fixed token mixes at each model's listed price, so per-token rates become a bill.
Chat turn
1K input + 500 output tokens
Gemma 3 12B costs less
Repository review
50K input + 3K output tokens
Gemma 3 12B costs less
Agent loop
200K input + 10K output tokens
Gemma 3 12B costs less
Listed OpenRouter rates, without cache or batch discounts. A workload that does not fit a model's context window would need splitting.
Specification differences
The leading value in a row is marked; a row with a missing value names no leader. Our two rows are marked in red.
| Specification | Google Gemma 3 12B | Google Gemma 4 31B |
|---|---|---|
| Maker | ||
| Weights | Open weights | Open weights |
| Released | March 13, 2025 | April 2, 2026 |
| Context window | 131K | 262K |
| Input $/Mtok | $0.075 | $0.14 |
| Output $/Mtok | $0.1625 | $0.40 |
| Coding index | Not benchmarked | 39.0 |
| Independent measurements behind it | 0 | 2 |
| Cost per solved task | Not computable | Not computable |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is Gemma 3 12B better than Gemma 4 31B for coding?
Only Gemma 4 31B has a coding score, 39.0 on the Vibe Coding Index; Gemma 3 12B has none, so the two cannot be ranked on coding. Pick Gemma 3 12B for output tokens 2.5x cheaper ($0.1625 vs $0.40 per million).
Which is cheaper, Gemma 3 12B or Gemma 4 31B?
Gemma 3 12B costs $0.075 per million input tokens and $0.1625 per million output tokens; Gemma 4 31B costs $0.14 and $0.40. Cost per solved task is known for neither of them.
Which has the bigger context window, Gemma 3 12B or Gemma 4 31B?
Gemma 4 31B, with 262K tokens against 131K.
How many benchmarks is the Gemma 3 12B vs Gemma 4 31B comparison based on?
37 benchmarks have a result for both models as of October 22, 2026. Gemma 3 12B's coding index rests on 0 independent benchmarks and Gemma 4 31B's on 2. Where only one model has a result, the page says so instead of naming a winner.