Gemma 4 31B vs Step 5 Preview
Updated October 21, 2026. 43 benchmarks have a result for both. Coding is ranked on the Vibe Coding Index; every other board is shown as published.
Decision reading
Pick Step 5 Preview for coding: it scores 71.4 on the Vibe Coding Index against 38.3 for Gemma 4 31B.
Pick Gemma 4 31B when output tokens cost 6.8x less, or when what matters is LMArena Text: hard prompts in English or LMArena Vision: homework, where it leads.
- Shared results
- 43
- Gemma 4 31B only
- 25
- Step 5 Preview only
- 18
- Like-for-like categories
- 8 of 9
Shared results can support a head-to-head reading. Results only one model has describe coverage, not superiority.
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, four public coding benchmarks
Step 5 Preview
Step 5 Preview scores 71.4 against 38.3.
2 benchmarks behind each score
Intelligence
AA Intelligence Index
Step 5 Preview
Step 5 Preview leads, 43.7 against 14.7.
Same board, both measured
Agentic work
AA Agentic Index
No call
Only Gemma 4 31B has a AA Agentic Index result, 4.2.
Directional only
Long documents
Listed context window
Step 5 Preview
Step 5 Preview takes 1.05M tokens against 262K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
Gemma 4 31B
Gemma 4 31B $0.0004, Step 5 Preview $0.0024 per turn.
List prices
At a glance
- Coding index
- Gemma 4 31B 38.3 from 2 benchmarks, Step 5 Preview 71.4 from 2 benchmarks.
- Widest category
- Intelligence, 29.0 points apart on average over 1 shared benchmark.
- Widest single gap
- LMArena WebDev: React: Gemma 4 31B 1359 against Step 5 Preview 1583, 37.3 points apart on a 0 to 100 scale.
- Output price
- Gemma 4 31B $0.40 against Step 5 Preview $2.70 per million output tokens, 6.8x apart.
- Context window
- Gemma 4 31B takes 262K tokens, Step 5 Preview 1.05M.
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 4 31B | Bars | Step 5 Preview | Ties |
|---|---|---|---|---|
| Coding | 0 | 7 | ||
| Intelligence | 0 | 1 | ||
| Agentic work | 0 | 3 | ||
| Reasoning | 2 | 5 | ||
| Math | 2 | 0 | ||
| Knowledge | 3 | 1 | ||
| Web design | 0 | 1 | ||
| General | 9 | 9 | ||
| All shared boards | 16 | 27 |
CodingLike-for-likeFour of these boards feed the Vibe Coding Index07
| Benchmark | Gemma 4 31B | Step 5 Preview | Leader |
|---|---|---|---|
| LiveBench coding | 49.0% | 85.6% | Step 5 Preview |
| LMArena WebDev | 1367 | 1567 | Step 5 Preview |
| AA Coding Index | 43.4 | Not measured | Only Gemma 4 31B measured |
| Terminal-Bench 4 (Vals AI) | Not measured | 31.8% | Only Step 5 Preview measured |
| Vibe Code Bench (Vals AI) | Not measured | 70.4% | Only Step 5 Preview measured |
| Code migration (Vals AI) | Not measured | 39.0% | Only Step 5 Preview measured |
| IOI (Vals AI) | Not measured | 38.1% | Only Step 5 Preview measured |
| SWE-bench Multimodal (Vals AI) | Not measured | 40.8% | Only Step 5 Preview measured |
| SciCode | 45.5% | 58.9% | Step 5 Preview |
| Terminal-Bench Hard (Artificial Analysis run) | 36.4% | Not measured | Only Gemma 4 31B measured |
| WeirdML | 52.3% | Not measured | Only Gemma 4 31B measured |
| ALE-Bench | 925.5 | Not measured | Only Gemma 4 31B measured |
| LMArena Text: coding | 1459 | 1485 | Step 5 Preview |
| LMArena Text: software and IT | 1463 | 1473 | Step 5 Preview |
| LMArena WebDev: React | 1359 | 1583 | Step 5 Preview |
| LMArena WebDev: HTML | 1380 | 1556 | Step 5 Preview |
| LMArena WebDev: from an image | Not measured | 1550 | Only Step 5 Preview measured |
| Neither model has a result on the other 21 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index01
| Benchmark | Gemma 4 31B | Step 5 Preview | Leader |
|---|---|---|---|
| AA Intelligence Index | 14.7 | 43.7 | Step 5 Preview |
| Epoch Capabilities Index | 143.0 | Not measured | Only Gemma 4 31B measured |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index03
| Benchmark | Gemma 4 31B | Step 5 Preview | Leader |
|---|---|---|---|
| AA Agentic Index | 4.2 | Not measured | Only Gemma 4 31B measured |
| LiveBench research | 44.4% | 72.8% | Step 5 Preview |
| LiveBench browser | 4.0% | 24.0% | Step 5 Preview |
| LiveBench long agentic | 13.3% | 27.5% | Step 5 Preview |
| SREBench (Vals AI) | Not measured | 0.4% | Only Step 5 Preview measured |
| Finance Agent (Vals AI) | Not measured | 50.7% | Only Step 5 Preview measured |
| Tax Agent Bench (Vals AI) | Not measured | 27.3% | Only Step 5 Preview measured |
| Legal Research Bench (Vals AI) | Not measured | 40.4% | Only Step 5 Preview measured |
| Excel Modeling Benchmark (Vals AI) | Not measured | 57.1% | Only Step 5 Preview measured |
| BioMysteryBench (Vals AI) | Not measured | 67.4% | Only Step 5 Preview measured |
| Terminal-Bench-Science (Vals AI) | Not measured | 2.9% | Only Step 5 Preview measured |
| Vals Science Index | Not measured | 29.6% | Only Step 5 Preview measured |
| τ²-Bench Telecom | 65.5% | Not measured | Only Gemma 4 31B measured |
| Neither model has a result on the other 26 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index25
| Benchmark | Gemma 4 31B | Step 5 Preview | Leader |
|---|---|---|---|
| LiveBench reasoning | 68.3% | 83.8% | Step 5 Preview |
| MysteryMechanism (Vals AI) | Not measured | 18.5% | Only Step 5 Preview measured |
| GPQA Diamond (Artificial Analysis run) | 85.7% | Not measured | Only Gemma 4 31B measured |
| Humanity's Last Exam (Artificial Analysis run) | 23.6% | 46.5% | Step 5 Preview |
| GPQA Diamond (Epoch AI run) | 75.8% | Not measured | Only Gemma 4 31B measured |
| LMArena Text: hard prompts | 1448 | 1461 | Step 5 Preview |
| LMArena Text: expert | 1464 | 1485 | Step 5 Preview |
| LMArena Text: hard prompts in English | 1456 | 1454 | Gemma 4 31B |
| LMArena Vision: diagrams | 1283 | 1308 | Step 5 Preview |
| LMArena Vision: homework | 1310 | 1265 | Gemma 4 31B |
| Neither model has a result on the other 11 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index20
| Benchmark | Gemma 4 31B | Step 5 Preview | Leader |
|---|---|---|---|
| OTIS Mock AIME | 73.3% | Not measured | Only Gemma 4 31B measured |
| ProofBench | Not measured | 42.0% | Only Step 5 Preview measured |
| LMArena Text: math | 1463 | 1458 | Gemma 4 31B |
| LMArena Text: mathematical work | 1465 | 1460 | Gemma 4 31B |
| Neither model has a result on the other 9 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index31
| Benchmark | Gemma 4 31B | Step 5 Preview | Leader |
|---|---|---|---|
| CaseLaw (Vals AI) | 52.6% | Not measured | Only Gemma 4 31B measured |
| MortgageTax (Vals AI) | 61.4% | Not measured | Only Gemma 4 31B measured |
| SimpleQA Verified | 10.4% | Not measured | Only Gemma 4 31B measured |
| LMArena Text: science | 1457 | 1454 | Gemma 4 31B |
| LMArena Text: medicine and healthcare | 1451 | 1439 | Gemma 4 31B |
| LMArena Text: legal and government | 1438 | 1437 | Gemma 4 31B |
| LMArena Text: business and finance | 1430 | 1435 | Step 5 Preview |
| Neither model has a result on the other 10 boards here | |||
Web designLike-for-likePublic boards, shown as published, kept out of the index01
| Benchmark | Gemma 4 31B | Step 5 Preview | Leader |
|---|---|---|---|
| LiveBench design | 58.4% | 76.9% | Step 5 Preview |
| Neither model has a result on the other 8 boards here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index99
| Benchmark | Gemma 4 31B | Step 5 Preview | Leader |
|---|---|---|---|
| LiveBench data analysis | 53.3% | 63.5% | Step 5 Preview |
| Public Benefits Bench (Vals AI) | Not measured | 58.1% | Only Step 5 Preview measured |
| SAGE (Vals AI) | 55.0% | Not measured | Only Gemma 4 31B measured |
| IFBench | 75.6% | Not measured | Only Gemma 4 31B measured |
| Long context reasoning (AA-LCR) | 69.7% | 88.3% | Step 5 Preview |
| DTBench | 82.7% | Not measured | Only Gemma 4 31B measured |
| LMCA | 39.3% | Not measured | Only Gemma 4 31B measured |
| Surface Evolver Bench | 30.6% | Not measured | Only Gemma 4 31B measured |
| LMArena Text | 1443 | 1440 | Gemma 4 31B |
| LMArena Text: instruction following | 1433 | 1438 | Step 5 Preview |
| LMArena Text: creative writing | 1417 | 1409 | Gemma 4 31B |
| LMArena Vision | 1276 | 1265 | Gemma 4 31B |
| LMArena Text: English | 1451 | 1447 | Gemma 4 31B |
| LMArena Text: non-English | 1430 | 1428 | Gemma 4 31B |
| LMArena Text: Chinese | 1475 | 1500 | Step 5 Preview |
| LMArena Text: Japanese | Not measured | 1383 | Only Step 5 Preview measured |
| LMArena Text: French | 1436 | 1438 | Step 5 Preview |
| LMArena Text: Spanish | 1442 | 1441 | Gemma 4 31B |
| LMArena Text: Russian | 1460 | 1433 | Gemma 4 31B |
| LMArena Text: Polish | 1433 | Not measured | Only Gemma 4 31B measured |
| LMArena Text: multi-turn | 1452 | 1437 | Gemma 4 31B |
| LMArena Text: longer queries | 1446 | 1449 | Step 5 Preview |
| LMArena Text: writing and language | 1432 | 1425 | Gemma 4 31B |
| LMArena Text: entertainment and media | 1405 | 1412 | Step 5 Preview |
| LMArena Vision: English | 1271 | 1290 | Step 5 Preview |
| LMArena Vision: Chinese | 1326 | Not measured | Only Gemma 4 31B measured |
| LMArena Vision: reading text | 1282 | 1282 | Step 5 Preview |
| LMArena Vision: humor | 1280 | Not measured | Only Gemma 4 31B measured |
| LMArena Vision: captioning | 1218 | Not measured | Only Gemma 4 31B measured |
| LMArena Vision: recognizing things | 1226 | Not measured | Only Gemma 4 31B measured |
| LMArena Vision: creative writing | 1285 | Not measured | Only Gemma 4 31B measured |
| LMArena Document | 1425 | Not measured | Only Gemma 4 31B measured |
| Neither model has a result on the other 14 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | Gemma 4 31B | Step 5 Preview | 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 4 31B costs less
Repository review
50K input + 3K output tokens
Gemma 4 31B costs less
Agent loop
200K input + 10K output tokens
Gemma 4 31B 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 4 31B | StepFun Step 5 Preview |
|---|---|---|
| Maker | StepFun | |
| Weights | Open weights | Open weights |
| Released | April 2, 2026 | October 8, 2026 |
| Context window | 262K | 1.05M |
| Input $/Mtok | $0.20 | $1.00 |
| Output $/Mtok | $0.40 | $2.70 |
| Coding index | 38.3 | 71.4 |
| Independent measurements behind it | 2 | 2 |
| Cost per solved task | Not computable | Not computable |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is Gemma 4 31B better than Step 5 Preview for coding?
Pick Step 5 Preview for coding: it scores 71.4 on the Vibe Coding Index against 38.3 for Gemma 4 31B. Pick Gemma 4 31B when output tokens cost 6.8x less, or when what matters is LMArena Text: hard prompts in English or LMArena Vision: homework, where it leads.
Which is cheaper, Gemma 4 31B or Step 5 Preview?
Gemma 4 31B costs $0.20 per million input tokens and $0.40 per million output tokens; Step 5 Preview costs $1.00 and $2.70. Cost per solved task is known for neither of them.
Which has the bigger context window, Gemma 4 31B or Step 5 Preview?
Step 5 Preview, with 1.05M tokens against 262K.
How many benchmarks is the Gemma 4 31B vs Step 5 Preview comparison based on?
43 benchmarks have a result for both models as of October 21, 2026. Gemma 4 31B's coding index rests on 2 independent benchmarks and Step 5 Preview's on 2. Where only one model has a result, the page says so instead of naming a winner.