Gemma 3 4B vs GPT-5.6 Luna
GPT-5.6 Luna leads Gemma 3 4B on the Artificial Analysis Intelligence Index, 37.3 to 4.8, and Gemma 3 4B costs 16.0x less per output token ($0.075 vs $1.20 per million). Updated October 22, 2026; 34 benchmarks have a result for both.
Gemma 3 4B or GPT-5.6 Luna: which to pick
Only GPT-5.6 Luna has a coding score, 58.0 on the Vibe Coding Index; Gemma 3 4B has none, so the two cannot be ranked on coding.
Pick Gemma 3 4B for output tokens 16.0x cheaper ($0.075 vs $1.20 per million).
- Shared results
- 34
- Gemma 3 4B only
- 8
- GPT-5.6 Luna only
- 85
- Like-for-like categories
- 6 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 GPT-5.6 Luna has a coding score, so no call is made.
One side not measured
Intelligence
AA Intelligence Index
GPT-5.6 Luna
GPT-5.6 Luna leads, 37.3 against 4.8.
Same board, both measured
Agentic work
AA Agentic Index
No call
Only GPT-5.6 Luna has a AA Agentic Index result, 42.1.
Directional only
Reasoning
GPQA Diamond
No call
Only GPT-5.6 Luna has a GPQA Diamond result, 91.7%.
Directional only
Knowledge
MMLU Pro
No call
Only GPT-5.6 Luna has a MMLU Pro result, 86.0%.
Directional only
Web design
Design Arena: websites
No call
Only GPT-5.6 Luna has a Design Arena: websites result, 1255.
Directional only
Long documents
Listed context window
GPT-5.6 Luna
GPT-5.6 Luna takes 1.05M tokens against 131K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
Gemma 3 4B
Gemma 3 4B $0.0001, GPT-5.6 Luna $0.0008 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
No call
Only GPT-5.6 Luna has the number, $0.04.
Directional only
At a glance
- Coding index
- Gemma 3 4B not benchmarked, GPT-5.6 Luna 58.0 from 6 benchmarks.
- Widest category
- Math, 54.4 points apart on average over 3 shared benchmarks.
- Widest single gap
- OTIS Mock AIME: Gemma 3 4B 7.5% against GPT-5.6 Luna 98.3%, 90.8 points apart on a 0 to 100 scale.
- Output price
- Gemma 3 4B $0.075 against GPT-5.6 Luna $1.20 per million output tokens, 16.0x apart.
- Context window
- Gemma 3 4B takes 131K tokens, GPT-5.6 Luna 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 3 4B | Bars | GPT-5.6 Luna | Ties |
|---|---|---|---|---|
| Coding | 0 | 4 | ||
| Intelligence | 0 | 2 | ||
| Reasoning | 0 | 7 | ||
| Math | 0 | 3 | ||
| Knowledge | 0 | 4 | ||
| General | 0 | 14 | ||
| All shared boards | 0 | 34 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index04
| Benchmark | Gemma 3 4B | GPT-5.6 Luna | Leader |
|---|---|---|---|
| LiveBench coding | Not measured | 73.6% | Only GPT-5.6 Luna measured |
| LMArena WebDev | Not measured | 1518 | Only GPT-5.6 Luna measured |
| AA Coding Index | 2.7 | 71.4 | GPT-5.6 Luna |
| SWE-bench Verified (Vals AI) | Not measured | 93.0% | Only GPT-5.6 Luna measured |
| Terminal-Bench 4 (Vals AI) | Not measured | 11.6% | Only GPT-5.6 Luna measured |
| Vibe Code Bench (Vals AI) | Not measured | 77.1% | Only GPT-5.6 Luna measured |
| Code migration (Vals AI) | Not measured | 44.5% | Only GPT-5.6 Luna measured |
| ProgramBench (Vals AI) | Not measured | 0.0% | Only GPT-5.6 Luna measured |
| IOI (Vals AI) | Not measured | 61.8% | Only GPT-5.6 Luna measured |
| Terminal-Bench 2.1 (Vals AI) | Not measured | 79.0% | Only GPT-5.6 Luna measured |
| Vibe Code Bench 1-100 (Vals AI) | Not measured | 22.6% | Only GPT-5.6 Luna measured |
| SkillsBench (Vals AI) | Not measured | 60.5% | Only GPT-5.6 Luna measured |
| CyberBench (Vals AI) | Not measured | 73.6% | Only GPT-5.6 Luna measured |
| LiveCodeBench (Artificial Analysis run) | 11.2% | Not measured | Only Gemma 3 4B measured |
| SciCode | Not measured | 53.6% | Only GPT-5.6 Luna measured |
| Terminal-Bench Hard (Artificial Analysis run) | 0.8% | Not measured | Only Gemma 3 4B measured |
| FrontierCode | Not measured | 39.8% | Only GPT-5.6 Luna measured |
| WeirdML | Not measured | 60.9% | Only GPT-5.6 Luna measured |
| ALE-Bench | Not measured | 1667.4 | Only GPT-5.6 Luna measured |
| LMArena Text: coding | 1230 | 1466 | GPT-5.6 Luna |
| LMArena Text: software and IT | 1269 | 1463 | GPT-5.6 Luna |
| LMArena WebDev: React | Not measured | 1508 | Only GPT-5.6 Luna measured |
| LMArena WebDev: HTML | Not measured | 1520 | Only GPT-5.6 Luna measured |
| LMArena WebDev: from an image | Not measured | 1517 | Only GPT-5.6 Luna measured |
| SciCode | 0.0% | 6.2% | GPT-5.6 Luna |
| Neither model has a result on the other 14 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | Gemma 3 4B | GPT-5.6 Luna | Leader |
|---|---|---|---|
| AA Intelligence Index | 4.8 | 37.3 | GPT-5.6 Luna |
| Epoch Capabilities Index | 115.9 | 156.5 | GPT-5.6 Luna |
Agentic workDirectional onlyPublic boards, shown as published, kept out of the index17 boards, none shared
| Benchmark | Gemma 3 4B | GPT-5.6 Luna | Leader |
|---|---|---|---|
| AA Agentic Index | Not measured | 42.1 | Only GPT-5.6 Luna measured |
| LiveBench research | Not measured | 51.0% | Only GPT-5.6 Luna measured |
| LiveBench browser | Not measured | 68.0% | Only GPT-5.6 Luna measured |
| LiveBench long agentic | Not measured | 30.4% | Only GPT-5.6 Luna measured |
| SREBench (Vals AI) | Not measured | 1.5% | Only GPT-5.6 Luna measured |
| Finance Agent (Vals AI) | Not measured | 55.0% | Only GPT-5.6 Luna measured |
| Tax Agent Bench (Vals AI) | Not measured | 24.8% | Only GPT-5.6 Luna measured |
| Legal Research Bench (Vals AI) | Not measured | 36.5% | Only GPT-5.6 Luna measured |
| Excel Modeling Benchmark (Vals AI) | Not measured | 67.1% | Only GPT-5.6 Luna measured |
| BioMysteryBench (Vals AI) | Not measured | 61.5% | Only GPT-5.6 Luna measured |
| Terminal-Bench-Science (Vals AI) | Not measured | 0.0% | Only GPT-5.6 Luna measured |
| Vals Science Index | Not measured | 25.3% | Only GPT-5.6 Luna measured |
| τ²-Bench Telecom | 5.0% | Not measured | Only Gemma 3 4B measured |
| APEX-Agents | Not measured | 43.0% | Only GPT-5.6 Luna measured |
| BALROG | Not measured | 45.6% | Only GPT-5.6 Luna measured |
| Vending-Bench 2 | Not measured | $4,095 | Only GPT-5.6 Luna measured |
| Berkeley Function Calling Leaderboard (BFCL) | 19.6% | Not measured | Only Gemma 3 4B measured |
| Neither model has a result on the other 24 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index07
| Benchmark | Gemma 3 4B | GPT-5.6 Luna | Leader |
|---|---|---|---|
| GPQA Diamond | Not measured | 91.7% | Only GPT-5.6 Luna measured |
| LiveBench reasoning | Not measured | 85.6% | Only GPT-5.6 Luna measured |
| MysteryMechanism (Vals AI) | Not measured | 14.4% | Only GPT-5.6 Luna measured |
| MMMU Pro (Vals AI) | Not measured | 85.0% | Only GPT-5.6 Luna measured |
| GPQA Diamond (Artificial Analysis run) | 29.1% | 91.1% | GPT-5.6 Luna |
| Humanity's Last Exam (Artificial Analysis run) | 5.3% | 39.5% | GPT-5.6 Luna |
| GPQA Diamond (Epoch AI run) | 23.2% | 91.6% | GPT-5.6 Luna |
| ARC-AGI | Not measured | 88.0% | Only GPT-5.6 Luna measured |
| ARC-AGI-2 | Not measured | 59.5% | Only GPT-5.6 Luna measured |
| SimpleBench | Not measured | 46.8% | Only GPT-5.6 Luna measured |
| Chess puzzles | Not measured | 43.0% | Only GPT-5.6 Luna measured |
| Mystery game puzzles | Not measured | 21.0% | Only GPT-5.6 Luna measured |
| LMArena Text: hard prompts | 1253 | 1451 | GPT-5.6 Luna |
| LMArena Text: expert | 1223 | 1478 | GPT-5.6 Luna |
| LMArena Text: hard prompts in English | 1253 | 1460 | GPT-5.6 Luna |
| LMArena Vision: diagrams | Not measured | 1273 | Only GPT-5.6 Luna measured |
| LMArena Vision: homework | Not measured | 1282 | Only GPT-5.6 Luna measured |
| Kagi LLM Benchmark | 25.2% | 49.1% | GPT-5.6 Luna |
| Neither model has a result on the other 4 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index03
| Benchmark | Gemma 3 4B | GPT-5.6 Luna | Leader |
|---|---|---|---|
| AA Math Index | 12.7 | Not measured | Only Gemma 3 4B measured |
| AIME 2025 (Artificial Analysis run) | 12.7% | Not measured | Only Gemma 3 4B measured |
| OTIS Mock AIME | 7.5% | 98.3% | GPT-5.6 Luna |
| FrontierMath, tiers 1 to 3 | Not measured | 82.1% | Only GPT-5.6 Luna measured |
| FrontierMath, tier 4 | Not measured | 61.0% | Only GPT-5.6 Luna measured |
| ProofBench | Not measured | 60.0% | Only GPT-5.6 Luna measured |
| LMArena Text: math | 1239 | 1459 | GPT-5.6 Luna |
| LMArena Text: mathematical work | 1256 | 1470 | GPT-5.6 Luna |
| Neither model has a result on the other 5 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index04
| Benchmark | Gemma 3 4B | GPT-5.6 Luna | Leader |
|---|---|---|---|
| MMLU Pro | Not measured | 86.0% | Only GPT-5.6 Luna measured |
| LegalBench (Vals AI) | Not measured | 84.0% | Only GPT-5.6 Luna measured |
| CorpFin (Vals AI) | Not measured | 64.2% | Only GPT-5.6 Luna measured |
| TaxEval (Vals AI) | Not measured | 76.2% | Only GPT-5.6 Luna measured |
| MortgageTax (Vals AI) | Not measured | 67.3% | Only GPT-5.6 Luna measured |
| MedCode (Vals AI) | Not measured | 42.4% | Only GPT-5.6 Luna measured |
| MMLU Pro (Artificial Analysis run) | 41.7% | Not measured | Only Gemma 3 4B measured |
| SimpleQA Verified | Not measured | 41.0% | Only GPT-5.6 Luna measured |
| LMArena Text: science | 1306 | 1433 | GPT-5.6 Luna |
| LMArena Text: medicine and healthcare | 1342 | 1399 | GPT-5.6 Luna |
| LMArena Text: legal and government | 1327 | 1446 | GPT-5.6 Luna |
| LMArena Text: business and finance | 1311 | 1439 | GPT-5.6 Luna |
| FACTS Benchmark Suite | Not measured | 35.0% | Only GPT-5.6 Luna measured |
| Neither model has a result on the other 5 boards here | |||
Web designDirectional onlyPublic boards, shown as published, kept out of the index9 boards, none shared
| Benchmark | Gemma 3 4B | GPT-5.6 Luna | Leader |
|---|---|---|---|
| Design Arena: websites | Not measured | 1255 | Only GPT-5.6 Luna measured |
| Design Arena: UI components | Not measured | 1248 | Only GPT-5.6 Luna measured |
| Design Arena: game dev | Not measured | 1194 | Only GPT-5.6 Luna measured |
| Design Arena: data viz | Not measured | 1211 | Only GPT-5.6 Luna measured |
| Design Arena: 3D | Not measured | 1121 | Only GPT-5.6 Luna measured |
| Design Arena: all code categories | Not measured | 1231 | Only GPT-5.6 Luna measured |
| Design Arena: SVG | Not measured | 1194 | Only GPT-5.6 Luna measured |
| Design Arena: ASCII art | Not measured | 1142 | Only GPT-5.6 Luna measured |
| LiveBench design | Not measured | 76.0% | Only GPT-5.6 Luna measured |
GeneralLike-for-likePublic boards, shown as published, kept out of the index014
| Benchmark | Gemma 3 4B | GPT-5.6 Luna | Leader |
|---|---|---|---|
| LiveBench data analysis | Not measured | 78.0% | Only GPT-5.6 Luna measured |
| Public Benefits Bench (Vals AI) | Not measured | 61.2% | Only GPT-5.6 Luna measured |
| SAGE (Vals AI) | Not measured | 44.2% | Only GPT-5.6 Luna measured |
| MedScribe (Vals AI) | Not measured | 84.4% | Only GPT-5.6 Luna measured |
| Vals Multimodal Index | Not measured | 69.1% | Only GPT-5.6 Luna measured |
| IFBench | 28.3% | Not measured | Only Gemma 3 4B measured |
| Long context reasoning (AA-LCR) | 6.7% | 83.7% | GPT-5.6 Luna |
| DTBench | 50.9% | 89.1% | GPT-5.6 Luna |
| LMCA | 2.8% | 48.5% | GPT-5.6 Luna |
| Surface Evolver Bench | Not measured | 61.9% | Only GPT-5.6 Luna measured |
| Furniture assembly | Not measured | 42.5% | Only GPT-5.6 Luna measured |
| LMArena Text | 1291 | 1431 | GPT-5.6 Luna |
| LMArena Text: instruction following | 1239 | 1437 | GPT-5.6 Luna |
| LMArena Text: creative writing | 1270 | 1396 | GPT-5.6 Luna |
| LMArena Vision | Not measured | 1257 | Only GPT-5.6 Luna measured |
| LMArena Text: English | 1300 | 1441 | GPT-5.6 Luna |
| LMArena Text: non-English | 1273 | 1417 | GPT-5.6 Luna |
| LMArena Text: Chinese | Not measured | 1470 | Only GPT-5.6 Luna measured |
| LMArena Text: German | 1281 | 1455 | GPT-5.6 Luna |
| LMArena Text: Japanese | Not measured | 1415 | Only GPT-5.6 Luna measured |
| LMArena Text: Korean | Not measured | 1415 | Only GPT-5.6 Luna measured |
| LMArena Text: French | Not measured | 1458 | Only GPT-5.6 Luna measured |
| LMArena Text: Spanish | Not measured | 1446 | Only GPT-5.6 Luna measured |
| LMArena Text: Russian | 1294 | 1427 | GPT-5.6 Luna |
| LMArena Text: Polish | Not measured | 1405 | Only GPT-5.6 Luna measured |
| LMArena Text: multi-turn | 1255 | 1434 | GPT-5.6 Luna |
| LMArena Text: longer queries | 1273 | 1436 | GPT-5.6 Luna |
| LMArena Text: writing and language | 1288 | 1419 | GPT-5.6 Luna |
| LMArena Text: entertainment and media | 1233 | 1402 | GPT-5.6 Luna |
| LMArena Vision: English | Not measured | 1263 | Only GPT-5.6 Luna measured |
| LMArena Vision: Chinese | Not measured | 1294 | Only GPT-5.6 Luna measured |
| LMArena Vision: reading text | Not measured | 1264 | Only GPT-5.6 Luna measured |
| LMArena Vision: humor | Not measured | 1190 | Only GPT-5.6 Luna measured |
| LMArena Vision: creative writing | Not measured | 1257 | Only GPT-5.6 Luna measured |
| LMArena Document | Not measured | 1457 | Only GPT-5.6 Luna measured |
| Neither model has a result on the other 11 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | Gemma 3 4B | GPT-5.6 Luna | 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 4B costs less
Repository review
50K input + 3K output tokens
Gemma 3 4B costs less
Agent loop
200K input + 10K output tokens
Gemma 3 4B 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 4B | OpenAI GPT-5.6 Luna |
|---|---|---|
| Maker | OpenAI | |
| Weights | Open weights | No public weights listed |
| Released | March 13, 2025 | July 9, 2026 |
| Context window | 131K | 1.05M |
| Input $/Mtok | $0.045 | $0.20 |
| Output $/Mtok | $0.075 | $1.20 |
| Coding index | Not benchmarked | 58.0 |
| Independent measurements behind it | 0 | 6 |
| Cost per solved task | Not computable | $0.04 |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is Gemma 3 4B better than GPT-5.6 Luna for coding?
Only GPT-5.6 Luna has a coding score, 58.0 on the Vibe Coding Index; Gemma 3 4B has none, so the two cannot be ranked on coding. Pick Gemma 3 4B for output tokens 16.0x cheaper ($0.075 vs $1.20 per million).
Which is cheaper, Gemma 3 4B or GPT-5.6 Luna?
Gemma 3 4B costs $0.045 per million input tokens and $0.075 per million output tokens; GPT-5.6 Luna costs $0.20 and $1.20. Cost per solved task is known for GPT-5.6 Luna only.
Which has the bigger context window, Gemma 3 4B or GPT-5.6 Luna?
GPT-5.6 Luna, with 1.05M tokens against 131K.
How many benchmarks is the Gemma 3 4B vs GPT-5.6 Luna comparison based on?
34 benchmarks have a result for both models as of October 22, 2026. Gemma 3 4B's coding index rests on 0 independent benchmarks and GPT-5.6 Luna's on 6. Where only one model has a result, the page says so instead of naming a winner.