Gemini 3.8 Flash vs Gemini 4 Carbon
Gemini 4 Carbon leads Gemini 3.8 Flash on coding, 86.6 to 60.3 on the Vibe Coding Index, and Gemini 3.8 Flash costs 2.7x less per output token ($3.75 vs $10.00 per million). Updated November 14, 2026; 71 benchmarks have a result for both.
vs boards won of 71 shared, 1 tied
Gemini 3.8 Flash or Gemini 4 Carbon: which to pick
Pick Gemini 4 Carbon for coding: it scores 86.6 on the Vibe Coding Index against 60.3 for Gemini 3.8 Flash, though on 4 benchmarks against 6.
Pick Gemini 3.8 Flash for output tokens 2.7x cheaper ($3.75 vs $10.00 per million) and work measured by Long context reasoning (AA-LCR), where it leads.
Model AGemini 3.8 FlashGoogle60.3/1006 independent benchmarks$2.74 per solved taskModel BGemini 4 CarbonGoogle86.6/1004 independent benchmarksNo cost per solved task
- Shared results
- 71
- Gemini 3.8 Flash only
- 78
- Gemini 4 Carbon only
- 2
- Like-for-like categories
- 8 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
Gemini 4 Carbon
Gemini 4 Carbon scores 86.6 against 60.3.
6 and 4 benchmarks behind the scores
Intelligence
AA Intelligence Index
Gemini 4 Carbon
Gemini 4 Carbon leads, 60.0 against 30.8.
Same board, both measured
Agentic work
AA Agentic Index
No call
Only Gemini 3.8 Flash has a AA Agentic Index result, 40.2.
Directional only
Reasoning
GPQA Diamond
No call
Only Gemini 3.8 Flash has a GPQA Diamond result, 94.4%.
Directional only
Knowledge
MMLU Pro
No call
Only Gemini 3.8 Flash has a MMLU Pro result, 90.2%.
Directional only
Web design
Design Arena: websites
Gemini 4 Carbon
Gemini 4 Carbon leads, 1319 against 1291.
Same board, both measured
General
Vals Index
Gemini 4 Carbon
Gemini 4 Carbon leads, 70.5% against 56.0%.
Same board, both measured
Long documents
Listed context window
No call
Both take 1.05M tokens.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
Gemini 3.8 Flash
Gemini 3.8 Flash $0.0026, Gemini 4 Carbon $0.0070 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
No call
Only Gemini 3.8 Flash has the number, $2.74.
Directional only
At a glance
- Coding index
- Gemini 3.8 Flash 60.3 from 6 benchmarks, Gemini 4 Carbon 86.6 from 4 benchmarks.
- Widest category
- Math, 30.8 points apart on average over 5 shared benchmarks.
- Widest single gap
- FrontierMath, tier 4: Gemini 3.8 Flash 22.0% against Gemini 4 Carbon 95.1%, 73.2 points apart on a 0 to 100 scale.
- Output price
- Gemini 3.8 Flash $3.75 against Gemini 4 Carbon $10.00 per million output tokens, 2.7x apart.
- Context window
- Both take 1.05M tokens.
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 | Gemini 3.8 Flash | Bars | Gemini 4 Carbon | Ties |
|---|---|---|---|---|
| Coding | 0 | 12 | ||
| Intelligence | 0 | 2 | ||
| Agentic work | 0 | 10 | ||
| Reasoning | 0 | 11 | 1 tie | |
| Math | 0 | 5 | ||
| Knowledge | 0 | 8 | ||
| Web design | 0 | 6 | ||
| General | 1 | 15 | ||
| All shared boards | 1 | 69 | 1 tie |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index012
| Benchmark | Gemini 3.8 Flash | Gemini 4 Carbon | Leader |
|---|---|---|---|
| LiveBench coding | 73.3% | 91.0% | Gemini 4 Carbon |
| LMArena WebDev | 1583 | 1768 | Gemini 4 Carbon |
| AA Coding Index | 76.3 | Not measured | Only Gemini 3.8 Flash measured |
| LiveCodeBench (Vals AI) | 89.5% | Not measured | Only Gemini 3.8 Flash measured |
| SWE-bench Verified (Vals AI) | 80.0% | Not measured | Only Gemini 3.8 Flash measured |
| Terminal-Bench 4 (Vals AI) | 19.7% | 63.1% | Gemini 4 Carbon |
| Vibe Code Bench (Vals AI) | 78.7% | 89.1% | Gemini 4 Carbon |
| Code migration (Vals AI) | 36.5% | Not measured | Only Gemini 3.8 Flash measured |
| ProgramBench (Vals AI) | 1.0% | Not measured | Only Gemini 3.8 Flash measured |
| IOI (Vals AI) | 56.9% | Not measured | Only Gemini 3.8 Flash measured |
| Terminal-Bench 2.1 (Vals AI) | 81.3% | Not measured | Only Gemini 3.8 Flash measured |
| SWE-bench Multimodal (Vals AI) | 40.8% | Not measured | Only Gemini 3.8 Flash measured |
| Vibe Code Bench 1-100 (Vals AI) | 18.8% | Not measured | Only Gemini 3.8 Flash measured |
| SkillsBench (Vals AI) | 58.0% | Not measured | Only Gemini 3.8 Flash measured |
| CyberBench (Vals AI) | 43.8% | Not measured | Only Gemini 3.8 Flash measured |
| SciCode | 56.6% | Not measured | Only Gemini 3.8 Flash measured |
| Terminal-Bench 4.0 (Artificial Analysis run) | 19.7% | 65.2% | Gemini 4 Carbon |
| FrontierCode 2.0 | 18.7% | 40.6% | Gemini 4 Carbon |
| FrontierCode | 41.2% | Not measured | Only Gemini 3.8 Flash measured |
| FrontierSWE | 19.6% | 61.2% | Gemini 4 Carbon |
| WeirdML | 84.8% | Not measured | Only Gemini 3.8 Flash measured |
| MirrorCode | 38.9% | Not measured | Only Gemini 3.8 Flash measured |
| ALE-Bench | 1269.7 | 2374.7 | Gemini 4 Carbon |
| LMArena Text: coding | 1510 | 1564 | Gemini 4 Carbon |
| LMArena Text: software and IT | 1509 | 1552 | Gemini 4 Carbon |
| LMArena WebDev: React | 1575 | 1792 | Gemini 4 Carbon |
| LMArena WebDev: HTML | 1608 | 1770 | Gemini 4 Carbon |
| LMArena WebDev: from an image | 1594 | Not measured | Only Gemini 3.8 Flash measured |
| SciCode | 12.3% | Not measured | Only Gemini 3.8 Flash measured |
| SWE-Bench Pro V2 | 94.9% | Not measured | Only Gemini 3.8 Flash measured |
| SWE Atlas: Codebase QnA | 47.0% | Not measured | Only Gemini 3.8 Flash measured |
| SWE Atlas: Refactoring | 44.8% | Not measured | Only Gemini 3.8 Flash measured |
| SWE Atlas: Test Writing | 53.7% | Not measured | Only Gemini 3.8 Flash measured |
| GameBench 2 | 64.0 | Not measured | Only Gemini 3.8 Flash measured |
| Terminal-Bench 4.0 | 19.1% | Not measured | Only Gemini 3.8 Flash measured |
| CursorBench 4.0 | Not measured | 60.4% | Only Gemini 4 Carbon measured |
| Neither model has a result on the other 16 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | Gemini 3.8 Flash | Gemini 4 Carbon | Leader |
|---|---|---|---|
| AA Intelligence Index | 30.8 | 60.0 | Gemini 4 Carbon |
| Epoch Capabilities Index | 158.8 | 169.8 | Gemini 4 Carbon |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index010
| Benchmark | Gemini 3.8 Flash | Gemini 4 Carbon | Leader |
|---|---|---|---|
| AA Agentic Index | 40.2 | Not measured | Only Gemini 3.8 Flash measured |
| Design Arena: agents building web apps | 1213 | Not measured | Only Gemini 3.8 Flash measured |
| Design Arena: agents building full-stack apps | 1217 | Not measured | Only Gemini 3.8 Flash measured |
| Design Arena: agents building mobile apps | 1225 | 1284 | Gemini 4 Carbon |
| Design Arena: agents building Android apps | 1085 | Not measured | Only Gemini 3.8 Flash measured |
| Design Arena: agents building Godot games | 1248 | Not measured | Only Gemini 3.8 Flash measured |
| Design Arena: PowerPoint slides | 1150 | Not measured | Only Gemini 3.8 Flash measured |
| Design Arena: spreadsheets | 1034 | 1224 | Gemini 4 Carbon |
| LiveBench research | 79.4% | 92.5% | Gemini 4 Carbon |
| LiveBench browser | 45.1% | 80.0% | Gemini 4 Carbon |
| LiveBench long agentic | 16.5% | 72.7% | Gemini 4 Carbon |
| SREBench (Vals AI) | 11.1% | Not measured | Only Gemini 3.8 Flash measured |
| Finance Agent (Vals AI) | 61.4% | Not measured | Only Gemini 3.8 Flash measured |
| Tax Agent Bench (Vals AI) | 32.1% | 51.8% | Gemini 4 Carbon |
| Legal Research Bench (Vals AI) | 38.9% | 55.8% | Gemini 4 Carbon |
| Harvey's Legal Agent Benchmark (Vals AI) | 12.1% | 17.5% | Gemini 4 Carbon |
| Excel Modeling Benchmark (Vals AI) | 72.2% | Not measured | Only Gemini 3.8 Flash measured |
| BioMysteryBench (Vals AI) | 62.2% | Not measured | Only Gemini 3.8 Flash measured |
| Terminal-Bench-Science (Vals AI) | 8.6% | Not measured | Only Gemini 3.8 Flash measured |
| Vals Science Index | 35.8% | Not measured | Only Gemini 3.8 Flash measured |
| Time Horizon Index (Vals AI) | 18.8% | Not measured | Only Gemini 3.8 Flash measured |
| CUA-bench (Vals AI) | 4.2% | Not measured | Only Gemini 3.8 Flash measured |
| APEX-Agents | 64.3% | 84.2% | Gemini 4 Carbon |
| Remote Labor Index | 5.8% | Not measured | Only Gemini 3.8 Flash measured |
| Vending-Bench 2 | $5,094 | $16,484 | Gemini 4 Carbon |
| Neither model has a result on the other 19 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index011
| Benchmark | Gemini 3.8 Flash | Gemini 4 Carbon | Leader |
|---|---|---|---|
| GPQA Diamond | 94.4% | Not measured | Only Gemini 3.8 Flash measured |
| LiveBench reasoning | 89.3% | 90.6% | Gemini 4 Carbon |
| MysteryMechanism (Vals AI) | 36.5% | Not measured | Only Gemini 3.8 Flash measured |
| MMMU Pro (Vals AI) | 89.1% | Not measured | Only Gemini 3.8 Flash measured |
| GPQA Diamond (Artificial Analysis run) | 95.3% | Not measured | Only Gemini 3.8 Flash measured |
| Humanity's Last Exam (Artificial Analysis run) | 47.8% | Not measured | Only Gemini 3.8 Flash measured |
| HLE-Diamond (Artificial Analysis run) | 33.1% | 69.1% | Gemini 4 Carbon |
| CritPt v2, hard subset (Artificial Analysis run) | 46.5% | 69.0% | Gemini 4 Carbon |
| GPQA Diamond (Epoch AI run) | 95.4% | 96.5% | Gemini 4 Carbon |
| ARC-AGI | 98.5% | 98.5% | Tie |
| ARC-AGI-2 | 89.2% | 94.2% | Gemini 4 Carbon |
| SimpleBench | 82.4% | 92.6% | Gemini 4 Carbon |
| Chess puzzles | 58.0% | 78.0% | Gemini 4 Carbon |
| Mystery game puzzles | 47.0% | 73.0% | Gemini 4 Carbon |
| EBR-bench | Not measured | 87.6% | Only Gemini 4 Carbon measured |
| LMArena Text: hard prompts | 1510 | 1551 | Gemini 4 Carbon |
| LMArena Text: expert | 1527 | 1582 | Gemini 4 Carbon |
| LMArena Text: hard prompts in English | 1508 | 1545 | Gemini 4 Carbon |
| LMArena Vision: diagrams | 1353 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Vision: homework | 1351 | Not measured | Only Gemini 3.8 Flash measured |
| Humanity's Last Exam (Scale run) | 34.3% | Not measured | Only Gemini 3.8 Flash measured |
| GoBench | 1615 | Not measured | Only Gemini 3.8 Flash measured |
| Neither model has a result on the other 4 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index05
| Benchmark | Gemini 3.8 Flash | Gemini 4 Carbon | Leader |
|---|---|---|---|
| OTIS Mock AIME | 98.9% | 100.0% | Gemini 4 Carbon |
| FrontierMath, tiers 1 to 3 | 68.4% | 90.5% | Gemini 4 Carbon |
| FrontierMath, tier 4 | 22.0% | 95.1% | Gemini 4 Carbon |
| ProofBench | 48.0% | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Text: math | 1525 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Text: mathematical work | 1519 | 1538 | Gemini 4 Carbon |
| MathArena | 31.0% | 85.4% | Gemini 4 Carbon |
| Neither model has a result on the other 7 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index08
| Benchmark | Gemini 3.8 Flash | Gemini 4 Carbon | Leader |
|---|---|---|---|
| MMLU Pro | 90.2% | Not measured | Only Gemini 3.8 Flash measured |
| LegalBench (Vals AI) | 87.0% | Not measured | Only Gemini 3.8 Flash measured |
| TaxEval (Vals AI) | 74.5% | Not measured | Only Gemini 3.8 Flash measured |
| MortgageTax (Vals AI) | 65.3% | Not measured | Only Gemini 3.8 Flash measured |
| MedCode (Vals AI) | 48.1% | Not measured | Only Gemini 3.8 Flash measured |
| AA-Omniscience | 23.7 | 53.5 | Gemini 4 Carbon |
| SimpleQA Verified | 69.7% | 77.7% | Gemini 4 Carbon |
| GeoBench | 86.0% | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Text: science | 1522 | 1535 | Gemini 4 Carbon |
| LMArena Text: medicine and healthcare | 1507 | 1525 | Gemini 4 Carbon |
| LMArena Text: legal and government | 1502 | 1542 | Gemini 4 Carbon |
| LMArena Text: business and finance | 1483 | 1520 | Gemini 4 Carbon |
| FACTS Benchmark Suite | 71.2% | 78.4% | Gemini 4 Carbon |
| PRBench Finance | 49.5% | Not measured | Only Gemini 3.8 Flash measured |
| PRBench Legal | 48.9% | 53.3% | Gemini 4 Carbon |
| Neither model has a result on the other 6 boards here | |||
Web designLike-for-likePublic boards, shown as published, kept out of the index06
| Benchmark | Gemini 3.8 Flash | Gemini 4 Carbon | Leader |
|---|---|---|---|
| Design Arena: websites | 1291 | 1319 | Gemini 4 Carbon |
| Design Arena: UI components | 1304 | 1357 | Gemini 4 Carbon |
| Design Arena: game dev | 1296 | 1326 | Gemini 4 Carbon |
| Design Arena: data viz | 1276 | Not measured | Only Gemini 3.8 Flash measured |
| Design Arena: 3D | 1274 | 1434 | Gemini 4 Carbon |
| Design Arena: all code categories | 1292 | 1348 | Gemini 4 Carbon |
| Design Arena: SVG | 1290 | Not measured | Only Gemini 3.8 Flash measured |
| Design Arena: ASCII art | 1269 | Not measured | Only Gemini 3.8 Flash measured |
| LiveBench design | 69.7% | 80.2% | Gemini 4 Carbon |
GeneralLike-for-likePublic boards, shown as published, kept out of the index115
| Benchmark | Gemini 3.8 Flash | Gemini 4 Carbon | Leader |
|---|---|---|---|
| Vals Index | 56.0% | 70.5% | Gemini 4 Carbon |
| LiveBench data analysis | 54.0% | 86.5% | Gemini 4 Carbon |
| Public Benefits Bench (Vals AI) | 65.3% | Not measured | Only Gemini 3.8 Flash measured |
| SAGE (Vals AI) | 88.2% | Not measured | Only Gemini 3.8 Flash measured |
| MedScribe (Vals AI) | 84.5% | Not measured | Only Gemini 3.8 Flash measured |
| Long context reasoning (AA-LCR) | 84.0% | 79.3% | Gemini 3.8 Flash |
| GDPval-AA | 1442 | 1733 | Gemini 4 Carbon |
| DTBench | 95.7% | Not measured | Only Gemini 3.8 Flash measured |
| LMCA | 52.9% | Not measured | Only Gemini 3.8 Flash measured |
| Surface Evolver Bench | 76.9% | Not measured | Only Gemini 3.8 Flash measured |
| Furniture assembly | 31.7% | 77.5% | Gemini 4 Carbon |
| LMArena Text | 1500 | 1538 | Gemini 4 Carbon |
| LMArena Text: instruction following | 1493 | 1539 | Gemini 4 Carbon |
| LMArena Text: creative writing | 1493 | 1533 | Gemini 4 Carbon |
| LMArena Vision | 1317 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Text: English | 1500 | 1530 | Gemini 4 Carbon |
| LMArena Text: non-English | 1492 | 1534 | Gemini 4 Carbon |
| LMArena Text: Chinese | 1557 | 1613 | Gemini 4 Carbon |
| LMArena Text: German | 1503 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Text: Japanese | 1515 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Text: Korean | 1470 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Text: French | 1497 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Text: Spanish | 1486 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Text: Russian | 1516 | 1560 | Gemini 4 Carbon |
| LMArena Text: Polish | 1509 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Text: multi-turn | 1503 | 1569 | Gemini 4 Carbon |
| LMArena Text: longer queries | 1508 | 1549 | Gemini 4 Carbon |
| LMArena Text: writing and language | 1496 | 1542 | Gemini 4 Carbon |
| LMArena Text: entertainment and media | 1484 | 1527 | Gemini 4 Carbon |
| LMArena Vision: English | 1315 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Vision: Chinese | 1411 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Vision: reading text | 1332 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Vision: humor | 1329 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Vision: creative writing | 1320 | Not measured | Only Gemini 3.8 Flash measured |
| LMArena Document | 1482 | Not measured | Only Gemini 3.8 Flash 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 | Gemini 3.8 Flash | Gemini 4 Carbon | 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
Gemini 3.8 Flash$0.0026Fits in one request
Gemini 4 Carbon$0.0070Fits in one request
Gemini 3.8 Flash costs less
Repository review
50K input + 3K output tokens
Gemini 3.8 Flash$0.049Fits in one request
Gemini 4 Carbon$0.130Fits in one request
Gemini 3.8 Flash costs less
Agent loop
200K input + 10K output tokens
Gemini 3.8 Flash$0.188Fits in one request
Gemini 4 Carbon$0.500Fits in one request
Gemini 3.8 Flash 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 Gemini 3.8 Flash | Google Gemini 4 Carbon |
|---|---|---|
| Maker | ||
| Weights | No public weights listed | No public weights listed |
| Released | September 2, 2026 | November 10, 2026 |
| Context window | 1.05M | 1.05M |
| Input $/Mtok | $0.75 | $2.00 |
| Output $/Mtok | $3.75 | $10.00 |
| Coding index | 60.3 | 86.6 |
| Independent measurements behind it | 6 | 4 |
| Cost per solved task | $2.74 | Not computable |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is Gemini 3.8 Flash better than Gemini 4 Carbon for coding?
Pick Gemini 4 Carbon for coding: it scores 86.6 on the Vibe Coding Index against 60.3 for Gemini 3.8 Flash, though on 4 benchmarks against 6. Pick Gemini 3.8 Flash for output tokens 2.7x cheaper ($3.75 vs $10.00 per million) and work measured by Long context reasoning (AA-LCR), where it leads.
Which is cheaper, Gemini 3.8 Flash or Gemini 4 Carbon?
Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens; Gemini 4 Carbon costs $2.00 and $10.00. Cost per solved task is known for Gemini 3.8 Flash only.
Which has the bigger context window, Gemini 3.8 Flash or Gemini 4 Carbon?
Neither: both take 1.05M tokens.
How many benchmarks is the Gemini 3.8 Flash vs Gemini 4 Carbon comparison based on?
71 benchmarks have a result for both models as of November 14, 2026. Gemini 3.8 Flash's coding index rests on 6 independent benchmarks and Gemini 4 Carbon's on 4. Where only one model has a result, the page says so instead of naming a winner.