Inference
EN

Gemini 3.8 Flash vs Gemini 4 Carbon

Change models
Gemini 3.8 FlashGoogle
Gemini 4 CarbonGoogle
IndexAll categories, out of 100.
67.8 ±2.5
82.7 ±5.1 best
GapIndex points below the top.
−14.9
Top
CoverageIndex weight that has results.
95% best
80%
SpeedOutput tokens per second.
68/s best
31/s
EloLMArena rating and rank.
1500
1538 best
ReleasedMonth of release.
2026-09
2026-10
Input / 1MUS dollars per 1M input tokens.
$0.825 best
$2.00
Output / 1MUS dollars per 1M output tokens.
$4.13 best
$10.00
Batch output / 1MBatch price per 1M output tokens.
$1.88 best
$5.00
1K chats2K tokens in, 500 out, each.
$3.71 best
$9.00
1K documents20K tokens in, 1K out, each.
$20.63 best
$50.00
1K agent tasks200K tokens in, 20K out, each.
$247.50 best
$600.00
ContextMaximum tokens in one request.
1.05M
1.05M
AgenticMulti-step tasks with tools.
65.3
83.2 best
CodingCode writing and repair.
66.2
80.5 best
ReasoningLogic problems and puzzles.
68.3
83.9 best
MultimodalTasks with images and text.
67.1
—
KnowledgeFacts and expert knowledge.
70.9
79.5 best
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
91.3
—
MathMath problems.
64.3
77.2 best

Benchmarks

29 shared
Only Gemini 3.8 Flash: 77Only Gemini 4 Carbon: 10
Gemini 3.8 Flash
Gemini 4 Carbon
Artificial Analysis Long Context ReasoningReasoning
81.3% best34 / 217
79.3%52 / 217
GPQA diamond · Epoch AIKnowledge
95.4%5 / 217
96.5% best1 / 217
OTIS Mock AIME 2024-2025Math
98.9%22 / 200
100.0% best1 / 200
Artificial Analysis Omniscience AccuracyKnowledge
50.5%19 / 164
58.3% best9 / 164
ARC-AGI-2 (semi-private)Reasoning
89.2%11 / 131
92.1% best5 / 131
ARC-AGI-1 (semi-private)Reasoning
98.5% best1 / 126
97.5%5 / 126
Vibe Code Bench v1.1Coding
78.7%34 / 123
89.1% best9 / 123
ARC-AGI-2 (public eval)Reasoning
92.5%11 / 122
97.9% best3 / 122
ARC-AGI-1 (public eval)Reasoning
98.0%11 / 116
98.5% best6 / 116
SimpleQA VerifiedKnowledge
69.7%10 / 91
77.7% best2 / 91
Chess PuzzlesReasoning
58.0%14 / 90
78.0% best2 / 90
FrontierMath-Tiers-1-3-v2-PrivateMath
68.4%32 / 90
90.5% best6 / 90
Legal Research BenchLegal
38.9%38 / 90
55.8% best1 / 90
Tax Agent BenchFinance
32.1%23 / 82
51.8% best1 / 82
Mystery Game PuzzlesReasoning
47.0%16 / 80
73.0% best5 / 80
FrontierMath-Tier-4-v2-PrivateMath
22.0%48 / 72
95.1% best4 / 72
Terminal-Bench 4.0Agentic
19.7%34 / 61
63.1% best4 / 61
Agent Arena steerabilityAgentic
−1.333 / 54
19.1 best2 / 54
Agent Arena task outcomeAgentic
1.329 / 54
18.4 best4 / 54
Furniture AssemblyReasoning
31.7%27 / 37
77.5% best6 / 37