Atlas

Models

← All models

Gemini 3.5 Flash

Google · 2026-05-19

Gemini 3.5 Flash is Google's generally available, natively multimodal reasoning model released at Google I/O 2026 for agentic workflows, long-horizon tasks, and coding. Google reports that it exceeds Gemini 3.1 Pro on Terminal-Bench 2.1, MCP Atlas, and GDPval-AA, while generating output tokens four times faster than other frontier models.

Benchmark scores

BenchmarkScore
AA-AnalystAgent (pass@1)58.5
AA-AnalystAgent (pass@5)76.3
AA-AnalystAgent (pass^5)45.0
AA-Briefcase Elo871
AA-Briefcase v1.1 Elo877
AA-LCR71.0
AA-LCR v1.174.3
AA-Omniscience Index21.2
ActiveVision8.2
AdvancedIF73.7
AdvancedIF (rubric mean score)94.1
Agent Red Teaming k=100 (CAIS)72.3
Aikido Known-CVE Detection (Pass@3 Union)53.8
Aikido Known-CVE Detection (Single-Run Average)44.9
AIME 202581.7
AIME 202695.0
ALE-Bench1712
Antidote: Everyday Edition1079
Apex (MathArena)32.3
APEX (Mercor)15.2
Apex Shortlist82.5
APEX-Agents 1.144.3
APEX-Agents 1.1 Pass@127.5
APEX-Agents-AA (avg@3)47.1
APEX-SWE36.1
APEX-v1-extended62.2
ARC-AGI-1 Semi-Private92.5
ARC-AGI-272.1
Arena Code WebDev — Brand & Marketing1493
Arena Code WebDev — Consumer Product1470
Loading Atlas data…