HOST DIAGNOSTIC · Google
Gemini 3.5 Flash
Gemini 3.5 Flash· Gemini· PROPRIETARY
ACTIVE
HOST IDENTIFICATION
Google
Gemini 3.5 Flash
ACTIVE
AVG SCORE
[15.5]
STRONG
STRONGEST
Speed
[18]
WEAKEST
Calibration
[13]
MAXED (20/20)
0
of 20 attributes
20-DIMENSION EVALUATION
ATTRIBUTE SCORES · 20 DIMENSIONS
Cognitive
Bulk Apperception
16
Reasoning
15
Mathematical Precision
15
World Knowledge
15
Scientific Acumen
15
Technical
Code Generation
16
Tool Use
17
Multimodal Fluency
17
Speed
18
Cost Efficiency
16
Behavioral
Candor
14
Creativity
14
Tenacity
15
Self-Correction
15
Calibration
13
Operational
Instruction Following
16
Context Fidelity
16
Multi-turn Coherence
15
Planning and Execution
17
Safety Alignment
14
SCORING METHODOLOGY
CALIBRATED. Google, launched 2026-05-19 at Google I/O 2026 (GA). Closed-weights. $1.50/$9 per M (3x the prior Gemini 3 Flash Preview); 1M context; text+image+audio+video in; ~280+ output t/s (Flash-tier, ~4x faster than frontier peers per Google). Thinking-level configurable (minimal/low/medium/high); benchmarked at 'high'. INDEPENDENT SOURCES: Artificial Analysis Intelligence Index = 55 (v4.0/v4.1), rank ~#7-8 of 147-148 models (+9 vs Gemini 3 Flash's 46; well above ~36 median for tier). Profile = agentic/coding/multimodal LEADER at Flash speed, but TRAILS frontier on pure reasoning. Agentic/coding: MCP Atlas 83.6% (LEADS field), Terminal-Bench 2.1 76.2% (2nd behind GPT-5.5 78.2%), SWE-Bench Pro 55.1% (edges 3.1 Pro 54.2%; Opus 4.7 leads at 64.3%), OSWorld 78.4% (~par GPT-5.5/Opus 4.7), Finance Agent v2 57.9% (LEADS, margin over GPT-5.5 51.8%), GDPval-AA 1656 Elo (well above 3 Flash 1204 + 3.1 Pro 1314; just behind GPT-5.4 1674). Multimodal: CharXiv Reasoning 84.2% (LEADS, edges GPT-5.5 84.1%). Reasoning (weaker): ARC-AGI-2 72.1% (huge +38.5 vs 3 Flash 33.6%, but GPT-5.5 leads 84.6%), Humanity's Last Exam 40.2% (trails 3.1 Pro 44.4% + Opus 4.7 46.9%), MRCR v2 128k 77.3% (trails 3.1 Pro). DIMENSION RATIONALE: Tool Use 17 + Planning 17 (MCP Atlas + Finance Agent leadership = best-in-Gemini-3-series agentic); Code 16 (between 3.1 Pro and GPT-5.5); Multimodal 17 (CharXiv lead, native A/V/image); Speed 18 (Flash-tier 280 t/s); Reasoning 15 + Scientific 15 (ARC-AGI-2 + HLE place it below the reasoning frontier despite agentic strength); Cost Efficiency 16 (cheaper than GPT-5.5 by ~3x but a 3x price hike vs prior Flash tempers this). Anchored above Gemini 3 Flash + 3.1 Flash Lite, below 3.1 Pro on reasoning but above it on agentic/coding. NOTE: 61% hallucination rate flagged by AA (Candor 14, Calibration 13). WATCH UPDATE (2026-08-11, CCH): Gemini 3.5 Pro STILL NOT GA — the widely-reported July 17 target slipped; as of 2026-08-06 the public Gemini API Pro-class tops out at Gemini 3.1 Pro, DeepMind's Pro page still shows '3.5 Pro coming soon', and 3.5 Pro has only surfaced in stealth Arena previews (Jul 31 ~30-min appearance; Aug 5-6 backend-tag + Arena sightings) — launch-imminent signal, remains a watch item, no record created per protocol. Sibling GA'd instead: Gemini 3.6 Flash + 3.5 Flash-Lite went GA 2026-07-21 per Gemini API release notes; Gemini 3.6 Flash record created 2026-08-11 (Parent Model = this record).