Google DeepMind

Gemini 3.6 Flash

Released Jul 21, 2026 · knowledge cutoff 2026-03

Status: Frontier
Location: United States / United Kingdom
Modality: Multimodal
Context window: 1M
Max output: 66K
Speed (high effort): 214 tok/s · 16s to first answer token
Price ($/MTok in / out): $1.50 / $7.50
Cost per task (high effort): $0.50
Open weights: No

Benchmarks4 of 8 reported

GPQA Diamond

92.8%

Terminal-Bench 2.1

78%

Humanity's Last Exam

38.3%

LMArena Elo

1485

MMLU-Pro

—

SWE-bench Verified

—

AIME (latest)

—

ARC-AGI-2

—

Strengths & weaknesses

Strengths

Sticks to the change you asked for instead of touching adjacent files during diagnostic work
Reaches the same result with fewer output tokens than the Flash before it
Available across Google's whole developer and enterprise surface on day one

Weaknesses

Flash-tier depth ceiling — it executes well but doesn't crack genuinely hard reasoning
Arrived with no matching Pro-tier model in its generation, so there was nothing to escalate to

Model card reports SWE-Bench Pro 58.7%, DeepSWE v1.1 49%, GDPval-AA v2 Elo 1421, and GDM-MRCR v2 91.8% at 128k; Google published none of the site's tracked benchmarks at launch. Artificial Analysis Intelligence Index 50 (third-party). GPQA Diamond 92.8% and HLE 38.3% are Artificial Analysis-measured, not Google-reported. LMArena Elo 1485 is the 'gemini-3.6-flash' text-arena listing (rank 12, arena.ai); it also sits #12 in the Frontend Code Arena at 1537. Announced 2026-07-21 alongside Gemini 3.5 Flash-Lite and the limited-access Gemini 3.5 Flash Cyber.

For developers

API model strings

googlegemini-3.6-flash

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor