Google DeepMindmodel systemcard

Gemini 3.6 Flash

#14of 27 · general

1Mcontext

Proprietarynon-commercial

2026-07-21released

2026-03-31knowledge cutoff

Standing

dot = this model · shaded = the field's middle 60% · track = full field

General

48.8#14/27

Math

45.1#12/27

Coding

41.2#13/27

Health

42.1#14/26

Engineering

37.5#11/17

Cyber

15.4#12/14

Legal

30.4#21/26

Instruct

64.4#6/27

Agents

49.2#12/27

Professional

64.6#7/27

Multimodal

51.2#8/19

Hallucination

47.3#8/27

Speed

57.9#8/22

Value

47.6#14/27

Measured

≈ imputed · amber dot = lab-self-reported among the sources

GPQA Diamond93.4%#5

HLE40.8%#16

MMLU-Pro89.3%#6

LiveBench73.6%#14

SWE-bench V79.6%#9

LiveCodeBench88.1%#6

SciCode52.7%#14

CritPt10.6%#17

Terminal-Bench75.7%#9

GDPval46.1%#13

Arena1482#7

AA Index51.6#13

Vals Index62.4#13

Epoch Capabilities Index154.0#13

Operations

$1.5 / $7.5$ per 1M tokens, in / out

$3blended

230tokens / sec

$0.24$ per solved task

1482arena elo · 6,065 votes

Series

DeepMind's models on the board, by release

Gemini 3.1 Pro2026-02-19 · 55Gemini 3.5 Flash2026-05-19 · 47Gemini 3.6 Flash2026-07-21 · 49

Neighbors

nearest capability fingerprints · deltas describe the neighbor

Qwen3.8 Max→ weaker multimodal Grok 4.5→ weaker multimodal Gemini 3.5 Flash→ near-twin Gemini 3.1 Pro→ follows instructions better · pricier MiMo V2.5 Pro→ weaker agent · cheaper