name = "Gemini 3.6 Flash"
description = "Fast Gemini model balancing multimodal reasoning, tool use, and cost"
family = "gemini-flash"
release_date = "2026-07-21"
last_updated = "2026-07-21"
attachment = true
reasoning = true
temperature = true
tool_call = true
structured_output = true
knowledge = "2026-03"
open_weights = false

[limit]
context = 1_048_576
output = 65_536

[modalities]
input = ["text", "image", "video", "audio", "pdf"]
output = ["text"]

[[benchmarks]]
name = "SWE-Bench Pro"
score = 58.7
metric = "resolve rate"
harness = "Antigravity"
source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
date = "2026-07-21"

[[benchmarks]]
name = "DeepSWE"
score = 49.0
metric = "resolve rate"
variant = "high reasoning"
version = "1.1"
source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
date = "2026-07-21"

[[benchmarks]]
name = "Terminal-Bench"
score = 78.0
metric = "accuracy"
harness = "Terminus 2"
version = "2.1"
source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
date = "2026-07-21"

[[benchmarks]]
name = "MLE-Bench"
score = 63.9
metric = "average position score"
dataset = "Partial 30"
source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
date = "2026-07-21"

[[benchmarks]]
name = "GDPval-AA"
score = 1421
metric = "Elo"
version = "v2"
source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
date = "2026-07-21"

[[benchmarks]]
name = "OSWorld-Verified"
score = 83.0
metric = "success rate"
source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
date = "2026-07-21"

[[benchmarks]]
name = "CharXiv Reasoning"
score = 85.2
metric = "accuracy"
variant = "no tools"
source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
date = "2026-07-21"

[[benchmarks]]
name = "CharXiv Reasoning"
score = 89.4
metric = "accuracy"
variant = "with tools"
source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
date = "2026-07-21"

[[benchmarks]]
name = "GDM-MRCR"
score = 91.8
metric = "accuracy"
variant = "128k average, 8-needle"
version = "v2"
source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
date = "2026-07-21"

[[benchmarks]]
name = "GDM-MRCR"
score = 54.0
metric = "accuracy"
variant = "1M pointwise, 8-needle"
version = "v2"
source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
date = "2026-07-21"