# Evaluation settings and benchmark-specific methodology are linked from the launch post.
# Methodology: https://deepmind.google/models/evals-methodology/gemini-3-8-flash
# Sources:
# - https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash
# - https://ai.google.dev/gemini-api/docs/latest-model
name = "Gemini 3.8 Flash"
description = "Google's most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows"
family = "gemini-flash"
release_date = "2026-09-02"
last_updated = "2026-09-02"
attachment = true
reasoning = true
temperature = true
tool_call = true
structured_output = true
open_weights = false

[limit]
context = 1_048_576
output = 65_536

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

[[benchmarks]]
name = "DeepSWE"
score = 73.7
metric = "pass@1"
version = "1.1"
harness = "mini-swe-agent"
variant = "high thinking"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "GDPval-AA"
score = 1545
metric = "Elo"
version = "2"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "Finance Agent"
score = 61.4
metric = "pass@1"
version = "2"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "Harvey Legal Agent Benchmark"
score = 10
metric = "all-pass rate"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "Terminal-Bench"
score = 89.4
metric = "pass@1"
version = "2.1"
harness = "Terminus 2"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "Terminal-Bench"
score = 19.1
metric = "pass@1"
version = "4.0"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "GDP.PDF"
score = 35
metric = "all-pass rate"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "CharXiv Reasoning"
score = 86.2
metric = "pass@1"
variant = "without tools"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "LVBench"
score = 87.8
metric = "pass@1"
variant = "agentic"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "LVBench"
score = 87.1
metric = "pass@1"
variant = "static; 1024 frames; without tools"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "HLE-Verified"
score = 54.9
metric = "accuracy"
dataset = "1811 verified and revised items"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "OSWorld"
score = 59
metric = "partial score"
version = "2.0"
dataset = "before 2026-08-08 patch"
harness = "Gemini CUA"
variant = "batched tool calls; best of 3 runs; 500 steps"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "BioMysteryBench"
score = 88.8
metric = "pass@1"
dataset = "human-solvable"
variant = "with tools"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "BioMysteryBench"
score = 56.5
metric = "pass@1"
dataset = "human-difficult"
variant = "with tools"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"

[[benchmarks]]
name = "LAB-Bench"
score = 86.2
metric = "macro-average"
version = "2"
dataset = "11 subtasks"
variant = "with tools"
source = "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/"