Thoughts on AI, technology, and the future we're building.
New posts every week
A model-level benchmark profile for Gemini 3 Flash, combining Arena category ranks, suite indexes, commercial price and speed data, SWE-bench status, and lab context.
Lab
Proprietary
Coverage
78%
7 of 9 tracked signals
Best Arena rank
#17
Vision Arena
Best suite
#14
Value / 73
Model profile
This view treats missing values as coverage gaps, then puts every available score, rank, cost, and speed signal into the same visual hierarchy.
Analyst view
Its best composite result is rank #14 in the value view, with a 73 value index. The strongest raw Arena signal is Vision Arena at rank #17 with 19,319 public votes.
Best current lane
Value
#14 / 73 value index
Strongest raw signal
Vision
#17, 1260 score
Commercial posture
$1.13
0.92s first token / 191 tok/s
Coverage gaps
Profile completeness
Arena categories
3/5
Vision, Text, Code
Commercial match
Live
Gemini 3 Flash Preview (Non-reasoning)
SWE-bench
#2
bash-only
Raw identifier
gemini-3-flash (thinking-minimal)
Performance cockpit
Arena rows show score, public rank, confidence interval, vote count, and publish date. Missing rows stay visible as coverage gaps.
Text Arena
Arena category
1458-1465 CI
WebDev Arena
Arena category
1381-1393 CI
Vision Arena
Arena category
1253-1267 CI
Search Arena
Arena category
Not listed
No confident public Arena row for this category.
Coverage gap
Document Arena
Arena category
Not listed
No confident public Arena row for this category.
Coverage gap
SWE-bench Verified
Software issue resolution
bash-only
Cost and speed
Price, latency, and throughput are only shown when a confident commercial source row is matched.
Blended price
$1.13
USD / 1M tokens
Input price
$0.500
USD / 1M input tokens
Output price
$3.00
USD / 1M output tokens
Latency
0.92s
Median time to first token
Output speed
191 tok/s
Median generation speed
Matched source row
Gemini 3 Flash Preview (Non-reasoning)Suite position
These views combine available public signals into purpose-specific indexes. A missing card means the model lacks enough source coverage for that lens.
Value
Quality, price, latency, and output speed
value index
Coding
Code Arena, SWE-bench, and output speed
coding index
Agents
Search, documents, code, and latency
agent proxy
Reasoning
Text, documents, vision, and mode signal
reasoning proxy
Long context
Document, search, and text benchmark proxy
context proxy
Lab context
Same-lab models make it easier to see whether this profile is a flagship, value play, specialist, or lower-coverage row.
Gemini 3.1 Pro Grounding
gemini-3.1-pro-grounding
Best rank
#4
Coverage
11%
Gemini 3.1 Pro Preview
gemini-3.1-pro-preview
Best rank
#6
Coverage
67%
Gemini 3 Pro
gemini-3-pro
Best rank
#6
Coverage
56%
Gemini 3 Pro Grounding
gemini-3-pro-grounding
Best rank
#7
Coverage
11%
Gemini 3.5 Flash
gemini-3.5-flash
Best rank
#10
Coverage
56%
Gemini 3 Flash Grounding
gemini-3-flash-grounding
Best rank
#10
Coverage
11%
Gemma 4 31b
gemma-4-31b
Best rank
#23
Coverage
56%
Gemini 2.5 Pro Grounding
gemini-2.5-pro-grounding
Best rank
#23
Coverage
11%