ModelsGemini 3.8 Flash

Gemini 3.8 Flash

Google’s Gemini 3.8 Flash model for multimodal reasoning and tool use.

GoogleReleased 2 September 2026Private weights
Context window
1M tokens
Maximum output
66k tokens
Input / output
$0.75 / $3.75USD per million tokens
Tests with results
5062 source-specific results
AA Intelligence Index
41.00Evaluator composite, own scale

Test coverage

  • 16aa
  • 1arc prize · standard
  • 2scale ai / seal
  • 29vals
  • 1datacurve
  • 1self
  • 115 no result

165 catalogue tests. Each test counts once; evaluator results take precedence where sources overlap.

Results

Filter the recorded results; open Details on a row for run conditions.

62 results

BenchmarkResultSourceRun
τ³-Banking44.95%aa
Details

Highest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.

Metric
τ³-Banking
Read
2026-09-30
Effort
high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
ITBench-AA52.54%aa
Details

Highest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.

Metric
ITBench-AA
Read
2026-09-30
Effort
high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
MLCR-AA21.67%aa
Details

Highest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.

Metric
MLCR-AA
Read
2026-09-30
Effort
high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
ARC-AGI-310.37%· Standardarc prize · standard
Details

Verified semi-private set; best published configuration per release within this scaffold. Public demo and community submissions excluded. Standard and adapter results remain separate.

Metric
ARC-AGI-3 · Standard
Read
2026-09-30
Effort
Standard
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
35.00%· Provider adapterarc prize · provider adapter
Details

Verified semi-private set; best published configuration per release within this scaffold. Public demo and community submissions excluded. Standard and adapter results remain separate.

Metric
ARC-AGI-3 · Provider adapter
Read
2026-09-30
Effort
Provider adapter
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
SWE-Bench Pro V2 Full94.86%scale ai / seal
Details

V2 FULL field, as published. Different agent scaffolds and effort settings are retained per row; this is a model-and-agent comparison.

Metric
SWE-Bench Pro V2 Full
Read
2026-09-22
Effort
(mini-swe-agent) high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
SWE-Bench Pro V2 HARD58.80%scale ai / seal
Details

V2 HARD field, as published. Different agent scaffolds and effort settings are retained per row; this is a model-and-agent comparison.

Metric
SWE-Bench Pro V2 HARD
Read
2026-09-22
Effort
(mini-swe-agent) high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
Vals RSI Index19.93%vals
Details

Can a model do the research that builds the next model? Autonomous LLM R&D scored against frozen published, human, and frontier-model references.

Metric
Vals RSI Index
Read
2026-09-29
Effort
high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
Vibe Code Bench 1–10018.77%vals
Details

Can models extend a working web application across a long sequence of dependent requests?

Metric
Vibe Code Bench 1-100
Read
2026-09-29
Effort
high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
CUA-bench4.17%vals
Details

Can an AI agent play six commercial video games with only a keyboard and a mouse? Half of them are held out.

Metric
CUA-bench
Read
2026-09-29
Effort
high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
AutomationBench-AA59.93%aa
Details

Highest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.

Metric
AutomationBench-AA
Read
2026-10-24
Effort
high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
SciCode56.60%aa
Details

Highest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.

Metric
SciCode
Read
2026-10-24
Effort
high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
GDP.pdf21.00%aa
Details

Highest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.

Metric
GDP.pdf
Read
2026-10-24
Effort
high
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →

API & pricing

Open one section for prompt tiers, runtime rates, endpoint measurements and effort configurations.

Published pricing, runtime and evaluator details
Current pricing

The listed $0.75 input / $3.75 output price applies through 31 December 2026. Google lists $1.50 / $7.50 from 1 January 2027.

Read source ↗

Documentation read 30 September 2026.

No separate hosted-runtime record was checked for this release.

AA API performance and evaluation cost
Output speed
238.736904635898 tokens/s
First answer token
20.7603934870001 seconds
500-token response
22.854749199463406 seconds
Index cost per task
$1.2427947606950427

Artificial Analysis endpoint measurements. Provider, effort and snapshot may differ from the official API specifications.

Evaluator source ↗

Original sources