RankingGemini 3.8 Flash

Gemini 3.8 Flash

Data updated 22 Oct 2026

33 published benchmark measures · 9 benchmark families contribute across 4 task areas. 4 capability estimates available in the full profile. See all results ↓ · Compare published benchmarks →

Model evidence summary

This profile combines published settings. It is not a runnable configuration or a leaderboard rank. Compare measured configurations →

Performance profile

Capabilities

Adjusted comparison score · 0–100

Filled points are supported; hollow points are preliminary. Lines stop at unknown capabilities. Exact values and sources follow below.

Agentic2 families · 0 with independent evidence · Supported55.0

2 core families; 5 direct opponents across 3 labs.

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: 51.4–59.2; 2/13 scenarios unsupported. Without one publisher: 45.0–58.9; 1/15 unsupported. Smoothing check: 54.2–55.0; 0/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts improved on neutral predictions within evaluated coverage; source sensitivity still requires inspection. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

GDPval: 40.0 observed win share
Google · Source 1
GDP.pdf: 60.0 observed win share
Google · Source 1

Hard reasoning3 families · 3 with independent evidence · Supported87.1

3 core families; 65 direct opponents across 11 labs.

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: 81.2–91.6; 0/6 scenarios unsupported. Without one publisher: 77.7–90.8; 0/16 unsupported. Smoothing check: 78.0–91.2; 0/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts improved on neutral predictions within evaluated coverage. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

GPQA: 85.0 observed win share
OpenAI, artificialanalysis.ai · Source 1 Source 2
ARC-AGI: 85.6 observed win share
arcprize.org · Source 1
Humanity’s Last Exam: 93.8 observed win share
Google, scale.com · Source 1 Source 2

Coding2 families · 2 with independent evidence · Supported58.3

2 core families; 35 direct opponents across 9 labs.

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: 58.3–67.0; 2/6 scenarios unsupported. Without one publisher: 53.7–64.3; 0/19 unsupported. Smoothing check: 57.3–58.3; 0/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts improved on neutral predictions within evaluated coverage. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Terminal-Bench: 22.9 observed win share
Google, OpenAI, tbench.ai · Source 1 Source 2 Source 3
DeepSWE: 92.1 observed win share
Google, OpenAI, deepswe.datacurve.ai · Source 1 Source 2 Source 3

Human prefNo comparable evidenceUnknown

0 core families; 0 direct opponents across 0 labs. No admitted core comparisons

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 1/1 scenarios unsupported. Without one publisher: No supported estimate; 1/1 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: cross-family and cross-publisher predictions could not be evaluated from the available comparison network. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

KnowledgeNo comparable evidenceUnknown

0 core families; 0 direct opponents across 0 labs. No admitted core comparisons

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 4/4 scenarios unsupported. Without one publisher: No supported estimate; 5/5 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: almost no publisher-held-out comparisons remained estimable. Cross-source validity is unresolved. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Multimodal2 families · 0 with independent evidence · Supported93.3

2 core families; 5 direct opponents across 3 labs.

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: 92.7–94.5; 2/8 scenarios unsupported. Without one publisher: 76.9–95.4; 1/8 unsupported. Smoothing check: 84.8–96.9; 0/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: limited publisher-held-out coverage. Cross-source validity is unresolved. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

CharXiv: 100.0 observed win share
Google · Source 1
Lvbench: 100.0 observed win share
Google · Source 1

Long contextNo comparable evidenceUnknown

0 core families; 0 direct opponents across 0 labs. No admitted core comparisons

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 5/5 scenarios unsupported. Without one publisher: No supported estimate; 5/5 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts underperformed neutral predictions. Treat this as an exploratory estimate. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Scores estimate outcomes against a shared reference panel; they are not accuracy percentages. Sparse or disconnected evidence cannot qualify an overall profile. Open a capability to inspect its evidence.

Reported effort · High thinking + unspecified

Settings reported in this model's published benchmark results, including results outside the aggregate. Effort names are provider-specific. These are not API defaults or equal compute budgets.

High thinking: 1 observations
Not specified: 36 observations

Mixed settings means multiple settings occur in the evidence. Best across efforts means the source selected its best reported result across settings; it does not mean Max. Unspecified settings stay unknown. This model-summary chart combines reported settings. The leaderboard keeps identified configurations separate and excludes unknown effort.

Inspect each result and its source ↓ · Download effort evidence

Score contributions and missing evidence

9 contributing families across 4 capabilities. Fixed reference panels do not change when the catalog expands.

Capability is fitted jointly across families. Capability estimates below describe different task areas; their weighted sum is not the Capability score.

Results without reviewed compatibility or a reference match remain in the raw evidence below. Coverage counts only contributing results.

Model information & shareable badge

Lab: Google
Catalog status: active
Availability: Public provider catalog; account and region restrictions may apply
Family: Gemini 3
Released: —
Context: —
License: proprietary
Model card: https://ai.google.dev/gemini-api/docs/models
Default Capability family coverage: /badge/gemini-3.8-flash.svg

Benchmark scores & sources

Original results, evaluation harnesses, and evidence behind this model.

Benchmark	Bucket	Score	Harness	Evidence	Source-recorded date
Humanity's Last Exam	Hard reasoning	44.52%	HLE no tools Effort: Not specified contributes to capability	official board	2026-09-09
Humanity's Sixth Sense	Supporting evidence	41.6%	HSS · high Effort: high Reported settings Full 522-task pass@1, task mean of three attempts; every rubric criterion must pass. 95% task-cluster bootstrap CI. Claude-Opus-5 judge. Scale AI + Elorian. Unversioned board; source hashes identify captures. Display only. Excluded from weighted Overall. Source-recorded date is the capture date, not a proven evaluation date.	official board Supporting evidence outside the reviewed capability core	2026-10-22
LABBench · 2	Supporting evidence	86.2% Reported settings & source Selfcomputed; Linuxterminal,bioinfotools,Python,R,network; macroaverage11subtasks Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Narrow specialist biological, medical or security evaluation; preserved as raw evidence outside general capability scoring.	Reviewed 2026-09-06
LMArena Agent Arena	Supporting evidence	3.57	LMArena Agent Arena · high Effort: Not specified	official board Supporting evidence outside the reviewed capability core	2026-10-20
LVBench · agentic	Multimodal	87.8% Reported settings & source Card labels agentic; linkedmethodology describes only no-tools static setup Agentic tool/protocol details unresolved; rawreportedresult only. Comparison limit: Card reports agentic result but linked methodology describes only static no-tools protocol. Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Card reports agentic result but linked methodology describes only static no-tools protocol.	Reviewed 2026-09-06
LVBench · static	Multimodal	87.1% Reported settings & source No tools;1024frames Gemini/GPT,300frames Claude dueAPIlimit; model-specific frame budget: 1024 Frame budgets differ; table labels Gemini3.8static explicitly. Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
OSWorld partial · 2.0 pre-08.08	Agentic	59% Reported settings & source Partialscore; batchtools;1080p/500steps; Gemini/Sonnet bestof3runs; screenshotonly; officialCUAharness Methodology says runs pre08.08patch but Opusvalue fromFable5.1card usesAugustfixedtasks; no controlledsameversionclaim. GPT values providerreports. Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Mixed source task revisions and best-of-three versus provider reporting prevent a uniform common-core unit.	Reviewed 2026-09-06
Real-SWE	Supporting evidence	38.75%	Real-SWE · Gemini 3.8 Flash · Gemini CLI Effort: Not specified	official board Supporting evidence outside the reviewed capability core	2026-10-13
Terminal-Bench · 2.1	Coding	89.4% Reported settings & source Terminus2only; Gemini selfcomputed, othermodels officialboard/ArtificialAnalysis Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Terminal-Bench coding slot admits version 4.0 only. Earlier versions remain stored and are not mixed into that slot.	Reviewed 2026-09-06
Terminal-Bench · 4.0	Coding	19.1% Reported settings & source Officialpublicboard highest scoring thinking level; nativeagents may differ Gemini3.8Flash Model Card · Model card page5 Results table · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
Terminal-Bench 4.0	Coding	19.09%	Terminal-Bench 4.0 reported Effort: Not specified contributes to capability	official board	2026-09-02
Terminal-Bench 4.0 · 4.0	Coding	19.1% Reported settings & source Maximum reported across reasoning efforts; OpenAI research environment or API Provider-published result; comparator measurements are not automatically independently reproduced. GPT-6 Astra: A new generation of intelligence · Coding table / Terminal-Bench 4.0 / Gemini 3.8 Flash · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
GDPval-AA	Agentic	—	—	—	—
LiveCodeBench	Coding	—	—	—	—
LMArena Text Arena	Human pref	—	—	—	—
MMLU-Pro	Knowledge	—	—	—	—
OSWorld-Verified	Agentic	—	—	—	—
SWE-bench Pro	Agentic	—	—	—	—
SWE-bench Verified	Agentic	—	—	—	—
Terminal-Bench 2.1	Agentic	—	—	—	—