Gemini 3.8 Flash: score and benchmark evidence
Historical release · · Google
102.8 points · Rank 9 among 19 ranked models · Stability range 90.4–113.4 points.
Rank follows the point estimate; it does not establish statistically significant superiority. The scale is anchored at mean 100, SD 15 in the frozen calibration cohort, not human IQ or a percentage.
One capability score combines evidence across seven equally weighted domains. Related benchmark variants share a family budget, and influence from one evaluator is capped.
Release tli-2026-v1.0-2026-09-06-db. Methodology and limitations for this release.
This page preserves historical evidence. See the current model record.
Capability domains
Domain scores use calibration-cohort standard-deviation units. Unmeasured domains are shown as unavailable, not zero.
| Domain | Score (SD units) | Observed benchmarks |
|---|---|---|
| Knowledge & reasoning | 0.33 | 6 |
| Coding & software engineering | 0.04 | 12 |
| Agentic tool & computer use | 0.26 | 4 |
| Professional & real-world work | -0.20 | 14 |
| Multimodal & vision | -0.04 | 4 |
| Cybersecurity | Unavailable | 0 |
| Preference & communication | Unavailable | 0 |
Published benchmark evidence
40 results contribute to this score, from 52 collected results and 39 contributing benchmark families. Counts are not independent sample sizes.
Results, source citations, review decisions and compatible reasoning modes are stored in a versioned relational database. Each publication uses a sealed, checksum-verified export; CSV files are optional exports, not editable scoring inputs. Previous revisions remain available for audit. Collected means a numeric result exists in the database. Used means its source, exact model checkpoint and benchmark protocol have been reviewed and its benchmark has a frozen calibration supported by multiple providers. For each benchmark we select the highest published aggregate across comparable reasoning modes, not necessarily maximum effort, and count it once. More tested settings provide more chances to record a high score; this selection advantage is not corrected by the conditional interval. Developer-reported and sole/unspecified settings remain eligible and disclosed. Unresolved does not mean false or useless: conflicting metrics, unverified citations and unmatched versions stay collected until reconciled.
The bar estimates score stability under the scoring assumptions: a fixed calibration and comparable benchmark availability. It resamples benchmark families and evaluators while keeping each domain’s evidence strength fixed, then adds a frozen correction learned from missing-evidence tests. Benchmark count and domain coverage both affect that correction. It does not add a second full capability-posterior draw or uncertainty from changing the reference calibration. It is not a guarantee under selective publication, new benchmark domains or future model releases.
Reporting sensitivity: ±1.3 points. Sensitivity to selectively published results. This stress range is not a confidence interval or a probability statement.
GDPval-AA v2 (Elo): 1464.96 native-points
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash (high); Artificial Analysis published evaluation harness
Metric: Elo
Original source · Retrieved 2026-09-06
Best published result among 3 reasoning configurations of the exact AA model release; fallback configurations excluded.
Terminal-Bench 2.1: 87.640449%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash (high); Artificial Analysis published evaluation harness
Metric: pass@1, AA harness
Original source · Retrieved 2026-09-06
Best published result among 3 reasoning configurations of the exact AA model release; fallback configurations excluded.
DeepSWE v1.1: 73.825503%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: gemini-3-8-flash; high; mini-swe-agent
Metric: pass@1 over scored rollout attempts
Original source · Retrieved 2026-09-06
Selected the best pass@1 across published reasoning efforts within mini-swe-agent. Pass@4 is not substituted.
LMArena (Chatbot Arena) — Text: 1494 native-points
Not used in the score. Source category not recorded. Review status: unresolved.
Not admitted after the September 6 model review: the collected 1494 for Gemini 3.8 Flash on LMArena (Chatbot Arena) — Text still lacks a matched source entry establishing its checkpoint and benchmark-specific metric, tools and retry setup. This is a citation/comparability gap, not a finding that the result is false or useless. More or maximum reasoning effort is not required.
AA Intelligence Index v4.1.1 (score): 59 native-points
Not used in the score. Source category not recorded. Review status: excluded.
Held-out composite, never a scoring input
GPQA Diamond: 95.252525%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash (high); Artificial Analysis published evaluation harness
Metric: accuracy, AA
Original source · Retrieved 2026-09-06
Best published result among 3 reasoning configurations of the exact AA model release; fallback configurations excluded.
AA-LCR: 82%
Not used in the score. Source category: independent. Review status: unresolved.
Configuration: Gemini 3.8 Flash (high)
Original source · Retrieved 2026-09-06
Current AA result is LCR v1.1. The inherited AA-LCR row has no reviewed version match; a new version cannot silently substantiate the older value.
Finance Agent v2 — Vals: 61.44%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: fabv2 overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
Excel Modeling Benchmark — Vals: 72.2%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: emb overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
SciCode: 56.597222%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash (high); Artificial Analysis published evaluation harness
Metric: accuracy, AA harness
Original source · Retrieved 2026-09-06
Best published result among 3 reasoning configurations of the exact AA model release; fallback configurations excluded.
Humanity's Last Exam (no tools): 47.822057%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash (high); Artificial Analysis published evaluation harness
Metric: accuracy, no tools, AA
Original source · Retrieved 2026-09-06
Best published result among 3 reasoning configurations of the exact AA model release; fallback configurations excluded.
Legal Research Bench — Vals: 38.94%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: legal_research overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
Harvey LAB (Vals): 10%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: hlab overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
AA-Briefcase Overall Elo: 1201.13 native-points
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash (high); Artificial Analysis published evaluation harness
Metric: overall Elo
Original source · Retrieved 2026-09-06
Best published result among 2 reasoning configurations of the exact AA model release; fallback configurations excluded.
MedScribe — Vals: 84.5%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: medscribe overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
SWE-bench Verified — Vals: 80%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: swebench overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
LegalBench — Vals: 86.99%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: legal_bench overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
TaxEval v2 — Vals: 74.45%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: tax_eval_v2 overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
Vals Index: 62.25%
Not used in the score. Source category not recorded. Review status: excluded.
Held-out composite, never a scoring input
Public Benefits Bench — Vals: 65.29%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: public-benefits-bench overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
MedCode — Vals: 48.13%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: medcode overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
AA-Omniscience: 29.55 native-points
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash (high); Artificial Analysis published evaluation harness
Metric: knowledge reliability index -100 to 100
Original source · Retrieved 2026-09-06
Best published result among 3 reasoning configurations of the exact AA model release; fallback configurations excluded.
MMLU Pro — Vals: 90.22%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: mmlu_pro overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
LiveBench: 75.83%
Not used in the score. Source category not recorded. Review status: unresolved.
Not admitted after the September 6 model review: the collected 75.83 for Gemini 3.8 Flash on LiveBench still lacks a matched source entry establishing its checkpoint and benchmark-specific metric, tools and retry setup. This is a citation/comparability gap, not a finding that the result is false or useless. More or maximum reasoning effort is not required.
MortgageTax — Vals: 65.34%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: mortgage_tax overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
SAGE — Vals: 35.06%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: sage overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
Code Migration †: 36.55%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: code-migration overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
LiveCodeBench: 89.48%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: lcb overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
ProgramBench v1 (Raw Pass Rate) — Vals: 71.899%
Contributes to the score. Source category: independent. Review status: reviewed-archive.
Configuration: google/gemini-3.8-flash · reasoning_effort=high
Metric: Raw Pass Rate: mean percent of hidden behavioral tests passed per task. Not Fully Resolved (all tests pass) and not Almost Resolved (≥95% of tests pass).
Original source · Retrieved 2026-09-05
Matched exact metric/value and the archived benchmark-specific configuration review. Unknown native effort stays explicitly unknown.
SkillsBench †: 57.98%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: skillsbench overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
CursorBench v3.2: 69.2%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash High; Cursor harness
Metric: CursorBench 3.2 score percentage
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
SWE-bench Pro: 61.6%
Not used in the score. Source category not recorded. Review status: unresolved.
Not admitted after the September 6 model review: the collected 61.6 for Gemini 3.8 Flash on SWE-bench Pro still lacks a matched source entry establishing its checkpoint and benchmark-specific metric, tools and retry setup. This is a citation/comparability gap, not a finding that the result is false or useless. More or maximum reasoning effort is not required.
CharXiv (with CI / RQ): 86.2%
Not used in the score. Source category not recorded. Review status: unresolved.
Not admitted after the September 6 model review: the collected 86.2 for Gemini 3.8 Flash on CharXiv (with CI / RQ) still lacks a matched source entry establishing its checkpoint and benchmark-specific metric, tools and retry setup. This is a citation/comparability gap, not a finding that the result is false or useless. More or maximum reasoning effort is not required.
MMMU-Pro: 85.606936%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash (high); Artificial Analysis published evaluation harness
Metric: accuracy, AA
Original source · Retrieved 2026-09-06
Best published result among 3 reasoning configurations of the exact AA model release; fallback configurations excluded.
FrontierCode v1.1 Extended: 53.45%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash; medium; chisel
Metric: v1.1 Extended cost-adjusted new_score, 150 tasks
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
LVBench (with Memory): 87.8%
Not used in the score. Source category not recorded. Review status: unresolved.
Not admitted after the September 6 model review: the collected 87.8 for Gemini 3.8 Flash on LVBench (with Memory) still lacks a matched source entry establishing its checkpoint and benchmark-specific metric, tools and retry setup. This is a citation/comparability gap, not a finding that the result is false or useless. More or maximum reasoning effort is not required.
MathArena — ArXivMath Jun 2026: 68.03%
Not used in the score. Source category not recorded. Review status: unresolved.
Not admitted after the September 6 model review: the collected 68.03 for Gemini 3.8 Flash on MathArena — ArXivMath Jun 2026 still lacks a matched source entry establishing its checkpoint and benchmark-specific metric, tools and retry setup. This is a citation/comparability gap, not a finding that the result is false or useless. More or maximum reasoning effort is not required.
MathArena — BrokenArXiv Jun 2026: 20.83%
Not used in the score. Source category not recorded. Review status: unresolved.
Not admitted after the September 6 model review: the collected 20.83 for Gemini 3.8 Flash on MathArena — BrokenArXiv Jun 2026 still lacks a matched source entry establishing its checkpoint and benchmark-specific metric, tools and retry setup. This is a citation/comparability gap, not a finding that the result is false or useless. More or maximum reasoning effort is not required.
FrontierCode v1.1 Main — Anthropic H2H: 43.6%
Not used in the score. Source category not recorded. Review status: unresolved.
Not admitted after the September 6 model review: the collected 43.6 for Gemini 3.8 Flash on FrontierCode v1.1 Main — Anthropic H2H still lacks a matched source entry establishing its checkpoint and benchmark-specific metric, tools and retry setup. This is a citation/comparability gap, not a finding that the result is false or useless. More or maximum reasoning effort is not required.
τ³-Banking: 45.773196%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash (medium); Artificial Analysis published evaluation harness
Metric: strict task success, AA
Original source · Retrieved 2026-09-06
Best published result among 3 reasoning configurations of the exact AA model release; fallback configurations excluded.
CritPt: 18.285714%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash (high); Artificial Analysis published evaluation harness
Metric: accuracy, AA official grading
Original source · Retrieved 2026-09-06
Best published result among 3 reasoning configurations of the exact AA model release; fallback configurations excluded.
Terminal-Bench 4.0: 19.1%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: Gemini 3.8 Flash; high; mini-SWE-agent
Metric: resolution rate, v4.0.0 66 tasks
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
ProofBench v1.1 — Vals: 48%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: proof_bench overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
Vibe Code Bench v1.1 — Vals: 78.65%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: vibe-code overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
MineBench: 1917 native-points
Not used in the score. Source category not recorded. Review status: excluded.
Unresolved Pro/base model identities in source catalog; exclude until matched.
HealthBench Professional (length-adjusted): 52.1%
Contributes to the score. Developer-reported. Review status: reviewed-developer.
Configuration: Gemini 3.8 Flash; publisher-reported maximum across reasoning efforts; OpenAI research/API evaluation
Metric: HealthBench Professional (length-adjusted) as labelled in OpenAI comparison table
Original source · Retrieved 2026-09-06
Developer-reported maximum across reasoning modes. Exact selected mode is not individually disclosed. Research prompts/tools may differ from production; protocol-specific rows remain separate.
BioMysteryBench v1 — Vals: 62.22%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: biomysterybench overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
Tax Agent Bench v1 — Vals: 66.77%
Contributes to the score. Source category: independent. Review status: reviewed-independent.
Configuration: google_gemini-3.8-flash; Vals high effort
Metric: tax_agent_bench overall percentage in Vals benchmark harness
Original source · Retrieved 2026-09-06
Reconciled against the exact model/configuration in the evaluator’s public source. Updated source value recorded explicitly; best published reasoning mode selected within the same benchmark protocol.
RuneBench (30m, mean ln(1 + XP/min)): 5.962732 native-points
Contributes to the score. Source category: independent. Review status: reviewed-archive.
Configuration: gemini38flash
Metric: Arithmetic mean of ln(1 + peakXpRate) over all 16 skills, exactly as the owner’s Heatmap.js default ranking. Data rates are already normalized; do not divide by 200 again.
Original source · Retrieved 2026-09-06
Matched exact metric/value and the archived benchmark-specific configuration review. Unknown native effort stays explicitly unknown.
VoxelBench (text-to-voxel, Glicko-2): 1811 native-points
Contributes to the score. Source category: independent. Review status: reviewed-archive.
Configuration: Gemini 3.8 Flash
Metric: Glicko-2 rating from the text-to-voxel board; not win percentage, deviation, image-input mode or MineBench Elo.
Original source · Retrieved 2026-09-06
Matched exact metric/value and the archived benchmark-specific configuration review. Unknown native effort stays explicitly unknown.
Roboflow Vision Evals (six-task mean, high tier): 86.516667%
Contributes to the score. Source category: independent. Review status: reviewed-archive.
Configuration: Require a complete six-task high-tier panel, the highest published benchmark tier. Retain high-tier scores even when lower than low-tier. Do not substitute the best of three repetitions: use the published mean. Each panel enters the fit once; do not also fit its components.
Metric: Arithmetic mean of six published high-tier task means; derived by The Latent, not Roboflow’s official low-tier overall score.
Original source · Retrieved 2026-09-06
Matched exact metric/value and the archived benchmark-specific configuration review. Unknown native effort stays explicitly unknown.
Code Arena WebDev (overall): 1567 native-points
Contributes to the score. Source category: independent. Review status: reviewed-archive.
Configuration: gemini-3.8-flash-high
Metric: Published overall WebDev preference rating, not task success percentage or Arena Text score.
Original source · Retrieved 2026-09-06
Matched exact metric/value and the archived benchmark-specific configuration review. Unknown native effort stays explicitly unknown.
Cite this model record
The Latent. “Gemini 3.8 Flash: score and benchmark evidence.” 2026-09-06. Release tli-2026-v1.0-2026-09-06-db.
Download release data and definitions (JSON) · Data usage terms
Related reporting
These dated stories provide reporting context; they may describe an earlier release.