Fronset Benchmark

Gemini 3.1 Pro Preview (archived)

Provider
Gemini
Model name
gemini-3.1-pro-preview
Last benchmarked
Qualified on
17 / 51 tasks (at the 90% bar)
Best value on
0 tasks

Good enough on 17/51 tasks at the 90% bar. Best value on 0 tasks. Doesn't qualify on any: Financial Analysis & Trading Decisions.

Cost vs quality across all tasks, as of 15 August 2026

0%25%50%75%100%78910Quality score (7–10)Cost-efficiency vs best value (1.0 = best value)Engagement Triage — quality 8.43, 18.3x the cost of the best-value good-enough optionAuthor Matching — quality 8.37, 15.5x the cost of the best-value good-enough optionX Post Selection — quality 8.55, 10.6x the cost of the best-value good-enough optionTopic Grouping and Client Matching — quality 8.02, 11.0x the cost of the best-value good-enough optionPublication Title Generation — quality 8.51, 11.6x the cost of the best-value good-enough optionAuthor Living-Person Safety Check — quality 8.76, 15.4x the cost of the best-value good-enough optionLanguage Detection — quality 9.92, 46.1x the cost of the best-value good-enough optionResearch Query Generation — quality 8.53, 14.9x the cost of the best-value good-enough optionSubstack Newsletter — quality 8.87, 20.8x the cost of the best-value good-enough optionS-1 TOC Extraction — quality 9.10, 9.8x the cost of the best-value good-enough optionEngagement Reply Draft — quality 7.97, 20.3x the cost of the best-value good-enough optionStructured Output Extraction — quality 9.84, 81.3x the cost of the best-value good-enough optionTranslation — quality 8.48, 27.5x the cost of the best-value good-enough optionPrompt Adaptation — quality 8.68, 26.6x the cost of the best-value good-enough optionClaim-Referenced Analyst Writing — quality 8.29, 8.0x the cost of the best-value good-enough optionActivity Feed Blurb Generation — quality 8.18, 22.9x the cost of the best-value good-enough optionTopic-to-Section Assignment — quality 8.09, 8.7x the cost of the best-value good-enough option

within ~1.3× of the best-value model · 1.3–2× · >2× · ★ this model is the best-value pick on that task. Top-right = best quadrant. Only tasks where this model qualifies at the 90% bar are plotted.

Per-task breakdown, as of 15 August 2026

TaskCategoryQuality (% of best)ConfidenceOverpay
Claim-Referenced Analyst WritingLong-form Content Generation90%RANKED8x
Topic-to-Section AssignmentTopic Organization & Clustering92%MEDIUM8.7x
S-1 TOC ExtractionStructured Data & Fact Extraction97%HIGH9.8x
X Post SelectionRelevance, Classification & Matching97%HIGH11x
Topic Grouping and Client MatchingRelevance, Classification & Matching94%RANKED11x
Publication Title GenerationContent Summarization & Synthesis93%RANKED12x
Research Query GenerationInfrastructure & Utility100%HIGH15x
Author Living-Person Safety CheckRelevance, Classification & Matching94%HIGH15x
Author MatchingRelevance, Classification & Matching98%HIGH15x
Engagement TriageRelevance, Classification & Matching91%RANKED18x
Engagement Reply DraftSocial & Promotional Content94%HIGH20x
Substack NewsletterLong-form Content Generation94%HIGH21x
Activity Feed Blurb GenerationSocial & Promotional Content94%RANKED23x
Prompt AdaptationInfrastructure & Utility98%RANKED27x
TranslationInfrastructure & Utility90%RANKED28x
Language DetectionRelevance, Classification & Matching99%RANKED46x
Structured Output Extraction bestStructured Data & Fact Extraction100%RANKED81x

Overpay — how much more you pay by running this model instead of the best-value model that clears the quality bar on that task (marked ★). "16x" means you overpay 16× — the same output for 16× the best-value good-enough option; ★ means this model is that option (no overpayment). Confidence — how sure we are about the quality score (more judgments + more agreement = higher confidence): RANKED many independent judges scored this model's outputs and their agreement is very high (most confident) — HIGH many judges have scored it and they mostly agree (well-pinned) — MEDIUM enough judges have weighed in to publish, but they disagree more than we'd like (treat with a small grain of salt). LOW-confidence cells are hidden everywhere on the site. See the methodology for the exact thresholds.