Google · Chat
Gemini 3 Pro Preview
Newer version: Gemini 3.1 Pro Preview. We compared release dates and version names in our catalogue.
Google's most advanced reasoning Gemini model
- llminfo IQ
- 136.4
- Maker's price
- $4.50 blended / 1M
- Max context
- 1M
- Listings
- 17
Evidence and sources
Advantages and trade-offs
Compared with 10 chat models, picked mainly by popularity and IQ
Advantages
Advertised audio input support
Audio input advertised by 2 of the 9 peers that list their features.
Trade-offs
Costs more than most peers
Maker's price: $4.50 blended / 1M, pricier than 7 of 9 priced peers.
Fewer provider choices
No open weights
The weights aren't public, so you can't run it yourself.
Input & output
as stated by the model's providersCapabilities
Supported operations and provider-stated features
Listings
| 1M | $2.00>128k $4.00>200k $4.00cache read $0.20batch $1.00 | $12.00>128k $18.00>200k $18.00batch $6.00 | |
| 1M | $2.00 | $12.00 | |
| —Not published by the provider | $2.00 | $12.00 | |
| —Not published by the provider | $2.00>200k $4.00 | $12.00>200k $18.00 | |
| —Not published by the provider | $2.00 | $12.00 | |
| —Not published by the provider | $4.00 | $18.00 | |
| 1M | $2.00>200k $4.00cache read $0.20 | $12.00>200k $18.00cache write $4.50 | |
| 1M | $2.50cache read $0.25 | $15.00cache write $2.50 | |
| 1M | $1.80 | $10.80 | |
| 1M | $0.57 | $3.43 |
7 listings without a price are hidden.
Benchmarks
Named effort levels run from lowest to highest. Select a result to see its source details.
Agentic
Show 6 more resultsHide 6 results
Human preference
Show 61 more resultsHide 61 results
Safety
| Benchmark | Score |
|---|---|
| Safety: BBQ - BBQ accuracy | |
| AIR-Bench: AIRBench 2024 - #1: Confidentiality | |
| AIR-Bench: AIRBench 2024 - #2: Integrity | |
| AIR-Bench: AIRBench 2024 - #3: Availability |
Show 62 more benchmarksHide 62 benchmarks
Reasoning
Vision
| Benchmark | w. sub @ 1fps | wo sub @ 1fps |
|---|---|---|
| Video-MME-v2 | ||
| Avg Acc | ||
| Capability Consistency | ||
| Level 1 |
Show 3 more benchmarksHide 3 benchmarks
| Benchmark | w. sub @ 1fps | wo sub @ 1fps |
|---|---|---|
| Level 2 | ||
| Level 3 | ||
| Reasoning Coherence |
| Benchmark | non-thinking low | thinking high |
|---|---|---|
| VITA-Bench | ||
| Delivery | ||
| In-store | ||
| OTA |
Math
Coding
Show 2 more resultsHide 2 results
OCR
Show 3 more resultsHide 3 results
Science
Factuality
Multilingual
General
Instruction following
Knowledge
Long context
Speech
Other
| Benchmark | Score |
|---|---|
| arak (bleu, 0-shot) | |
| conversation (bleu, 0-shot) | |
| dream (exact_match_mcqa, 0-shot) | |
| email (bleu, 0-shot) |
Show 18 more benchmarksHide 18 benchmarks
Show 30 more resultsHide 30 results
| Benchmark | google/gemini-3-pro-preview |
|---|---|
| ALT_バイアス | |
| ALT_倫理・道徳 | |
| ALT_制御性 | |
| ALT_堅牢性 |
Show 24 more benchmarksHide 24 benchmarks
Show 6 more resultsHide 6 results
Show 1 more resultHide 1 result
Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links. · Data by LLM Stats · Epoch AI Benchmarking Hub, CC BY 4.0 (licence) · LMArena, CC BY 4.0 (licence) · ArmBench-LLM by Metric-AI. Public numeric primary task results; no separate results licence asserted. · Steel.dev Agent Leaderboards, MIT (licence) · BullshitBench by Peter Gostev, MIT (licence) · EQ-Bench (Samuel J. Paech) — cite paech2023eqbench · HAL (Princeton) — cite the HAL paper · HalluHard: Fan, Delsad, Flammarion and Andriushchenko (EPFL), arXiv:2602.01031 · Stanford HELM (Stanford CRFM); scenario scores from Capabilities, Safety, Long Context, MedHELM, AIR-Bench, MMLU and Finance · Horangi 4, Weights & Biases Korea; public publisher-selected numeric evaluation results · LLM Chess by Maxim Saplin, Apache-2.0 (Copyright [2024] [Maxim Saplin]) (licence) · MathArena (ETH SRI / INSAIT), MIT — cite dekoninck2026matharena (licence) · Clinical Benchmarks (clinicalbenchmarks.ai), CC BY 4.0; fractions converted to percentages; source documents retain their own terms (licence) · MLE-bench (OpenAI), MIT (licence) · Nejumi Leaderboard 4, Weights & Biases Japan; public publisher-selected numeric evaluation results · OmniDocBench (OpenDataLab), Apache-2.0 (licence) · Open Agent Leaderboard results (Exgentic), CDLA-Permissive-2.0; agreement text at https://cdla.dev/permissive-2-0/. Fractions shown as percentages. (licence) · OpenHands Index (All Hands AI), Apache-2.0 (licence) · Phare (Giskard), CC BY 4.0 (licence) · Shaberi v2.1, Shisa.AI, Apache-2.0. GPT-5.1-judge scores only; judges are not comparable. (licence) · SnakeBench (Greg Kamradt / ARC Prize Foundation), MIT (licence) · SWE-fficiency (swefficiency.com) leaderboard materials, CC BY 4.0; Human Expert reference row omitted (licence) · Sierra Research tau2-bench leaderboard submissions (github.com/sierra-research/tau2-bench), MIT (licence) · Terminal-Bench (Laude Institute), Apache-2.0 (licence) · UzLiB, Tahrirchi (Abror Shopulatov). Public LEADERBOARD.md numbers; repository MIT covers software and documentation, no separate results licence stated and none asserted. · Vectara Hallucination Leaderboard, Apache-2.0 (licence)
Around the web
- Discussion
- Google Grapples with Employee Skepticism About New Gemini ModelDiscussion
23 points, 2 comments
news.ycombinator.com
- Show HN: Benchmark: AI doesn't find bugs unless you tell it what's wrongDiscussion
5 points, 2 comments
news.ycombinator.com
- Show HN: I'm a dermatologist and I vibe coded a 3D biophysical skin modelDiscussion
8 points, 0 comments
news.ycombinator.com
- Discussion
Also known as: databricks-gemini-3-pro, databricks/databricks-gemini-3-pro, gemini-3-pro, gemini-3-pro-grounding, gemini-3-pro-preview-thinking, gemini-3.0-pro-preview, google/gemini-3-pro, google/gemini-3-pro-preview, perplexity/google/gemini-3-pro-preview, replicate/google/gemini-3-pro, vertex/gemini-3-pro-preview, vertex_ai/gemini-3-pro-preview
Like this one. Tap or hover any dotted value, label or icon to see where it comes from and how it is measured. open details when tapped or hovered. How this data works