Gemini 3.1 Pro vs Gemini 4 Carbon
Full profiles: Gemini 3.1 Pro (Google) and Gemini 4 Carbon (Google): every benchmark, price and rank.
Updated · Data from 11 sources
As of 29 Oct 2026, Gemini 3.1 Pro and Gemini 4 Carbon are neck and neck on quality (60.6 vs 65.1 on the BenchLeader Index, inside the ±7.3 margin), and Gemini 3.1 Pro is 1.8× cheaper at $4.50 vs $8.00 per 1M tokens blended. Choose 3.1 Pro for long context (66 vs 65) and fast first answers (19.9 s vs 103.1 s), where it also streams ~1.5× faster; choose 4 Carbon for knowledge (76 vs 56), coding (72 vs 57), and reasoning (79 vs 66).
Key facts (as of 29 Oct 2026)
- Gemini 3.1 Pro scores 60.6 on the BenchLeader Index and Gemini 4 Carbon scores 65.1 (BenchLeader, 29 Oct 2026).
- Gemini 3.1 Pro costs $4.50 per 1M tokens blended (list price, 29 Oct 2026).
- Gemini 4 Carbon costs $8.00 per 1M tokens blended (list price, 29 Oct 2026).
- Gemini 3.1 Pro streams ~1.5× faster than Gemini 4 Carbon (Artificial Analysis, 7-day median, 29 Oct 2026).
- Of the 10 benchmarks both models have, Gemini 4 Carbon leads 9 (BenchLeader, 29 Oct 2026).
Gemini 3.1 Pro benchmarks Which to pick for…
The pick's numbers come first. Category scores on the BenchLeader scale (50 is the average model, 15 points one standard deviation). High confidence means a lead of half a standard deviation or more; under one point is called either.
| Use case | Pick | Why | Confidence |
|---|---|---|---|
| Coding | Gemini 4 Carbon | 72 vs 57 category score; SciCode 64.2% vs 58.7% | High |
| Reasoning | Gemini 4 Carbon | 79 vs 66 category score; HLE-Diamond (AA) 69.1% vs 31.4% | High |
| Knowledge | Gemini 4 Carbon | 76 vs 56 category score | High |
| Long documents | Gemini 3.1 Pro | 66 vs 65 category score; AA-LCR v1.1 82.0% vs 79.3% | Medium |
| Chat quality (human preference) | – | Only Gemini 3.1 Pro has an Arena rating (58 category score) | Not comparable |
| Lowest cost | Gemini 3.1 Pro | $4.50 vs $8.00 per 1M blended; agentic coding session $0.168 vs $0.320 | – |
| Fastest first answer | Gemini 3.1 Pro | 19.9 s (Gemini 3.1 Pro) vs 103.1 s (Gemini 4 Carbon (high)) to the first answer token | – |
| High volume on a budget | Neither flagship | A smaller model is usually the better fit: Cheapest good LLMs, Fastest LLMs | – |
Key differences
- Gemini 3.1 Pro and Gemini 4 Carbon (high) are neck and neck on quality (60.6 vs 65.1, within the ±7.3 margin).
- Gemini 3.1 Pro is stronger in agents & tools, human preference, instruction following, long context, maths, multimodal.
- Gemini 4 Carbon (high) is stronger in coding, knowledge, reasoning.
- Gemini 3.1 Pro is 1.8× cheaper ($4.50 vs $8.00 per 1M blended).
- Gemini 3.1 Pro streams 1.5× faster (151 vs 100 tokens per second).
At a glance
| Spec | Gemini 3.1 Pro | Gemini 4 Carbon |
|---|---|---|
| Developer | ||
| Released | 19 Feb 2026 | 28 Oct 2026 |
| Context window (tokens) | 1,048,576 | 1,000,000 |
| Compared at | Gemini 3.1 Pro | Gemini 4 Carbon (high) |
| Open weights | No | No |
| Price in / out / cached input per 1M | $2.00 / $12.00 / $0.200 | $4.00 / $20.00 / – |
Category by category
Category scores on the BenchLeader scale: 50 is the average evaluated model, 15 points is one standard deviation. A gap under one point is called level.
| Category | Gemini 3.1 Pro | Gemini 4 Carbon (high) | Leader |
|---|---|---|---|
| Reasoning | 66.3 | 78.9 | Gemini 4 Carbon |
| Coding | 56.8 | 71.7 | Gemini 4 Carbon |
| Agents & tools | 57.1 | – | not comparable |
| Maths | 60.7 | – | not comparable |
| Knowledge | 55.9 | 75.8 | Gemini 4 Carbon |
| Instruction following | 66.9 | – | not comparable |
| Human preference | 58.4 | – | not comparable |
| Multimodal | 64.6 | – | not comparable |
| Long context | 65.9 | 64.5 | Gemini 3.1 Pro |
Where sources disagree
Across the 10 benchmarks both have, Gemini 3.1 Pro leads 1, Gemini 4 Carbon leads 9.
| Overall index | Gemini 3.1 Pro | Gemini 4 Carbon (high) | Leader |
|---|---|---|---|
| BenchLeader Index | 60.6 | 65.1 | neck and neck |
Each publisher runs its own tasks, harnesses and effort settings, so their overall verdicts can differ; BenchLeader reconciles them on one scale. † measured at another effort setting of the same model (the publisher did not test the one compared here).
Price, speed and latency
List prices blended at three input tokens per output token. Per-task prices use the token counts on the methodology page; suite cost is what Artificial Analysis paid to run its whole Intelligence Index.
| Measure | Gemini 3.1 Pro | Gemini 4 Carbon (high) |
|---|---|---|
| Blended price per 1M tokens | $4.50 | $8.00 |
| Input / output per 1M | $2.00 / $12.00 | $4.00 / $20.00 |
| Cached input per 1M | $0.200 | – |
| Price per task: chat reply | $0.0044 | $0.0076 |
| Price per task: summarize a 30-page report | $0.031 | $0.060 |
| Price per task: code edit | $0.030 | $0.054 |
| Price per task: agentic coding session | $0.168 | $0.320 |
| Price per task: structured extraction | $0.0064 | $0.012 |
| Cost to run a full benchmark suite | – | $7.76 |
| Output speed | 151 tok/s | 100 tok/s |
| Time to first answer | 19.9 s | 103.1 s |
| Time to a complete answer | 23.3 s | 108.1 s |
Each model is compared at its best-ranked reasoning-effort setting. Gemini 3.1 Pro scores 60.6 and takes 20 s to its first answer token; Gemini 3.1 Pro (high) scores 60.3 in 20 s. Gemini 4 Carbon (high) scores 65.1 and takes 103 s to its first answer token. Higher effort means more thinking before the answer: better scores, longer waits and more tokens billed. How reasoning effort changes scores and cost.
Every benchmark
158 benchmarks, grouped by category; each name links to the board and its leaders. † measured at another effort setting of the same model, shown for reference and not counted as a win.
| Metric | Publisher | Gemini 3.1 Pro | Gemini 4 Carbon (high) |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 60.6 | 65.1 |
| Agents & tools score | BenchLeader | 57.1 | – |
| Coding score | BenchLeader | 56.8 | 71.7 |
| Human preference score | BenchLeader | 58.4 | – |
| Instruction following score | BenchLeader | 66.9 | – |
| Knowledge score | BenchLeader | 55.9 | 75.8 |
| Long context score | BenchLeader | 65.9 | 64.5 |
| Maths score | BenchLeader | 60.7 | – |
| Multimodal score | BenchLeader | 64.6 | – |
| Reasoning score | BenchLeader | 66.3 | 78.9 |
| Blended price $/M | BenchLeader | $4.50 | $8.00 |
| Output speed | BenchLeader | 151 tok/s | 100 tok/s |
| Time to first answer | BenchLeader | 19.9 s | 103.1 s |
| Context window | BenchLeader | 1,048,576 tokens | 1,000,000 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | 94.1% | – |
| Humanity's Last Exam | Scale AI / CAIS | 46.4% | – |
| SimpleBench | SimpleBench | 79.6% | – |
| Arena Hard Prompts | Arena (LMArena) | 1508 | – |
| LiveBench 1.0 Reasoning | LiveBench | 84.0%† | – |
| GPQA Diamond (AA) | Artificial Analysis | 94.1% | – |
| Humanity's Last Exam (AA) | Artificial Analysis | 47.0% | 60.4% |
| HLE-Diamond (AA) | Artificial Analysis | 31.4% | 69.1% |
| GPQA Diamond (Vals) | Vals AI | 95.5%† | – |
| EnigmaEval | Scale AI SEAL | 36.8%† | – |
| ARC-AGI-1 | ARC Prize | 98.0% | – |
| ARC-AGI-2 | ARC Prize | 77.1% | – |
| ARC-AGI-3 | ARC Prize | 0.4% | – |
| CritPt | Artificial Analysis | 17.7% | 36.0% |
| CritPt v2 hard subset (AA) | Artificial Analysis | – | 63.0% |
| MysteryMechanism | Vals AI | 20.7%† | – |
| Chess Puzzles | Epoch AI Benchmarking Hub | 51.0%† | – |
| Mystery Game Puzzles | Epoch AI Benchmarking Hub | 34.0%† | – |
| CL-bench | Epoch AI Benchmarking Hub | 20.8% | – |
| CL-bench Life | Epoch AI Benchmarking Hub | 16.9% | – |
| LMCA | Epoch AI Benchmarking Hub | 53.8%† | – |
| DTBench | Epoch AI Benchmarking Hub | 95.5%† | – |
| ForecastBench | ForecastBench | 64.3% | – |
| Coding | |||
| SciCode | SciCode | 58.7% | 64.2% |
| WeirdML | WeirdML | 72.1% | – |
Show all 158 benchmarks
| Metric | Publisher | Gemini 3.1 Pro | Gemini 4 Carbon (high) |
|---|---|---|---|
| Coding | |||
| GSO-Bench | GSO-Bench | 22.6% | – |
| Arena Coding | Arena (LMArena) | 1522 | – |
| Arena WebDev | Arena (LMArena) | 1447 | – |
| LiveBench 1.0 Coding | LiveBench | 76.5%† | – |
| LiveCodeBench | Vals AI | 88.5%† | – |
| IOI | Vals AI | 51.8%† | – |
| SWE-bench (Vals) | Vals AI | 78.8%† | – |
| SWE-Bench Pro | Scale AI SEAL | 46.1%† | – |
| Code Migration | Vals AI | 17.3%† | – |
| ProgramBench | Vals AI | 0.0%† | – |
| Vibe Code Bench 1-100 | Vals AI | 6.7%† | – |
| Vibe Code Bench v1.1 | Vals AI | 32.0%† | 89.1% |
| SWE-Bench Pro (private) | Scale AI SEAL | 46.1%† | – |
| SWE Atlas: Codebase QnA | Scale AI SEAL | 13.5% | – |
| SWE Atlas: Refactoring | Scale AI SEAL | 33.8% | – |
| SWE Atlas: Test Writing | Scale AI SEAL | 29.8% | – |
| MirrorCode | Epoch AI Benchmarking Hub | 8.9%† | – |
| DeepSWE v1.1 | Epoch AI Benchmarking Hub | 11.7% | – |
| ALE-Bench | Epoch AI Benchmarking Hub | 1160.6 | – |
| AlgoTune | AlgoTune | 2.0 | – |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | 4.0% | 65.2% |
| Terminal-Bench 2.1 (AA) | Artificial Analysis | 73.8% | – |
| FrontierCode 1.1 (AA) | Artificial Analysis | – | 48.5% |
| FrontierCode 2.0 (AA) | Artificial Analysis | – | 40.6% |
| Agents & tools | |||
| Terminal-Bench | Terminal-Bench | 80.2% | – |
| APEX-Agents | Mercor | 35.1% | – |
| Arena Agent | Arena (LMArena) | -4.4 | – |
| LiveBench Agentic Coding | LiveBench | 44.1%† | – |
| Terminal-Bench Hard | Artificial Analysis | 53.8% | – |
| τ²-Bench Telecom (AA) | Artificial Analysis | 95.6% | – |
| Terminal-Bench 2.1 (Vals) | Vals AI | 70.8%† | – |
| MCP Atlas | Scale AI SEAL | 78.2%† | – |
| HiL-Bench | Scale AI SEAL | 35.3% | – |
| τ³-bench Banking | τ²-bench | 26.0%† | – |
| GDPval-AA v2.1 | Artificial Analysis | 14.6% | – |
| τ³-Banking (AA) | Artificial Analysis | 21.4% | – |
| ITBench SRE (AA) | Artificial Analysis | 30.3% | – |
| Analyst Agent (AA) | Artificial Analysis | 41.3% | – |
| APEX-Agents (AA) | Artificial Analysis | 32.0% | – |
| Finance Agent v2 | Vals AI | 43.0%† | – |
| Finance Agent v2.1 | Vals AI | – | 67.2% |
| Harvey's Legal Agent Benchmark v1.1 | Vals AI | – | 17.5% |
| Harvey's Legal Agent Benchmark v1 | Vals AI | 0.0%† | – |
| Legal Research Bench | Vals AI | 20.7%† | 55.8% |
| SREBench | Vals AI | 0.0%† | – |
| Tax Agent Bench | Vals AI | 9.3%† | 48.9% |
| Terminal-Bench 2.0 (Vals) | Vals AI | 67.4%† | – |
| Terminal-Bench 4.0 (Vals) | Vals AI | 2.5%† | 63.1% |
| Terminal-Bench Science | Vals AI | 1.4%† | – |
| BALROG | BALROG | 57.0% | – |
| DeepResearch Bench | DeepResearch Bench | 47.8%† | – |
| PostTrainBench | Epoch AI Benchmarking Hub | 22.0% | – |
| ExploitBench | Epoch AI Benchmarking Hub | 26.1% | – |
| METR Time Horizons | METR | 77.0% | – |
| GBAEval | Epoch AI Benchmarking Hub | 0.8% | – |
| Vending-Bench 2 | Andon Labs | 911.2 | 16484.2† |
| AutomationBench | Artificial Analysis | 35.4% | 77.8% |
| GDP.pdf (AA) | Artificial Analysis | 17.8% | 28.4% |
| Harvey LAB-AA v1.1 | Harvey | 0.0% | – |
| EnterpriseOps-Gym | Artificial Analysis | 42.2% | – |
| AA-Briefcase v1.1 | Artificial Analysis | 457 | 1609 |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | – | 64.8% |
| BrowseComp | OpenAI | 31.2% | – |
| DeepSearchQA | 60.2% | – | |
| FACTS Search | 83.7% | – | |
| Maths | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | 59.6% | – |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | 26.8% | – |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | 95.6% | – |
| ProofBench | Vals AI | 26.0% | – |
| LiveBench Mathematics | LiveBench | 91.0%† | – |
| AIME (Vals) | Vals AI | 98.1%† | – |
| AIME 2026 | MathArena | 98.3% | – |
| HMMT February 2026 | MathArena | 94.7% | – |
| USAMO 2026 | MathArena | 74.4% | – |
| MathArena Apex | MathArena | 60.9% | – |
| Arena Maths | Arena (LMArena) | 1489 | – |
| Knowledge | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | 73.5%† | – |
| LiveBench 1.0 Data Analysis | LiveBench | 78.5%† | – |
| AA-Omniscience v1.1 | Artificial Analysis | – | 53.5 |
| MMLU-Pro | Vals AI | 91.0%† | – |
| LegalBench | Vals AI | 87.4%† | – |
| CorpFin | Vals AI | 64.5%† | – |
| TaxEval | Vals AI | 72.9%† | – |
| MedQA | Vals AI | 96.4%† | – |
| PRBench Finance | Scale AI SEAL | 41.9% | – |
| PRBench Legal | Scale AI SEAL | 44.0% | – |
| MultiNRC | Scale AI SEAL | 64.7% | – |
| BioMysteryBench | Vals AI | 60.0%† | – |
| CaseLaw v2 | Vals AI | 64.8%† | – |
| Excel Modeling Benchmark | Vals AI | 52.6%† | – |
| Excel Modeling Benchmark v1.1 | Vals AI | – | 76.5% |
| MedCode | Vals AI | 59.1%† | – |
| MedScribe | Vals AI | 76.1%† | – |
| Public Benefits Bench v1.1 | Vals AI | 53.8%† | – |
| DrugDiscoveryBench | Scale AI SEAL | 41.9%† | – |
| EBR-bench | Epoch AI Benchmarking Hub | 14.3% | – |
| GDP.pdf | Epoch AI Benchmarking Hub | 17.0% | – |
| AA-Omniscience v1.1: accuracy | Artificial Analysis | – | 58.3% |
| AA-Omniscience v1.1: non-hallucination | Artificial Analysis | – | 88.6% |
| Global-MMLU-Lite (AA) | Artificial Analysis | 93.2% | – |
| Instruction following | |||
| LiveBench Language | LiveBench | 85.4%† | – |
| LiveBench Instruction Following | LiveBench | 79.1%† | – |
| IFBench | Artificial Analysis | 77.1% | – |
| MultiChallenge | Scale AI SEAL | 71.4% | – |
| TutorBench | Scale AI SEAL | 53.0% | – |
| Arena Instruction Following | Arena (LMArena) | 1480 | – |
| Human preference | |||
| Arena Text | Arena (LMArena) | 1487 | – |
| EQ-Bench 4 | EQ-Bench | 1132 | – |
| Arena Creative Writing | Arena (LMArena) | 1481 | – |
| Arena Multi-turn | Arena (LMArena) | 1496 | – |
| Arena Longer Queries | Arena (LMArena) | 1499 | – |
| Multimodal | |||
| Arena Vision | Arena (LMArena) | 1296 | – |
| MMMU-Pro | Artificial Analysis | 82.4% | – |
| MMMU-Pro (official) | MMMU | 80.5%† | – |
| MMMU-Pro (Vals) | Vals AI | 88.2%† | – |
| MortgageTax | Vals AI | 69.4%† | – |
| SAGE | Vals AI | 83.2%† | – |
| VTB | Scale AI SEAL | 29.0% | – |
| Blueprint-Bench 2 | Epoch AI Benchmarking Hub | 26.5% | 58.6%† |
| Long context | |||
| AA-LCR v1.1 | Artificial Analysis | 82.0% | 79.3% |
| Arena Document | Arena (LMArena) | 1459 | – |
| MLCR | Artificial Analysis | 11.1% | – |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | 155.3 | – |
| LiveBench 1.0 | LiveBench | 77.0%† | – |
| AA Intelligence Index v5.0 | Artificial Analysis | – | 59.4 |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 29.7 | – |
| AA Progress Index v1 | Artificial Analysis | 296.1 | 351.6 |
| Vals Index | Vals AI | – | 70.4 |
| Vals Finance Index | Vals AI | – | 66.3% |
| Vals Science Index | Vals AI | 27.4%† | – |
| Vals Multimodal Index | Vals AI | 56.1%† | – |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | 29.8% | – |
| MASK | Scale AI SEAL | 42.4% | – |
Data as of 2026-10-29. Best configuration of each model; every score links to its source on the model pages.
What changed
- Gemini 3.1 Pro: BenchLeader Index 61.7 → 60.6.
- Gemini 3.1 Pro: 2 new results (AA Progress Index v1, HLE-Diamond (AA)).
- Gemini 3.1 Pro: BenchLeader Index 59.4 → 61.7.
- Gemini 3.1 Pro: Gemini 3.1 Pro cost per task moved from $1.30/M to $1.44/M (11% up).
- Gemini 3.1 Pro: 2 score revisions.
- Gemini 3.1 Pro: 2 score revisions.
- Gemini 3.1 Pro: 1 score revision.
- Gemini 3.1 Pro: BenchLeader Index 62.3 → 59.4.
- Gemini 3.1 Pro: 8 new results (LiveBench 2.0, LiveBench Long Agentic, LiveBench 2.0 Reasoning…).
- Gemini 3.1 Pro: 1 new result (USAMO 2026).
Related comparisons
- Gemini 3.1 Pro vs Claude Opus 5.5
- Gemini 3.1 Pro vs GPT-6 Astra
- Gemini 3.1 Pro vs GPT-6.1 Sol
- Gemini 4 Carbon vs Claude Opus 5.5
- Gemini 4 Carbon vs GPT-6 Astra
- Gemini 4 Carbon vs GPT-6.1 Sol
- Gemini 4 Carbon vs Claude Fable 5.1
- Gemini 3 Flash vs Gemini 4 Carbon
Rankings: Best LLM overall · Best LLM for coding · Best LLM for agents · Best cheap LLMs
Gemini 3.1 Pro vs Gemini 4 Carbon: questions
- Is Gemini 3.1 Pro better than Gemini 4 Carbon?
- Gemini 3.1 Pro and Gemini 4 Carbon (high) are neck and neck on quality (60.6 vs 65.1, within the ±7.3 margin). The BenchLeader Index combines every independent quality benchmark; the two are within the margin of error as of 29 Oct 2026, so check the category scores for your use.
- Gemini 3.1 Pro vs Gemini 4 Carbon: which is cheaper, and will that change?
- Gemini 3.1 Pro is cheaper today: $4.50 against $8.00 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Is Gemini 3.1 Pro or Gemini 4 Carbon better for coding?
- Gemini 4 Carbon scores higher in coding (72 vs 57 on the category index, where 50 is average). On SciCode, the top shared coding benchmark, it is 58.7% for Gemini 3.1 Pro against 64.2% for Gemini 4 Carbon.
- Which is faster, Gemini 3.1 Pro or Gemini 4 Carbon?
- Gemini 3.1 Pro streams ~1.5× faster (151 vs 100 output tokens per second, Artificial Analysis 7-day median).
- Which is better for long documents?
- Gemini 3.1 Pro scores higher on long-context benchmarks (66 vs 65). Both take about 1M tokens: Gemini 3.1 Pro 1,048,576, Gemini 4 Carbon 1,000,000.
More questions (3)
- Which has lower latency?
- Gemini 3.1 Pro answers first: 19.9 s against 103.1 s to the first answer token, as Artificial Analysis measures it. Reasoning effort drives most of this: a higher setting thinks longer before it answers.
- Do independent benchmarks agree on Gemini 3.1 Pro vs Gemini 4 Carbon?
- Across the 10 benchmarks both have, Gemini 3.1 Pro leads 1, Gemini 4 Carbon leads 9.
- When was Gemini 4 Carbon released?
- Gemini 4 Carbon was released on 28 Oct 2026.