Gemini 3 Pro vs Gemini 4 Carbon
Full profiles: Gemini 3 Pro (Google) and Gemini 4 Carbon (Google): every benchmark, price and rank.
Updated · Data from 16 sources
As of 10 Nov 2026, Gemini 4 Carbon is ahead of Gemini 3 Pro on quality (68.5 vs 60.2 on the BenchLeader Index). Gemini 4 Carbon leads in every category we measure both in.
Key facts (as of 10 Nov 2026)
- Gemini 3 Pro scores 60.2 on the BenchLeader Index and Gemini 4 Carbon scores 68.5 (BenchLeader, 10 Nov 2026).
- Gemini 4 Carbon costs $8.00 per 1M tokens blended (list price, 10 Nov 2026).
- Of the 7 benchmarks both models have, Gemini 4 Carbon leads 7 (BenchLeader, 10 Nov 2026).
- On Arena WebDev, Gemini 3 Pro is rated 1441 and Gemini 4 Carbon 1768 (LMArena, as published by 10 Nov 2026).
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 | 75 vs 58 category score; Arena WebDev 1768 vs 1441 | High |
| Agents & tool use | Gemini 4 Carbon | 78 vs 61 category score | High |
| Reasoning | Gemini 4 Carbon | 77 vs 66 category score; Arena Hard Prompts 1546 vs 1502 | High |
| Knowledge | Gemini 4 Carbon | 74 vs 56 category score | High |
| Chat quality (human preference) | Gemini 4 Carbon | 73 vs 68 on Arena votes | Medium |
| High volume on a budget | Neither flagship | A smaller model is usually the better fit: Cheapest good LLMs, Fastest LLMs | – |
Key differences
- Gemini 4 Carbon (high) leads on quality: 68.5 vs 60.2.
- Gemini 3 Pro is stronger in instruction following, maths, multimodal.
- Gemini 4 Carbon (high) is stronger in agents & tools, coding, human preference, knowledge, reasoning, long context.
At a glance
| Spec | Gemini 3 Pro | Gemini 4 Carbon |
|---|---|---|
| Developer | ||
| Released | 18 Nov 2025 | 28 Oct 2026 |
| Context window (tokens) | – | 1,048,576 |
| Compared at | Gemini 3 Pro | Gemini 4 Carbon (high) |
| Open weights | No | No |
| Price in / out / cached input per 1M | – | $4.00 / $20.00 / $0.100 |
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 Pro | Gemini 4 Carbon (high) | Leader |
|---|---|---|---|
| Reasoning | 65.5 | 76.9 | Gemini 4 Carbon |
| Coding | 57.8 | 74.5 | Gemini 4 Carbon |
| Agents & tools | 61.3 | 78.2 | Gemini 4 Carbon |
| Maths | 54.2 | – | not comparable |
| Knowledge | 56.2 | 74.4 | Gemini 4 Carbon |
| Instruction following | 60.1 | – | not comparable |
| Human preference | 67.7 | 72.6 | Gemini 4 Carbon |
| Multimodal | 65.6 | – | not comparable |
| Long context | – | 64.3 | not comparable |
On one chart
Each domain on the index's own scale, where 50 is an average model. Click a domain to see the benchmarks behind it. Pick up to 5 models from the list.
Not drawn, measured for fewer than two of these models: Maths, Instruction following, Long context, Multimodal.
Where sources disagree
Across the 7 benchmarks both have, Gemini 3 Pro leads 0, Gemini 4 Carbon leads 7. Every overall index that rates both puts Gemini 4 Carbon ahead.
| Overall index | Gemini 3 Pro | Gemini 4 Carbon (high) | Leader |
|---|---|---|---|
| BenchLeader Index | 60.2 | 68.5 | Gemini 4 Carbon |
| Arena Text | 1486 | 1526 | Gemini 4 Carbon |
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. Cost, tokens and time per benchmark task are measured: Artificial Analysis’ averages over its whole Intelligence Index run.
| Measure | Gemini 3 Pro | Gemini 4 Carbon (high) |
|---|---|---|
| Blended price per 1M tokens | – | $8.00 |
| Input / output per 1M | – / – | $4.00 / $20.00 |
| Cached input per 1M | – | $0.100 |
| Price per task: chat reply | – | $0.0076 |
| Price per task: summarize a 30-page report | – | $0.060 |
| Price per task: code edit | – | $0.054 |
| Price per task: agentic coding session | – | $0.320 |
| Price per task: structured extraction | – | $0.012 |
| Agentic coding session, input mostly cached | – | $0.144 |
| Cost per benchmark task (AA Index) | – | $7.76 |
| Output tokens per task (AA Index) | – | 123k |
| Of which reasoning | – | 72k |
| Time per task (AA Index) | – | 26 min |
| Output speed | – | 80 tok/s |
| Time to first answer | – | 66.8 s |
| Time to a complete answer | – | 156.1 s |
Each model is compared at its best-ranked reasoning-effort setting. Gemini 3 Pro scores 60.2; Gemini 3 Pro (high) scores 60.1. Gemini 4 Carbon (high) scores 68.5 and takes 67 s to its first answer token; Gemini 4 Carbon (medium) scores 64.1 in 23 s. 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
131 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 Pro | Gemini 4 Carbon (high) |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 60.2 | 68.5 |
| Agents & tools score | BenchLeader | 61.3 | 78.2 |
| Coding score | BenchLeader | 57.8 | 74.5 |
| Human preference score | BenchLeader | 67.7 | 72.6 |
| Instruction following score | BenchLeader | 60.1 | – |
| Knowledge score | BenchLeader | 56.2 | 74.4 |
| Maths score | BenchLeader | 54.2 | – |
| Multimodal score | BenchLeader | 65.6 | – |
| Reasoning score | BenchLeader | 65.5 | 76.9 |
| Long context score | BenchLeader | – | 64.3 |
| Blended price $/M | BenchLeader | – | $8.00 |
| Output speed | BenchLeader | – | 80 tok/s |
| Time to first answer | BenchLeader | – | 66.8 s |
| Context window | BenchLeader | – | 1,048,576 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | 92.6% | – |
| Humanity's Last Exam | Scale AI / CAIS | 37.5% | – |
| SimpleBench | SimpleBench | 76.4% | – |
| Arena Hard Prompts | Arena (LMArena) | 1502 | 1546 |
| GPQA Diamond (AA) | Artificial Analysis | 90.8%† | – |
| Humanity's Last Exam (AA) | Artificial Analysis | 39.7%† | – |
| HLE-Diamond (AA) | Artificial Analysis | – | 69.1% |
| GPQA Diamond (Vals) | Vals AI | 91.7%† | – |
| Kagi LLM Benchmark | Kagi LLM Benchmark | 80.1% | – |
| GPQA Diamond (HELM) | HELM Capabilities | 80.3% | – |
| ARC-AGI-1 | ARC Prize | 75.0% | – |
| ARC-AGI-2 | ARC Prize | 54.0% | – |
| CritPt | Artificial Analysis | 9.1%† | – |
| CritPt v2 hard subset (AA) | Artificial Analysis | – | 69.0% |
| SciPredict | Scale AI SEAL | 25.3% | – |
| CL-bench | Epoch AI Benchmarking Hub | 15.8% | – |
| ForecastBench | ForecastBench | 65.0% | – |
| Coding | |||
| SWE-bench Verified (Epoch) | Epoch AI Benchmarking Hub | 72.9% | – |
| SciCode | SciCode | – | 64.2% |
| WeirdML | WeirdML | 69.9% | – |
| GSO-Bench | GSO-Bench | 18.6% | – |
| Arena Coding | Arena (LMArena) | 1517 | 1567 |
| Arena WebDev | Arena (LMArena) | 1441 | 1768 |
| LiveCodeBench | Vals AI | 86.4% | – |
| SWE-bench (Vals) | Vals AI | 76.4%† | – |
Show all 131 benchmarks
| Metric | Publisher | Gemini 3 Pro | Gemini 4 Carbon (high) |
|---|---|---|---|
| Coding | |||
| SWE-Bench Pro | Scale AI SEAL | 43.3% | – |
| SWE-bench Verified (bash only) | SWE-bench | 74.2% | – |
| SWE-bench Verified (any scaffold) | SWE-bench | 77.4% | – |
| Vibe Code Bench v1.1 | Vals AI | 14.3%† | 89.1% |
| SWE-Bench Pro (private) | Scale AI SEAL | 43.3% | – |
| ALE-Bench | Epoch AI Benchmarking Hub | 1176.8 | – |
| AlgoTune | AlgoTune | 1.8 | – |
| FrontierSWE | Epoch AI Benchmarking Hub | – | 61.2% |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | – | 65.2% |
| Design Arena Website | Design Arena | 1249 | – |
| FrontierCode 2.0 (AA) | Artificial Analysis | – | 40.6% |
| LiveCodeBench (AA) | Artificial Analysis | 91.8%† | – |
| Agents & tools | |||
| Terminal-Bench | Terminal-Bench | 69.4% | – |
| GDPval | OpenAI | 40.3% | – |
| Remote Labor Index | Scale AI / CAIS | 1.3% | – |
| APEX-Agents | Mercor | – | 84.2%† |
| Arena Agent | Arena (LMArena) | – | 16.6 |
| Terminal-Bench Hard | Artificial Analysis | 41.7%† | – |
| τ²-Bench Telecom (AA) | Artificial Analysis | 87.1%† | – |
| MCP Atlas | Scale AI SEAL | 70.3% | – |
| BFCL Overall | Berkeley Function Calling Leaderboard | 72.5% | – |
| τ²-bench | τ²-bench | 82.5%† | – |
| τ³-bench Banking | τ²-bench | 18.0%† | – |
| GDPval-AA v2.1 | Artificial Analysis | – | 1733 |
| Finance Agent v2.1 | Vals AI | – | 67.2% |
| Harvey's Legal Agent Benchmark v1.1 | Vals AI | – | 17.5% |
| Legal Research Bench | Vals AI | – | 55.8% |
| Poker Agent | Vals AI | 1078.9% | – |
| Tax Agent Bench | Vals AI | – | 51.8% |
| Terminal-Bench 2.0 (Vals) | Vals AI | 55.1%† | – |
| Terminal-Bench 4.0 (Vals) | Vals AI | – | 63.1% |
| Arena Search | Arena (LMArena) | 1201 | – |
| BALROG | BALROG | 58.1% | – |
| DeepResearch Bench | DeepResearch Bench | 46.3%† | – |
| METR Time Horizons | METR | 71.0% | – |
| Vending-Bench 2 | Andon Labs | 5478.2 | 16484.2† |
| AutomationBench | Artificial Analysis | – | 77.8% |
| GDP.pdf (AA) | Artificial Analysis | – | 28.4% |
| AA-Briefcase v1.1 | Artificial Analysis | – | 1611 |
| GDPx | GDPx by SemiAnalysis | – | 10.9% |
| Design Arena Agentic | Design Arena | 1073 | – |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | – | 64.8% |
| FACTS Search | – | 94.4%† | |
| Maths | |||
| OTIS Mock AIME | Epoch AI Benchmarking Hub | 91.4% | – |
| ProofBench | Vals AI | 20.0% | – |
| AIME (Vals) | Vals AI | 96.7%† | – |
| MGSM | Vals AI | 93.9%† | – |
| AIME 2026 | MathArena | 91.7% | – |
| HMMT February 2026 | MathArena | 86.4% | – |
| MathArena Apex | MathArena | 23.4% | – |
| Omni-MATH (HELM) | HELM Capabilities | 55.6% | – |
| MATH 500 | Vals AI | 96.4%† | – |
| Arena Maths | Arena (LMArena) | 1476 | – |
| Knowledge | |||
| AA-Omniscience v1.1 | Artificial Analysis | – | 53.5 |
| MMLU-Pro | Vals AI | 90.1% | – |
| LegalBench | Vals AI | 87.0%† | – |
| CorpFin | Vals AI | 63.7%† | – |
| TaxEval | Vals AI | 72.6%† | – |
| MedQA | Vals AI | 96.0%† | – |
| PRBench Finance | Scale AI SEAL | 39.2% | – |
| PRBench Legal | Scale AI SEAL | 40.6% | – |
| MultiNRC | Scale AI SEAL | 59.0% | – |
| MMLU-Pro (HELM) | HELM Capabilities | 90.3% | – |
| CaseLaw v2 | Vals AI | 53.1%† | – |
| Excel Modeling Benchmark v1.1 | Vals AI | – | 76.5% |
| MedCode | Vals AI | 52.2%† | – |
| MedScribe | Vals AI | 72.0%† | – |
| 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 | 92.2%† | – |
| Instruction following | |||
| IFBench | Artificial Analysis | 70.4%† | – |
| MultiChallenge | Scale AI SEAL | 65.7% | – |
| TutorBench | Scale AI SEAL | 53.7% | – |
| IFEval (HELM) | HELM Capabilities | 87.6% | – |
| WildBench (HELM) | HELM Capabilities | 85.9% | – |
| Arena Instruction Following | Arena (LMArena) | 1473 | 1525 |
| Human preference | |||
| Arena Text | Arena (LMArena) | 1486 | 1526 |
| Arena Creative Writing | Arena (LMArena) | 1484 | 1515 |
| Arena Multi-turn | Arena (LMArena) | 1496 | 1564 |
| Arena Longer Queries | Arena (LMArena) | 1491 | 1537 |
| Multimodal | |||
| Arena Vision | Arena (LMArena) | 1305 | – |
| MMMU-Pro | Artificial Analysis | 80.2%† | 89.3%† |
| VISTA | Scale AI SEAL | 51.5% | – |
| MMMU-Pro (official) | MMMU | 81.0% | – |
| MMMU-Pro (Vals) | Vals AI | 87.5%† | – |
| MortgageTax | Vals AI | 69.1%† | – |
| AudioMC | Scale AI SEAL | 54.6%† | – |
| AudioMC (text output) | Scale AI SEAL | 54.6%† | – |
| VTB | Scale AI SEAL | 26.9% | – |
| GeoBench | GeoBench | 84.0% | – |
| VPCT | Epoch AI Benchmarking Hub | 91.0% | – |
| Blueprint-Bench 2 | Epoch AI Benchmarking Hub | – | 58.6%† |
| Long context | |||
| AA-LCR v1.1 | Artificial Analysis | 76.0%† | 79.3% |
| Arena Document | Arena (LMArena) | 1451 | – |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | 152.9 | – |
| AA Intelligence Index v5.0 | Artificial Analysis | – | 60.0 |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 28.0† | – |
| AA Progress Index v1 | Artificial Analysis | 290.7† | 350.8 |
| Vals Index | Vals AI | – | 70.5 |
| HELM Capabilities mean | HELM Capabilities | 79.9% | – |
| Vals Finance Index | Vals AI | – | 67.0% |
| Vals RSI Index | Vals AI | – | 27.7% |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | 41.7% | – |
| MASK | Scale AI SEAL | 42.6% | – |
| PropensityBench | Scale AI SEAL | 52.9% | – |
| Metadata | |||
| AA Openness Index | Artificial Analysis | 5.6† | – |
Data as of 2026-11-10. Best configuration of each model; every score links to its source on the model pages.
What changed
- Gemini 4 Carbon (high): 2 new results (FrontierSWE, Vals RSI Index).
- Gemini 4 Carbon (high): BenchLeader Index 69.3 → 68.5.
- Gemini 4 Carbon (high): BenchLeader Index 68.8 → 69.3.
- Gemini 3 Pro: 1 new result (Arena Search).
- Gemini 4 Carbon (high): 1 score revision.
- Gemini 4 Carbon (high): 2 score revisions.
- Gemini 4 Carbon (high): 1 score revision.
- Gemini 4 Carbon (high): 2 score revisions.
- Gemini 4 Carbon (high): 1 score revision.
- Gemini 3 Pro: 1 score revision.
Related comparisons
- Gemini 3 Pro vs Claude Fable 5.2
- Gemini 3 Pro vs Claude Opus 5.5
- Gemini 3 Pro vs GPT-6 Astra
- Gemini 4 Carbon vs Claude Fable 5.2
- Gemini 4 Carbon vs Claude Opus 5.5
- Gemini 4 Carbon vs GPT-6 Astra
- Gemini 4 Carbon vs GPT-6.1 Sol
- Gemini 3 Pro vs Gemini 4 Argon
Rankings: Best LLM overall · Best LLM for coding · Best LLM for agents · Best cheap LLMs
Gemini 3 Pro vs Gemini 4 Carbon: questions
- Is Gemini 4 Carbon better than Gemini 3 Pro?
- Gemini 4 Carbon (high) leads on quality: 68.5 vs 60.2. The BenchLeader Index combines every independent quality benchmark; Gemini 4 Carbon (high) is ahead overall as of 10 Nov 2026, but check the category scores for your use.
- Is Gemini 3 Pro or Gemini 4 Carbon better for coding?
- Gemini 4 Carbon scores higher in coding (75 vs 58 on the category index, where 50 is average). On Arena WebDev, the top shared coding benchmark, it is 1441 for Gemini 3 Pro against 1768 for Gemini 4 Carbon.
- Is Gemini 3 Pro or Gemini 4 Carbon better for agents and tool use?
- Gemini 4 Carbon scores higher in agentic tasks (78 vs 61 on the category index, where 50 is average).
- Do independent benchmarks agree on Gemini 3 Pro vs Gemini 4 Carbon?
- Across the 7 benchmarks both have, Gemini 3 Pro leads 0, Gemini 4 Carbon leads 7. Every overall index that rates both puts Gemini 4 Carbon ahead.
- When was Gemini 4 Carbon released?
- Gemini 4 Carbon was released on 28 Oct 2026.