Part of Best LLM for RAG & Long Context

Best LLM for Grounded RAG

Gemini 4 Carbon ranks first for Grounded RAG as of November 15, 2026.[1][2][3][4][5][6][7] Leads on LMArena factuality (Elo 1521, #1 of 112) and overall preference (Elo 1526, #1 of 147) while scoring 77.7% on SimpleQA Verified (#2 of 73) and 77.2% on FACTS Grounding. GPT-6 Astra scores within two points, so read the top two as close rather than as a measured gap. GPT-6 Astra is the next-ranked alternative. Ranks #2 of 15 on LMArena search factuality (Elo 1239) and scores 80.1% on FACTS Grounding (#3 of 25), with strong source-faithfulness despite middling overall arena placement.

About this recommendation

Updated
Evidence through
Sources
24
Revision
v134

Decision audit

Why this result

Inspect the inputs and the computed order behind the recommendation.

Models screened

24

live candidates

Evaluation feeds

7

task-weighted

Winner coverage

62%

intended feed weight

Largest provider share

4 of 8

Google

Established source breadth. 4 citation families and 0 practitioner families provide attributed capability reports; 0 cautionary threads retained. The largest citation family contributes 41%.

Google: Gemini 4 Carbon leads the measured task comparison with 3 independent evaluation families. Community volume, release age and usage popularity add no ranking points. The leading scores are close under this policy.

Real-world track record is still developing. The recommendation can lead on independent measurements while practitioner reports accumulate.

Sources evaluated

These are the weights used in scoring. Missing measurements are omitted from each model’s average. A feed with no result for any model released in the last 60 days is stale and does not score until it catches up.

winner: Gemini 4 Carbon
Evaluation feedWeightWinner resultField measured
FACTS GroundingObserved 2026-10-29 · evaluated 2026-10-29 · Gemini 4 Carbon
41%
#617/24
SimpleQA Verified (Epoch AI)Observed 2026-11-02 · evaluated 2026-11-02 · high effort · standard error 1.30
18%
#221/24
LMArena Text FactualityObserved 2026-11-11 · evaluation date not supplied · high effort
18%
#123/24
CRAGdisabled; last successful check unknownunavailable
0%
feed unavailable0/24
LMArena Search
12%
not measured5/24
LMArena Search Factuality
12%
not measured5/24
LongBench v2disabled; last successful check unknownunavailable
0%
feed unavailable0/24

Provider concentration

Each exact model is scored separately; provider identity is not a ranking input.

Google50%
  • Google4 models
  • OpenAI3 models
  • Anthropic1 model

Decision table

Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.

  1. Google · coverage 62%

    #1 LMArena Text Factuality · #2 SimpleQA Verified (Epoch AI)

  2. OpenAI · coverage 81% · close score

    #2 LMArena Search Factuality · #3 FACTS Grounding

  3. OpenAI · coverage 48%

    #1 FACTS Grounding · #21 SimpleQA Verified (Epoch AI)

  4. OpenAI · coverage 81%

    #3 LMArena Search Factuality · #4 FACTS Grounding

  5. Google · coverage 62%

    #3 SimpleQA Verified (Epoch AI) · #4 LMArena Text Factuality

  6. Google · coverage 48%

    #5 FACTS Grounding · #60 LMArena Text Factuality

  7. Anthropic · coverage 81%

    #2 LMArena Text Factuality · #7 SimpleQA Verified (Epoch AI)

  8. Google · coverage 62%

    #10 SimpleQA Verified (Epoch AI) · #11 FACTS Grounding

Task score averages the available, task-weighted measurements. Benchmark positions compare models matched to our catalog. Models measured by at least two independent evaluator families are considered first. Missing tests, practitioner report counts, popularity and release age add no points. Coverage describes evidence breadth, not a probability of correctness. Scores within two points are shown as close under this policy; this is not a statistical confidence interval. When the leading scores are that close, the model measured by more evaluator families ranks first.

  1. 01

    Gemini 4 CarbonGoogle · $2 in / $10 out per M tokens · 1.0M context

    Leads on LMArena factuality (Elo 1521, #1 of 112) and overall preference (Elo 1526, #1 of 147) while scoring 77.7% on SimpleQA Verified (#2 of 73) and 77.2% on FACTS Grounding.

    Best when: Deploy when verified short-answer accuracy matters: it answered 77.7% of SimpleQA Verified correctly, second only in a 73-model field.
    Source 3
    “Within the 73 catalog models matched to this evaluation: Answers 77.7% of SimpleQA Verified correctly in Epoch AI's independent run (#2 of 73), short factual questions with a single verified answer.”
    SimpleQA Verified (Epoch AI)Open original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  2. 02

    GPT-6 AstraOpenAI · $10 in / $50 out per M tokens · 1.1M context

    Ranks #2 of 15 on LMArena search factuality (Elo 1239) and scores 80.1% on FACTS Grounding (#3 of 25), with strong source-faithfulness despite middling overall arena placement.

    Best when: Choose for search-grounded RAG: its search factuality Elo of 1239 is second-best among 15 models evaluated with search augmentation.
    Source 5
    “Within the 15 catalog models matched to this evaluation: Ranks #2 of 15 on LMArena's search factuality view (Elo 1239), measuring the factual accuracy of search-grounded answers.”
    LMArena search factualityOpen original ↗

    More tips

    • Prioritize when FACTS Grounding is your reference: 80.1% source-faithfulness ranks third of 25, ahead of Gemini 4 Carbon.
      Source 6
      “Within the 25 catalog models matched to this evaluation: Scores 80.1% on FACTS Grounding (#3 of 25), measuring factual support against supplied source material.”
      FACTS GroundingOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  3. 03

    GPT-6.1 TerraOpenAI · $1 in / $5 out per M tokens · 1.1M context

    Tops FACTS Grounding at 84.2% (#1 of 25) for source-faithfulness, though it drops to #21 of 73 on SimpleQA Verified at 57.3%.

    Best when: Select when supplied-source fidelity is paramount: its 84.2% on FACTS Grounding is the highest measured against explicit source material.
    Source 9
    “Within the 25 catalog models matched to this evaluation: Scores 84.2% on FACTS Grounding (#1 of 25), measuring factual support against supplied source material.”
    FACTS GroundingOpen original ↗

    Watch out for

    • Avoid for ungrounded factual recall: 57.3% on SimpleQA Verified places it at #21 of 73, indicating weaker closed-book accuracy.
      Source 10
      “Within the 73 catalog models matched to this evaluation: Answers 57.3% of SimpleQA Verified correctly in Epoch AI's independent run (#21 of 73), short factual questions with a single verified answer.”
      SimpleQA Verified (Epoch AI)Open original ↗
  4. 04

    GPT-6.1 SolOpenAI · $2 in / $10 out per M tokens · 1.1M context

    Scores 78% on FACTS Grounding (#4 of 25) and ranks #3 of 15 on LMArena search factuality (Elo 1237), showing reliable source adherence in grounded settings.

    Best when: Use for search-augmented RAG where its #3 search factuality rank (Elo 1237) indicates strong factual accuracy with retrieved context.
    Source 11
    “Within the 15 catalog models matched to this evaluation: Ranks #3 of 15 on LMArena's search factuality view (Elo 1237), measuring the factual accuracy of search-grounded answers.”
    LMArena search factualityOpen original ↗

    More tips

    • Deploy when source material is provided: 78% on FACTS Grounding demonstrates above-average faithfulness to supplied documents.
      Source 12
      “Within the 25 catalog models matched to this evaluation: Scores 78% on FACTS Grounding (#4 of 25), measuring factual support against supplied source material.”
      FACTS GroundingOpen original ↗

    Watch out for

    • SimpleQA Verified performance of 73.9% (#5 of 73) lags behind top closed-book models, suggesting weaker standalone factual recall.
      Source 13
      “Within the 73 catalog models matched to this evaluation: Answers 73.9% of SimpleQA Verified correctly in Epoch AI's independent run (#5 of 73), short factual questions with a single verified answer.”
      SimpleQA Verified (Epoch AI)Open original ↗
  5. 05

    Gemini 4 ArgonGoogle · $2 in / $10 out per M tokens · 1.0M context

    Ranks #2 of 147 on LMArena overall (Elo 1522) and #4 of 112 on factuality (Elo 1497), with 76% SimpleQA Verified (#3 of 73) but weaker 73.4% on FACTS Grounding.

    Best when: Choose for general factual QA where human preference matters: #2 overall and #4 factuality ranks indicate broad reliability.
    Source 14
    “Within the 147 catalog models matched to this evaluation: Ranks #2 of 147 on LMArena's overall text arena (Elo 1522), based on blind human preference votes.”
    LMArena text arenaOpen original ↗
    Source 15
    “Within the 112 catalog models matched to this evaluation: Ranks #4 of 112 on LMArena's factuality text view (Elo 1497), where answers are judged for factual accuracy in blind comparison.”
    LMArena text factualityOpen original ↗

    More tips

    • Use for short verified-answer questions: 76% on SimpleQA Verified places it third of 73 models.
      Source 16
      “Within the 73 catalog models matched to this evaluation: Answers 76% of SimpleQA Verified correctly in Epoch AI's independent run (#3 of 73), short factual questions with a single verified answer.”
      SimpleQA Verified (Epoch AI)Open original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  6. 06

    Gemma 4 26B A4B Google · $0.068 in / $0.22 out per M tokens · 262K context · open-weight

    Gemma 4 26B A4B scores 77.8% on FACTS Grounding (#5 of 25), measuring factual support against supplied source material.

    Best when: Its strongest result for this task is on FACTS Grounding, which measures staying faithful to supplied source documents.
    Source 8
    “Within the 25 catalog models matched to this evaluation: Scores 77.8% on FACTS Grounding (#5 of 25), measuring factual support against supplied source material.”
    FACTS GroundingOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  7. 07

    Claude Opus 5.5Anthropic · $4 in / $20 out per M tokens · 1M context

    Ranks #2 of 112 on LMArena factuality (Elo 1505) with strong human preference scores.

    Best when: Its strongest result for this task is on LMArena factuality, which measures answering accurately when judged for facts.
    Source 17
    “Within the 112 catalog models matched to this evaluation: Ranks #2 of 112 on LMArena's factuality text view (Elo 1505), where answers are judged for factual accuracy in blind comparison.”
    LMArena text factualityOpen original ↗

    Watch out for

    • Avoid for strict source-faithfulness RAG: 62.4% on FACTS Grounding places it #16 of 25, below all Gemini and GPT-6 variants tested.
      Source 18
      “Within the 25 catalog models matched to this evaluation: Scores 62.4% on FACTS Grounding (#16 of 25), measuring factual support against supplied source material.”
      FACTS GroundingOpen original ↗
    • Search-grounded deployments risk poor results: ranks #9 of 15 on both search arena and search factuality, the lowest among major closed-weight options.
      Source 19
      “Within the 15 catalog models matched to this evaluation: Ranks #9 of 15 on LMArena's search arena (score 1213), measuring answers produced with search.”
      LMArena search arenaOpen original ↗
      Source 20
      “Within the 15 catalog models matched to this evaluation: Ranks #9 of 15 on LMArena's search factuality view (Elo 1206), measuring the factual accuracy of search-grounded answers.”
      LMArena search factualityOpen original ↗
  8. 08

    Gemini 3.8 FlashGoogle · $0.83 in / $4.13 out per M tokens · 1.0M context

    Mid-tier benchmark scores with severe reliability issues in production RAG: multi-turn degradation from 89% to 10% accuracy and document extraction failures with duplicate conflicting entries.

    Best when: Consider only for single-turn, low-stakes queries where 69.7% SimpleQA Verified and mid-tier arena ranks suffice.
    Source 21
    “Within the 112 catalog models matched to this evaluation: Ranks #25 of 112 on LMArena's factuality text view (Elo 1470), where answers are judged for factual accuracy in blind comparison.”
    LMArena text factualityOpen original ↗
    Source 22
    “Within the 73 catalog models matched to this evaluation: Answers 69.7% of SimpleQA Verified correctly in Epoch AI's independent run (#10 of 73), short factual questions with a single verified answer.”
    SimpleQA Verified (Epoch AI)Open original ↗

    Watch out for

    • Multi-turn conversations collapse after round one: accuracy dropped from 89% to 36% to 10% across five rounds, with answers diverting to generic responses.
      Source 23
      “現象 同一份 60 題題組跑三輪、五個模型,只有 Google 供應商的答案品質隨輪次崩潰: | 模型 | 第1輪 | 第2輪 | 第3輪 | 第4輪 | 第5輪 | |---|---|---|---|---|---| | gemini-3.8-flash | 89% | 36% | 18% | 10% | 17% | | gpt-5.6-luna | 77% | 82% | 100% | 83% | 100% | | gpt-5.6-terra | 89% | 82% | 100% | 81% | 100% | | qwen3.6:35b-a3b-q8_0 | 81% | 79% | 97% | 81% | 100% | | qwen3.8:27b-q8_0 | 89% | 82% | 95% | 83% | 100% | (數字=金標準事實覆蓋率,三輪合計 180 筆/臂) 第二輪起 Gemini 的回答脫離知識庫,改用通用常識作答,甚至明說不知道品牌: > 問:客服電話是多少? > 答:抱歉,由於我目前**沒有您所在…”
    • Document ingestion produces unreliable chunks: observed 92 product entries from 5 pages with 18 duplicates containing contradictory brand and specification data.
      Source 24
      “現象 評測租戶知識庫「DM」的 DM_家樂福_采集漫旅_商品頁_p11-15.pdf(5 頁)以 Gemini 3.8 Flash 走混合 OCR(切片 2x3 + 整頁)處理後,落庫內容出現: - 16 組 【頁面標題】(其中 10 組為「不詳」),但原始 PDF 只有 5 頁 - 92 筆 商品: 條目,只有 74 筆唯一 — 18 筆重複 - 重複條目的欄位互相矛盾,例如同一商品「鮮剖100%椰子汁」出現兩次: - 一處 品牌:半天水、規格:含椰肉/純天然 520毫升×3、原價:120~129 - 另一處 品牌:鮮剖、規格:520毫升×3、原價:120-129元、備註:含椰肉/純天然 - 前者所在區段有完整頁面標題「單筆購買第12~15頁…」,後者所在區段頁面標題為「不詳」 已排除 chunk overlap 造成的假重複:把 chunk 依 chunk_index 串接後長度等於各 chunk 長度總和(無重疊),上述統計是文件本體的真實內容。 也已排除 worker 跑舊版:VM `poc-rag-v…”

Frequently asked

What is the best model for Grounded RAG?
Gemini 4 Carbon ranks first as of November 15, 2026. Leads on LMArena factuality (Elo 1521, #1 of 112) and overall preference (Elo 1526, #1 of 147) while scoring 77.7% on SimpleQA Verified (#2 of 73) and 77.2% on FACTS Grounding.[1][2][3][4]
What is the best alternative to Gemini 4 Carbon?
GPT-6 Astra ranks second. Ranks #2 of 15 on LMArena search factuality (Elo 1239) and scores 80.1% on FACTS Grounding (#3 of 25), with strong source-faithfulness despite middling overall arena placement.[5][6][7]
Which open-weight model ranks highest for Grounded RAG?
Gemma 4 26B A4B is the highest-ranked open-weight model, at #6. Gemma 4 26B A4B scores 77.8% on FACTS Grounding (#5 of 25), measuring factual support against supplied source material.[8]

Sources

  1. 1

    “Within the 147 catalog models matched to this evaluation: Ranks #1 of 147 on LMArena's overall text arena (Elo 1526), based on blind human preference votes.”

    LMArena text arena · Benchmark · Nov 11, 2026
  2. 2

    “Within the 112 catalog models matched to this evaluation: Ranks #1 of 112 on LMArena's factuality text view (Elo 1521), where answers are judged for factual accuracy in blind comparison.”

    LMArena text factuality · Benchmark · Nov 11, 2026
  3. 3

    “Within the 73 catalog models matched to this evaluation: Answers 77.7% of SimpleQA Verified correctly in Epoch AI's independent run (#2 of 73), short factual questions with a single verified answer.”

    SimpleQA Verified (Epoch AI) · Benchmark · Nov 2, 2026
  4. 4

    “Within the 25 catalog models matched to this evaluation: Scores 77.2% on FACTS Grounding (#6 of 25), measuring factual support against supplied source material.”

    FACTS Grounding · Benchmark · Oct 29, 2026
  5. 5

    “Within the 15 catalog models matched to this evaluation: Ranks #2 of 15 on LMArena's search factuality view (Elo 1239), measuring the factual accuracy of search-grounded answers.”

    LMArena search factuality · Benchmark · Nov 9, 2026
  6. 6

    “Within the 25 catalog models matched to this evaluation: Scores 80.1% on FACTS Grounding (#3 of 25), measuring factual support against supplied source material.”

    FACTS Grounding · Benchmark · Oct 24, 2026
  7. 7

    “Within the 73 catalog models matched to this evaluation: Answers 75.6% of SimpleQA Verified correctly in Epoch AI's independent run (#4 of 73), short factual questions with a single verified answer.”

    SimpleQA Verified (Epoch AI) · Benchmark · Aug 30, 2026
  8. 8

    “Within the 25 catalog models matched to this evaluation: Scores 77.8% on FACTS Grounding (#5 of 25), measuring factual support against supplied source material.”

    FACTS Grounding · Benchmark · Oct 24, 2026
  9. 9

    “Within the 25 catalog models matched to this evaluation: Scores 84.2% on FACTS Grounding (#1 of 25), measuring factual support against supplied source material.”

    FACTS Grounding · Benchmark · Nov 11, 2026
  10. 10

    “Within the 73 catalog models matched to this evaluation: Answers 57.3% of SimpleQA Verified correctly in Epoch AI's independent run (#21 of 73), short factual questions with a single verified answer.”

    SimpleQA Verified (Epoch AI) · Benchmark · Oct 26, 2026
  11. 11

    “Within the 15 catalog models matched to this evaluation: Ranks #3 of 15 on LMArena's search factuality view (Elo 1237), measuring the factual accuracy of search-grounded answers.”

    LMArena search factuality · Benchmark · Nov 9, 2026
  12. 12

    “Within the 25 catalog models matched to this evaluation: Scores 78% on FACTS Grounding (#4 of 25), measuring factual support against supplied source material.”

    FACTS Grounding · Benchmark · Oct 24, 2026
  13. 13

    “Within the 73 catalog models matched to this evaluation: Answers 73.9% of SimpleQA Verified correctly in Epoch AI's independent run (#5 of 73), short factual questions with a single verified answer.”

    SimpleQA Verified (Epoch AI) · Benchmark · Sep 29, 2026
  14. 14

    “Within the 147 catalog models matched to this evaluation: Ranks #2 of 147 on LMArena's overall text arena (Elo 1522), based on blind human preference votes.”

    LMArena text arena · Benchmark · Nov 11, 2026
  15. 15

    “Within the 112 catalog models matched to this evaluation: Ranks #4 of 112 on LMArena's factuality text view (Elo 1497), where answers are judged for factual accuracy in blind comparison.”

    LMArena text factuality · Benchmark · Nov 11, 2026
  16. 16

    “Within the 73 catalog models matched to this evaluation: Answers 76% of SimpleQA Verified correctly in Epoch AI's independent run (#3 of 73), short factual questions with a single verified answer.”

    SimpleQA Verified (Epoch AI) · Benchmark · Oct 13, 2026
  17. 17

    “Within the 112 catalog models matched to this evaluation: Ranks #2 of 112 on LMArena's factuality text view (Elo 1505), where answers are judged for factual accuracy in blind comparison.”

    LMArena text factuality · Benchmark · Nov 11, 2026
  18. 18

    “Within the 25 catalog models matched to this evaluation: Scores 62.4% on FACTS Grounding (#16 of 25), measuring factual support against supplied source material.”

    FACTS Grounding · Benchmark · Oct 24, 2026
  19. 19

    “Within the 15 catalog models matched to this evaluation: Ranks #9 of 15 on LMArena's search arena (score 1213), measuring answers produced with search.”

    LMArena search arena · Benchmark · Nov 9, 2026
  20. 20

    “Within the 15 catalog models matched to this evaluation: Ranks #9 of 15 on LMArena's search factuality view (Elo 1206), measuring the factual accuracy of search-grounded answers.”

    LMArena search factuality · Benchmark · Nov 9, 2026
  21. 21

    “Within the 112 catalog models matched to this evaluation: Ranks #25 of 112 on LMArena's factuality text view (Elo 1470), where answers are judged for factual accuracy in blind comparison.”

    LMArena text factuality · Benchmark · Nov 11, 2026
  22. 22

    “Within the 73 catalog models matched to this evaluation: Answers 69.7% of SimpleQA Verified correctly in Epoch AI's independent run (#10 of 73), short factual questions with a single verified answer.”

    SimpleQA Verified (Epoch AI) · Benchmark · Sep 2, 2026
  23. 23

    fix(agent): 多輪對話送出兩個 SystemMessage,Gemini 只保留一個 → 檢索內容被丟棄,第二輪起答案脫離知識庫

    Original-language excerpt:“現象 同一份 60 題題組跑三輪、五個模型,只有 Google 供應商的答案品質隨輪次崩潰: | 模型 | 第1輪 | 第2輪 | 第3輪 | 第4輪 | 第5輪 | |---|---|---|---|---|---| | gemini-3.8-flash | 89% | 36% | 18% | 10% | 17% | | gpt-5.6-luna | 77% | 82% | 100% | 83% | 100% | | gpt-5.6-terra | 89% | 82% | 100% | 81% | 100% | | qwen3.6:35b-a3b-q8_0 | 81% | 79% | 97% | 81% | 100% | | qwen3.8:27b-q8_0 | 89% | 82% | 95% | 83% | 100% | (數字=金標準事實覆蓋率,三輪合計 180 筆/臂) 第二輪起 Gemini 的回答脫離知識庫,改用通用常識作答,甚至明說不知道品牌: > 問:客服電話是多少? > 答:抱歉,由於我目前**沒有您所在…”

    larry610881 · GitHub · Sep 8, 2026
  24. 24

    fix(knowledge): 混合 OCR 產生重複頁面與矛盾商品條目(5 頁 PDF 出現 16 組頁面標題、92/74 商品重複)

    Original-language excerpt:“現象 評測租戶知識庫「DM」的 DM_家樂福_采集漫旅_商品頁_p11-15.pdf(5 頁)以 Gemini 3.8 Flash 走混合 OCR(切片 2x3 + 整頁)處理後,落庫內容出現: - 16 組 【頁面標題】(其中 10 組為「不詳」),但原始 PDF 只有 5 頁 - 92 筆 商品: 條目,只有 74 筆唯一 — 18 筆重複 - 重複條目的欄位互相矛盾,例如同一商品「鮮剖100%椰子汁」出現兩次: - 一處 品牌:半天水、規格:含椰肉/純天然 520毫升×3、原價:120~129 - 另一處 品牌:鮮剖、規格:520毫升×3、原價:120-129元、備註:含椰肉/純天然 - 前者所在區段有完整頁面標題「單筆購買第12~15頁…」,後者所在區段頁面標題為「不詳」 已排除 chunk overlap 造成的假重複:把 chunk 依 chunk_index 串接後長度等於各 chunk 長度總和(無重疊),上述統計是文件本體的真實內容。 也已排除 worker 跑舊版:VM `poc-rag-v…”

    larry610881 · GitHub · Sep 8, 2026

Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.