Part of Best LLM for Vision & Documents

Best LLM for OCR & Documents

Gemini 3.8 Flash is the provisional first pick for OCR & Documents as of November 15, 2026, while independent task measurements are still limited.[1][2][3][4] Ranks #2 on LMArena Vision's OCR category with Elo 1315, though practitioner tests show mixed real-world accuracy on license plates and structured documents. Claude Opus 5.5 scores within two points, so read the top two as close rather than as a measured gap. Watch out: Watch for hallucinated readings in license plate OCR, where it invented 3 false plates out of 10 in a WhatsApp bot test, achieving only 3/10 correct. Claude Opus 5.5 is the next-ranked alternative. It ranks #4 of 79 on LMArena's vision arena (Elo 1295), based on human preference on image-understanding tasks.

About this recommendation

Updated
Evidence through
Sources
15
Revision
v130

Decision audit

Why this result

Inspect the inputs and the computed order behind the recommendation.

Models screened

24

live candidates

Evaluation feeds

5

task-weighted

Winner coverage

45%

intended feed weight

Largest provider share

3 of 8

Google

Provisional source breadth. 4 citation families and 3 practitioner families provide attributed capability reports; 3 cautionary threads retained. The largest citation family contributes 83%.

Google: Gemini 3.8 Flash leads the measured task comparison with limited independent task measurements. Community volume, release age and usage popularity add no ranking points. The leading scores are close under this policy.

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 3.8 Flash
Evaluation feedWeightWinner resultField measured
LMArena Vision OCRObserved 2026-11-09 · evaluation date not supplied · high effort
67%
#224/24
LMArena DocumentObserved 2026-11-09 · evaluation date not supplied · high effortstale
0%
no current result0/24
VLMEvalKit tasksdisabled; last successful check unknownunavailable
0%
feed unavailable0/24
LMArena VisionObserved 2026-11-09 · evaluation date not supplied · high effort
33%
#724/24
Structured-output evaldisabled; last successful check unknownunavailable
0%
feed unavailable0/24

Provider concentration

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

Google38%
  • Google3 models
  • Meta2 models
  • OpenAI2 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 45% · provisional evidence

    #2 LMArena Vision OCR · #7 LMArena Vision

  2. Anthropic · coverage 45% · close score · provisional evidence

    #4 LMArena Vision · #4 LMArena Vision OCR

  3. Google · coverage 45% · provisional evidence

    #5 LMArena Vision · #7 LMArena Vision OCR

  4. Meta · coverage 45% · provisional evidence

    #6 LMArena Vision OCR · #9 LMArena Vision

  5. Meta · coverage 45% · provisional evidence

    #5 LMArena Vision · #8 LMArena Vision OCR

  6. OpenAI · coverage 45% · provisional evidence

    #7 LMArena Vision · #10 LMArena Vision OCR

  7. OpenAI · coverage 45% · provisional evidence

    #14 LMArena Vision · #16 LMArena Vision OCR

  8. Google · coverage 45% · provisional evidence

    #13 LMArena Vision OCR · #18 LMArena Vision

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

    Ranks #2 on LMArena Vision's OCR category with Elo 1315, though practitioner tests show mixed real-world accuracy on license plates and structured documents.

    Best when: Use for general image text extraction where LMArena's blind preference rankings matter, as it places second among 74 models in the OCR category.
    Source 1
    “Within the 74 catalog models matched to this evaluation: Ranks #2 of 74 on LMArena Vision's OCR category (Elo 1315), based on blind human preference for reading text out of images.”
    LMArena Vision OCROpen original ↗

    More tips

    • Consider for document tasks where it ranks #7 of 36 on LMArena's document arena with score 1482.
      Source 5
      “Within the 36 catalog models matched to this evaluation: Ranks #7 of 36 on LMArena's document arena (score 1482), based on blind preference for document tasks.”
      LMArena document arenaOpen original ↗

    Watch out for

    • Watch for hallucinated readings in license plate OCR, where it invented 3 false plates out of 10 in a WhatsApp bot test, achieving only 3/10 correct.
      Source 2
      “Problema observado El bot de WhatsApp (bacollab-whatsapp) lee las patentes mandando la foto entera a un VLM. El 1 de noviembre medí eso con 9 fotos reales del grupo (10 chapas, autos y motos, con las lecturas correctas verificadas a mano): | modelo | correctas | patentes inventadas | |---|---|---| | x-ai/grok-4.3 | 5/10 | 8 | | qwen/qwen3.5-plus | 4/10 | 6 | | google/gemini-3.8-flash | 3/10 | 3 | | z-ai/glm-5.3-flash | 3/10 | 6 | | deepseek/deepseek-v4.1-flash (el que usa el bot hoy) | 2/10…”
      bradlymcgregorhaleOpen original ↗
    • Expect failures on German and French OCR router rules, with wrong text exceeding the 2% threshold in Embassy of the Free Mind's tests.
      Source 6
      “Step 1 result: a fresh 3.8 Flash-low read fails the router rule on German and French. Latin is not measured. No lane built. PR #7520, record scripts/eval/experiments/2026-10-27-flash-low-fresh-vs-lite-7517.md. Same pages as router tests 2026-10-26 (one per book), Opus route-judge against the image. Cut reads count as not accepted. Rule: P(acc ≥ best − 0.05) ≥ 0.9, wrong text ≤ best + 2%. | cell | route | k/n | wrong text | verdict | |---|---|---|---|---| | German | lite | 21/30 | 5/30 | f…”
    • Audit outputs for duplicate and contradictory entries when processing product catalogs, as seen in a 5-page PDF that produced 92 entries with 18 duplicates and conflicting brand/specification data.
      Source 3
      “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…”
  2. 02

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

    Claude Opus 5.5 ranks #4 of 79 on LMArena's vision arena (Elo 1295), based on human preference on image-understanding tasks.

    Best when: Its strongest result for this task is on LMArena Vision, which measures reading images, screenshots, and documents.
    Source 4
    “Within the 79 catalog models matched to this evaluation: Ranks #4 of 79 on LMArena's vision arena (Elo 1295), based on human preference on image-understanding tasks.”
    LMArena vision arenaOpen original ↗

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

  3. 03

    Gemini 3.7 FlashGoogle · $0.75 in / $3.75 out per M tokens · 1.0M context

    Ranks #5 on LMArena's vision arena and #7 on OCR category, with solid document arena placement at #10, showing balanced performance across evaluation types.

    Best when: Use for document processing workflows where it scores 1476 on LMArena's document arena, ranking #10 of 36.
    Source 7
    “Within the 36 catalog models matched to this evaluation: Ranks #10 of 36 on LMArena's document arena (score 1476), based on blind preference for document tasks.”
    LMArena document arenaOpen original ↗

    More tips

    • Consider for image text extraction where it places #7 of 74 on LMArena Vision's OCR category with Elo 1308.
      Source 8
      “Within the 74 catalog models matched to this evaluation: Ranks #7 of 74 on LMArena Vision's OCR category (Elo 1308), based on blind human preference for reading text out of images.”
      LMArena Vision OCROpen original ↗

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

  4. 04

    Muse Spark 1.3Meta · $1.25 in / $4.25 out per M tokens · 1.0M context

    Muse Spark 1.3 ranks #6 of 74 on LMArena Vision's OCR category (Elo 1309), based on blind human preference for reading text out of images.

    Best when: Its strongest result for this task is on LMArena Vision OCR, which measures reading text out of images and scans.
    Source 9
    “Within the 74 catalog models matched to this evaluation: Ranks #6 of 74 on LMArena Vision's OCR category (Elo 1309), based on blind human preference for reading text out of images.”
    LMArena Vision OCROpen original ↗

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

  5. 05

    Muse Spark 1.2Meta · $1.25 in / $4.25 out per M tokens · 1.0M context

    Ties for #5 on LMArena's vision arena with Elo 1294, though OCR-specific performance lags slightly at #8, suggesting strength in general image understanding over text extraction.

    Best when: Use for vision-heavy workflows where general image understanding matters, as it ranks #5 of 79 on LMArena's vision arena.
    Source 10
    “Within the 79 catalog models matched to this evaluation: Ranks #5 of 79 on LMArena's vision arena (Elo 1294), based on human preference on image-understanding tasks.”
    LMArena vision arenaOpen original ↗

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

  6. 06

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

    Ties for #7 on LMArena's vision arena but drops to #10 on OCR category, indicating weaker specialized text extraction relative to general image capabilities.

    Best when: Use for general vision tasks where it ranks #7 of 79 with Elo 1293 on LMArena's vision arena.
    Source 11
    “Within the 79 catalog models matched to this evaluation: Ranks #7 of 79 on LMArena's vision arena (Elo 1293), based on human preference on image-understanding tasks.”
    LMArena vision arenaOpen original ↗

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

  7. 07

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

    Ranks #15 on document arena and #16 on OCR category, placing in the middle third of evaluated models for document and text extraction tasks.

    Best when: Its strongest result for this task is on LMArena Vision, which measures reading images, screenshots, and documents.
    Source 12
    “Within the 79 catalog models matched to this evaluation: Ranks #14 of 79 on LMArena's vision arena (Elo 1283), based on human preference on image-understanding tasks.”
    LMArena vision arenaOpen original ↗

    Watch out for

    • Consider stronger alternatives for OCR-heavy workflows, as it ranks #16 of 74 on LMArena Vision's OCR category with Elo 1297 and #15 of 36 on document arena with score 1467.
      Source 13
      “Within the 36 catalog models matched to this evaluation: Ranks #15 of 36 on LMArena's document arena (score 1467), based on blind preference for document tasks.”
      LMArena document arenaOpen original ↗
      Source 14
      “Within the 74 catalog models matched to this evaluation: Ranks #16 of 74 on LMArena Vision's OCR category (Elo 1297), based on blind human preference for reading text out of images.”
      LMArena Vision OCROpen original ↗
  8. 08

    Gemini 3.6 FlashGoogle · $0.75 in / $3.75 out per M tokens · 1.0M context

    Gemini 3.6 Flash ranks #13 of 74 on LMArena Vision's OCR category (Elo 1298), based on blind human preference for reading text out of images.

    Best when: Its strongest result for this task is on LMArena Vision OCR, which measures reading text out of images and scans.
    Source 15
    “Within the 74 catalog models matched to this evaluation: Ranks #13 of 74 on LMArena Vision's OCR category (Elo 1298), based on blind human preference for reading text out of images.”
    LMArena Vision OCROpen original ↗

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

Frequently asked

What is the best model for OCR & Documents?
Gemini 3.8 Flash ranks first as of November 15, 2026. Ranks #2 on LMArena Vision's OCR category with Elo 1315, though practitioner tests show mixed real-world accuracy on license plates and structured documents.[1][2][3]
What should I watch out for with Gemini 3.8 Flash?
Watch for hallucinated readings in license plate OCR, where it invented 3 false plates out of 10 in a WhatsApp bot test, achieving only 3/10 correct.[2]
What is the best alternative to Gemini 3.8 Flash?
Claude Opus 5.5 ranks second. Claude Opus 5.5 ranks #4 of 79 on LMArena's vision arena (Elo 1295), based on human preference on image-understanding tasks.[4]

Sources

  1. 1

    “Within the 74 catalog models matched to this evaluation: Ranks #2 of 74 on LMArena Vision's OCR category (Elo 1315), based on blind human preference for reading text out of images.”

    LMArena Vision OCR · Benchmark · Nov 9, 2026
  2. 2

    “Problema observado El bot de WhatsApp (bacollab-whatsapp) lee las patentes mandando la foto entera a un VLM. El 1 de noviembre medí eso con 9 fotos reales del grupo (10 chapas, autos y motos, con las lecturas correctas verificadas a mano): | modelo | correctas | patentes inventadas | |---|---|---| | x-ai/grok-4.3 | 5/10 | 8 | | qwen/qwen3.5-plus | 4/10 | 6 | | google/gemini-3.8-flash | 3/10 | 3 | | z-ai/glm-5.3-flash | 3/10 | 6 | | deepseek/deepseek-v4.1-flash (el que usa el bot hoy) | 2/10…”

    bradlymcgregorhale · GitHub · Nov 1, 2026
  3. 3

    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
  4. 4

    “Within the 79 catalog models matched to this evaluation: Ranks #4 of 79 on LMArena's vision arena (Elo 1295), based on human preference on image-understanding tasks.”

    LMArena vision arena · Benchmark · Nov 9, 2026
  5. 5

    “Within the 36 catalog models matched to this evaluation: Ranks #7 of 36 on LMArena's document arena (score 1482), based on blind preference for document tasks.”

    LMArena document arena · Benchmark · Nov 9, 2026
  6. 6

    “Step 1 result: a fresh 3.8 Flash-low read fails the router rule on German and French. Latin is not measured. No lane built. PR #7520, record scripts/eval/experiments/2026-10-27-flash-low-fresh-vs-lite-7517.md. Same pages as router tests 2026-10-26 (one per book), Opus route-judge against the image. Cut reads count as not accepted. Rule: P(acc ≥ best − 0.05) ≥ 0.9, wrong text ≤ best + 2%. | cell | route | k/n | wrong text | verdict | |---|---|---|---|---| | German | lite | 21/30 | 5/30 | f…”

    JDerekLomas · GitHub · Oct 27, 2026
  7. 7

    “Within the 36 catalog models matched to this evaluation: Ranks #10 of 36 on LMArena's document arena (score 1476), based on blind preference for document tasks.”

    LMArena document arena · Benchmark · Nov 9, 2026
  8. 8

    “Within the 74 catalog models matched to this evaluation: Ranks #7 of 74 on LMArena Vision's OCR category (Elo 1308), based on blind human preference for reading text out of images.”

    LMArena Vision OCR · Benchmark · Nov 9, 2026
  9. 9

    “Within the 74 catalog models matched to this evaluation: Ranks #6 of 74 on LMArena Vision's OCR category (Elo 1309), based on blind human preference for reading text out of images.”

    LMArena Vision OCR · Benchmark · Nov 9, 2026
  10. 10

    “Within the 79 catalog models matched to this evaluation: Ranks #5 of 79 on LMArena's vision arena (Elo 1294), based on human preference on image-understanding tasks.”

    LMArena vision arena · Benchmark · Nov 9, 2026
  11. 11

    “Within the 79 catalog models matched to this evaluation: Ranks #7 of 79 on LMArena's vision arena (Elo 1293), based on human preference on image-understanding tasks.”

    LMArena vision arena · Benchmark · Nov 9, 2026
  12. 12

    “Within the 79 catalog models matched to this evaluation: Ranks #14 of 79 on LMArena's vision arena (Elo 1283), based on human preference on image-understanding tasks.”

    LMArena vision arena · Benchmark · Nov 9, 2026
  13. 13

    “Within the 36 catalog models matched to this evaluation: Ranks #15 of 36 on LMArena's document arena (score 1467), based on blind preference for document tasks.”

    LMArena document arena · Benchmark · Nov 9, 2026
  14. 14

    “Within the 74 catalog models matched to this evaluation: Ranks #16 of 74 on LMArena Vision's OCR category (Elo 1297), based on blind human preference for reading text out of images.”

    LMArena Vision OCR · Benchmark · Nov 9, 2026
  15. 15

    “Within the 74 catalog models matched to this evaluation: Ranks #13 of 74 on LMArena Vision's OCR category (Elo 1298), based on blind human preference for reading text out of images.”

    LMArena Vision OCR · Benchmark · Nov 9, 2026

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