Skip to main content

BenchLM recommendation

Best LLMs for Translation in 2026

Data verified

As of July 13, 2026, the top model in best llms for translation on the BenchLM leaderboard is Claude Fable 5 with a score of 100.

Last verified: July 13, 2026

Translation quality tracks the multilingual category: benchmarks that test comprehension and generation across languages. The frontier models at the top of this table are effectively tied on major-language pairs; the gaps show up in lower-resource languages, idiom, and domain terminology.

Unless noted otherwise, ranking surfaces on this page use BenchLM's provisional leaderboard lane rather than the stricter sourced-only verified leaderboard.

Bottom line: Claude Fable 5, Claude Mythos 5, and Gemini 3.1 Pro are tied at the top of the multilingual category — pick by price and ecosystem for major-language translation, and test low-resource pairs yourself.

According to BenchLM.ai, Claude Fable 5 leads this ranking with a score of 100, followed by Claude Mythos 5 (100) and Gemini 3.1 Pro (100). The top three are separated by just a few points — any of them would perform well for this use case.

The best open-weight option is Qwen3.5 397B (Reasoning) (ranked #18 with a score of 81.2). Proprietary models hold a clear advantage in this category, though open-weight options may suffice for less demanding use cases.

This ranking is based on provisional weighted averages across the scoring benchmarks in multilingual tracked by BenchLM.ai. For detailed model profiles, click any model name below. To compare two specific models head-to-head, use the "vs #" links.

What changed

Claude Fable 5 tied at the top of the multilingual category.

Gemini 3.1 Pro matches the leaders at a fraction of the price — the value translation pick.

Claude Mythos 5 tied for the top multilingual score.

How to choose

Full Rankings (99 models)

1
Claude Fable 5
Anthropic·Proprietary·1M+

100

prov. avg

2
Claude Mythos 5
Anthropic·Proprietary·1M+

100

prov. avg

3
Gemini 3.1 Pro
Google·Proprietary·1M

100

prov. avg

4
GPT-5.4
OpenAI·Proprietary·1.05M

100

prov. avg

5
Claude Opus 4.6
Anthropic·Proprietary·1M

100

prov. avg

6
GPT-5.3 Codex
OpenAI·Proprietary·400K

95.2

prov. avg

7
GPT-5.2
OpenAI·Proprietary·400K

95.2

prov. avg

Grok 4.1
xAI·Proprietary·1M

95.2

tracked

9
Claude Sonnet 4.6
Anthropic·Proprietary·200K

89.6

prov. avg

10
Qwen3.7 Max
Alibaba·Proprietary·1M

84

prov. avg

11
GPT-5.1
OpenAI·Proprietary·200K

84

prov. avg

12
GPT-5.2-Codex
OpenAI·Proprietary·400K

84

prov. avg

13
GPT-5.1-Codex-Max
OpenAI·Proprietary·400K

84

prov. avg

14
Claude Sonnet 4.5
Anthropic·Proprietary·200K

84

prov. avg

GPT-5 (medium)
OpenAI·Proprietary·128K

84

tracked

16
Kimi K2.5 (Reasoning)
Moonshot AI·Proprietary·128K

81.2

prov. avg

o1-preview
OpenAI·Proprietary·200K

81.2

tracked

Qwen3.5 397B (Reasoning)
Alibaba·Open Weight·128K

81.2

tracked

19
Claude Opus 4.5
Anthropic·Proprietary·200K

80.4

prov. avg

20
Qwen3.7 Plus
Alibaba·Proprietary·1M

79.6

prov. avg

21
Gemini 3 Pro Deep Think
Google·Proprietary·2M

78.4

prov. avg

22
Gemini 3 Pro
Google·Proprietary·2M

78.4

prov. avg

23
GLM-5 (Reasoning)
Z.AI·Open Weight·200K

78.4

prov. avg

24
GPT-5 (high)
OpenAI·Proprietary·128K

78.4

prov. avg

25
Qwen3.6 Plus
Alibaba·Proprietary·1M

77.6

prov. avg

26
Qwen3.5 397B
Alibaba·Open Weight·128K

77.6

prov. avg

27
GLM-5
Z.AI·Open Weight·200K

73.1

prov. avg

28
Nemotron 3 Ultra
NVIDIA·Open Weight·1M

72.8

prov. avg

Grok 4.1 Fast
xAI·Proprietary·1M

72.8

tracked

30
Kimi K2.5
Moonshot AI·Open Weight·256K

70.9

prov. avg

31
Qwen3.5-122B-A10B
Alibaba·Open Weight·262K

70.6

prov. avg

32
Qwen3.5-27B
Alibaba·Open Weight·262K

70.6

prov. avg

33
Gemini 2.5 Pro
Google·Proprietary·1M

70

prov. avg

34
DeepSeek V3.2
DeepSeek·Open Weight·128K

67.2

prov. avg

35
Qwen3.5-35B-A3B
Alibaba·Open Weight·262K

67.2

prov. avg

36
Claude 4 Sonnet
Anthropic·Proprietary·200K

67.2

prov. avg

37
Qwen3 235B 2507 (Reasoning)
Alibaba·Open Weight·128K

67.2

prov. avg

38
o4-mini (high)
OpenAI·Proprietary·200K

67.2

prov. avg

39
o3
OpenAI·Proprietary·200K

64.4

prov. avg

40
Claude 4.1 Opus
Anthropic·Proprietary·200K

64.4

prov. avg

o3-pro
OpenAI·Proprietary·200K

64.4

tracked

Qwen2.5-1M
Alibaba·Open Weight·1M

64.4

tracked

DeepSeekMath V2
DeepSeek·Open Weight·128K

64.4

tracked

44
Qwen3 235B 2507
Alibaba·Open Weight·128K

62.7

prov. avg

45
DeepSeek V3.2 (Thinking)
DeepSeek·Open Weight·128K

61.6

prov. avg

46
Grok 4
xAI·Proprietary·128K

61.6

prov. avg

47
Claude Haiku 4.5
Anthropic·Proprietary·200K

61.6

prov. avg

Qwen2.5-72B
Alibaba·Open Weight·128K

61.6

tracked

49
GLM-4.7
Z.AI·Open Weight·200K

58.8

prov. avg

50
Gemini 3 Flash
Google·Proprietary·1M

58.8

prov. avg

51
Claude 3.5 Sonnet
Anthropic·Proprietary·200K

58.8

prov. avg

DeepSeek Coder 2.0
DeepSeek·Open Weight·128K

58.8

tracked

Llama 3.1 405B
Meta·Open Weight·128K

58.8

tracked

Mistral Large 2
Mistral·Proprietary·128K

58.8

tracked

55
MiMo-V2-Flash
Xiaomi·Open Weight·256K

56

prov. avg

56
o1
OpenAI·Proprietary·200K

56

prov. avg

57
Nemotron 3 Super 100B
NVIDIA·Open Weight·1M

56

prov. avg

DeepSeek LLM 2.0
DeepSeek·Open Weight·128K

56

tracked

Mistral Large 3
Mistral·Proprietary·128K

56

tracked

60
o3-mini
OpenAI·Proprietary·200K

44.8

prov. avg

Claude 4.1 Opus Thinking
Anthropic·Proprietary·200K

44.8

tracked

Grok Code Fast 1
xAI·Proprietary·256K

44.8

tracked

63
GPT-4.1 mini
OpenAI·Proprietary·1M

42

prov. avg

64
GPT-4o
OpenAI·Proprietary·128K

42

prov. avg

Z-1
Z·Proprietary·128K

42

tracked

Mistral 8x7B
Mistral·Open Weight·32K

39.2

tracked

Nemotron-4 15B
NVIDIA·Open Weight·32K

39.2

tracked

68
GPT-OSS 120B
OpenAI·Open Weight·128K

36.4

prov. avg

Nemotron 3 Nano 30B
NVIDIA·Open Weight·32K

36.4

tracked

Claude 3 Haiku
Anthropic·Proprietary·200K

36.4

tracked

71
GPT-4.1
OpenAI·Proprietary·1M

33.6

prov. avg

72
Gemini 2.5 Flash
Google·Proprietary·1M

33.6

prov. avg

73
Gemini 3.1 Flash-Lite
Google·Proprietary·1M

30.8

prov. avg

GPT-4o mini
OpenAI·Proprietary·128K

30.8

tracked

Claude 3 Opus
Anthropic·Proprietary·200K

30.8

tracked

Nemotron Ultra 253B
NVIDIA·Open Weight·32K

30.8

tracked

Moonshot v1
Moonshot AI·Proprietary·128K

30.8

tracked

Gemini 1.5 Pro
Google·Proprietary·2M

25.2

tracked

Llama 3 70B
Meta·Open Weight·128K

22.4

tracked

GPT-4 Turbo
OpenAI·Proprietary·128K

22.4

tracked

Gemini 1.0 Pro
Google·Proprietary·32K

19.6

tracked

DeepSeek V3.1 (Reasoning)
DeepSeek·Open Weight·128K

11.2

tracked

Llama 4 Behemoth
Meta·Open Weight·32K

11.2

tracked

DeepSeek-R1
DeepSeek·Open Weight·128K

8.4

tracked

Phi-4
Microsoft·Open Weight·16K

8.4

tracked

Nova Pro
Amazon·Proprietary·128K

8.4

tracked

Mistral 7B v0.3
Mistral·Open Weight·32K

8.4

tracked

88
GPT-4.1 nano
OpenAI·Proprietary·1M

5.6

prov. avg

89
GPT-OSS 20B
OpenAI·Open Weight·128K

5.6

prov. avg

DeepSeek V3.1
DeepSeek·Open Weight·128K

5.6

tracked

91
Grok 3 [Beta]
xAI·Proprietary·128K

2.8

prov. avg

92
Llama 4 Scout
Meta·Open Weight·10M

2.8

prov. avg

93
Llama 4 Maverick
Meta·Open Weight·1M

2.8

prov. avg

o1-pro
OpenAI·Proprietary·200K

1

tracked

GLM-4.5
Z.AI·Proprietary·128K

1

tracked

DBRX Instruct
Databricks·Open Weight·32K

1

tracked

Mixtral 8x22B Instruct v0.1
Mistral·Open Weight·64K

1

tracked

GLM-4.5-Air
Z.AI·Proprietary·128K

1

tracked

Mistral 8x7B v0.2
Mistral·Open Weight·32K

1

tracked

These rankings update weekly

Get notified when models move. One email a week with what changed and why.

Free. No spam. Unsubscribe anytime.

Key Takeaways

The top model is Claude Fable 5 by Anthropic with a provisional score of 100.

The best open-weight model is Qwen3.5 397B (Reasoning) at position #18.

99 models are included in this ranking.

Score in Context

What these scores mean

The multilingual score blends cross-language comprehension and generation benchmarks like MGSM and MMLU-ProX. It is the closest measured proxy for translation strength BenchLM tracks.

Known limitations

Benchmarks over-represent high-resource languages. For low-resource pairs, dialects, or domain terminology (legal, medical), run your own evaluation set — leaderboard gaps do not transfer reliably.

Best LLMs for Translation FAQ

What is the best LLM for translation?

Claude Fable 5, Claude Mythos 5, and Gemini 3.1 Pro are tied at the top of BenchLM's multilingual category. For most translation work Gemini 3.1 Pro is the practical pick — leader-tier quality at $2/$12 per million tokens, roughly a fifth of Fable 5's price.

Are LLMs better than Google Translate?

For context-heavy translation — documents, marketing copy, anything where tone matters — frontier LLMs generally produce more natural output because they use surrounding context and follow style instructions. For quick single sentences, dedicated translation tools remain faster and cheaper.

What is the best open-source model for translation?

Alibaba's Qwen rows are the strongest open-weight multilingual performers BenchLM tracks, which fits their heavily multilingual training focus. See the Qwen rankings and the open-source leaderboard for current scores per model.

How should I evaluate translation quality myself?

Build a 30-50 segment test set from your real content across your target pairs, translate with 2-3 shortlisted models, and have a native speaker rank blind. Benchmark scores shortlist correctly, but domain terminology and tone preferences are yours to verify.

Last updated: July 13, 2026

The AI models change fast. We track them for you.

A weekly brief for engineers and researchers covering new models, ranking shifts, and pricing changes.

Free. No spam. Unsubscribe anytime.