Talk
Choosing an LLM for English–Japanese translation
Ranked by scores for translating English news sentences into Japanese, measured by closeness to reference translations.
- Highest rated with a known price
- GPT-4o (2024-08) ChatGPT $4.4
- On a budget
- Gemma 4 26B A4B Gemini Open weights $0.26
Score and price
The closer to the top left, the cheaper and better rated the model.
OpenAIGoogleAlibabaMetaNVIDIAOther
Best-rated model at each price
First line: model / Score, second line: price estimate
| Model | Developer | Score | Price |
|---|---|---|---|
| GPT-4o (2024-08) | OpenAI | $4.4 (in $2.5, out $10) | |
| Gemma 4 26B A4B | $0.26 (in $0.15, out $0.6) |
All models (highest score first)
Second line: price (input / output) / context / released / supports Tools Image JSON □ Compare
| Pick | Compare | Model | Developer | Score | JA→EN (score) | Japanese (human rating) | Input $ /1M | Output $ /1M | Context | Supports | Weights | Released | Price and specs |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Gemma 4 31B | 26.0 | — | — | — | 262.1K | 262.1K 2026-04 | |||||||
| GPT-4o (2024-08) | OpenAI | 26.5 | 1292 | $2.5 | $10 | 128K | $2.5 / $10 128K 2024-08 | ||||||
| Gemma 4 26B A4B | 26.1 | — | $0.15 | $0.6 | 262.1K | $0.15 / $0.6 262.1K 2026-04 | |||||||
| GPT-4.1 | OpenAI | 26.0 | 1347 | $2 | $8 | 1.05M | $2 / $8 1.05M 2025-04 | ||||||
| o3 | OpenAI | 21.6 | 1424 | $2 | $8 | 200K | $2 / $8 200K 2025-04 | ||||||
| Llama 3.3 Swallow 70B (instruct, v0.4) | Swallow (Science Tokyo) | 25.1 | — | — | — | — | — | — | |||||
| GPT-5 | OpenAI | 23.6 | 1431 | $1.25 | $10 | 400K | $1.25 / $10 400K 2025-08 | ||||||
| GPT-5.4 | OpenAI | 24.9 | 1480 | $2.5 | $15 | 1.05M | $2.5 / $15 1.05M 2026-03 | ||||||
| GPT-5 mini | OpenAI | 23.9 | 1364 | $0.25 | $2 | 400K | $0.25 / $2 400K 2025-08 | ||||||
| GPT-OSS-Swallow-120B-SFT-v0.1 | Swallow (Science Tokyo) | 21.3 | — | — | — | — | — | — | |||||
| GPT-OSS-Swallow-120B-RL-v0.1 | Swallow (Science Tokyo) | 21.8 | — | — | — | — | — | — | |||||
| Qwen3.5 122B A10B | Alibaba | 24.7 | 1361 | $0.4 | $3.2 | 262.1K | $0.4 / $3.2 262.1K 2026-02 | ||||||
| GPT-OSS 120B | OpenAI | 20.8 | 1325 | — | — | 128K | 128K 2025-08 | ||||||
| Gemma 4 E4B | 22.8 | — | — | — | 131.1K | — | — | 131.1K | |||||
| Qwen3 235B A22B (instruct) | Alibaba | 23.0 | 1392 | — | — | 128K | 128K 2025-07 | ||||||
| Gemma 4 E2B | 22.7 | — | — | — | 131.1K | — | — | 131.1K | |||||
| Qwen3 235B A22B (thinking) | Alibaba | 22.9 | 1376 | — | — | 262.1K | 262.1K 2025-07 | ||||||
| Qwen3.5 27B | Alibaba | 24.1 | 1339 | $0.3 | $2.4 | 262.1K | $0.3 / $2.4 262.1K 2026-02 | ||||||
| Gemma 3 27B | 23.8 | 1308 | — | — | 128K | 128K 2025-03 | |||||||
| Qwen3.5 35B A3B | Alibaba | 23.9 | 1315 | $0.25 | $2 | 262.1K | $0.25 / $2 262.1K 2026-02 | ||||||
| Llama 3.1 Swallow 8B (instruct, v0.5) | Swallow (Science Tokyo) | 22.2 | — | — | — | — | — | — | |||||
| GPT-OSS-Swallow-20B-RL-v0.1 | Swallow (Science Tokyo) | 21.0 | — | — | — | — | — | — | |||||
| Qwen3-Swallow-32B-RL-v0.2 | Swallow (Science Tokyo) | 21.5 | — | — | — | — | — | — | |||||
| Qwen3-Swallow-32B-SFT-v0.2 | Swallow (Science Tokyo) | 21.0 | — | — | — | — | — | — | |||||
| GPT-OSS-Swallow-20B-SFT-v0.1 | Swallow (Science Tokyo) | 21.6 | — | — | — | — | — | — | |||||
| ELYZA-Shortcut-1.0-Qwen-32B | ELYZA | 22.8 | — | — | — | — | — | — | |||||
| Llama 3.3 70B (instruct) | Meta | 24.6 | 1214 | — | — | 131.1K | 131.1K 2024-12 | ||||||
| NVIDIA Nemotron 3 Super 120B A12B | NVIDIA | 22.2 | — | — | — | 262.1K | 262.1K 2026-03 | ||||||
| Qwen3-Swallow-32B-CPT-v0.2 | Swallow (Science Tokyo) | 22.8 | — | — | — | — | — | — | |||||
| o3-mini | OpenAI | 22.1 | 1315 | $1.1 | $4.4 | 200K | $1.1 / $4.4 200K 2025-01 | ||||||
| Qwen3 Next 80B A3B | Alibaba | 19.5 | 1293 | $0.5 | $6 | 131.1K | $0.5 / $6 131.1K 2025-09 | ||||||
| Qwen3 Next 80B A3B (instruct) | Alibaba | 22.8 | 1370 | $0.5 | $2 | 129K | $0.5 / $2 129K 2025-09 | ||||||
| Qwen3-Swallow-30B-A3B-CPT-v0.2 | Swallow (Science Tokyo) | 22.2 | — | — | — | — | — | — | |||||
| Qwen3-Swallow-30B-A3B-SFT-v0.2 | Swallow (Science Tokyo) | 20.7 | — | — | — | — | — | — | |||||
| Qwen3-Swallow-30B-A3B-RL-v0.2 | Swallow (Science Tokyo) | 20.9 | — | — | — | — | — | — | |||||
| ABEJA-QwQ32b-Reasoning-Japanese-v1.0 | ABEJA | 21.0 | — | — | — | — | — | — | |||||
| ELYZA-Thinking-1.0-Qwen-32B | ELYZA | 21.6 | — | — | — | — | — | — | |||||
| Llama 4 Scout | Meta | 23.0 | 1259 | — | — | 128K | 128K 2025-04 | ||||||
| LLM-jp-4 32B A3B | LLM-jp | 19.0 | — | — | — | — | — | — | |||||
| GPT-OSS 20B | OpenAI | 20.5 | 1256 | — | — | 131K | 131K 2025-08 |
How to read this page
- Also shown (not used for ranking): scores on Japanese-to-English translation; human ratings on questions in Japanese. Most models measured have published weights. Only the 68 models measured on this evaluation are listed (the newest was released in 2026-04).
- The values are closeness to reference translations on a 100-point scale, and even the top models score around 30. Differences under 1 point are not clear differences. Bars are drawn from 0 to 30.
- The human rating comes from people comparing two answers and picking the better one (higher is better).
- Only some models have human ratings for Japanese, mostly models with published weights.
- Only models rated higher than every cheaper model are listed, best first. For about the same rating, pick the cheapest row in this table. The top row within the price you can pay is the model with the highest score at that price.
- “On a budget” is the highest-rated model at $2 or less per million tokens (or at your price cap, when that is lower).
- Best at each price
- Prices are what developers pay to use a model through its API, separate from app subscriptions. The unit is one million tokens. A token is a small piece of text (about three quarters of an English word); one million tokens is roughly 8 to 10 novels. Reading (input) and writing (output) are priced differently; here they are blended 3 : 1. For example, at $4.4, one everyday question and answer (about 1,000 tokens) costs about 0.4 cents. Models without a known price are left out.
- In the Supports column the icons are, from left: can use tools / can read images / can answer in a fixed format (JSON); faded means not supported or unknown.
- Open-weights models have their contents (the weights) published, so you can also run them on your own computer or server. Models without the open-weights mark are normally used through the developer's own services (apps or API); for some models the weights status is unknown.
- Names such as ChatGPT next to a model are the consumer services its developer offers. Check on the service itself whether that model is available there.
- Vertical is Score, horizontal is price (the horizontal scale gets coarser to the right). Points joined by the step line are the best-rated models at each price. Hover or tap a dot to see the model's name.
Data: Arena, Epoch AI, Swallow LLM team (CC BY 4.0), Models.dev (MIT), updated 2026-10-24, data sources