LLM-jp-3.1 vs Claude Haiku 4.5: open Japanese model vs fast closed model

If your product works in Japanese, there are two very different ways to go. LLM-jp-3.1 is an open-weight model built for Japanese by Japan's National Institute of Informatics (NII). Claude Haiku 4.5 is Anthropic's fast, closed, general-purpose model.

The short version: LLM-jp-3.1 for Japanese writing and humanities tasks, Claude Haiku 4.5 for coding, speed and everything else.

Specs and price

Prices are FastMetal's USD list as of October 26, 2026; current figures are on the models page. On a task of 100,000 input tokens and 10,000 output tokens, that's about $0.015 on LLM-jp-3.1 and $0.18 on Haiku 4.5, less than a tenth of the cost.

Both run in Japan on FastMetal: LLM-jp-3.1 on Sakura Internet's AI Engine, Haiku 4.5 on Amazon Bedrock in Tokyo.

Where each is stronger

LLM-jp-3.1 is a Japanese specialist. The flagship llm-jp-3.1-8x13b-instruct4 is reported to beat GPT-4-0613 on Japanese evaluations, with its highest scores in the humanities category. Instruction following improved substantially from one generation to the next. Because the weights are public, you can host it yourself or use it for research, and for teams that care about Japanese nuance or need a model whose provenance is fully open, that matters.

Claude Haiku 4.5 is the all-rounder. It's closed, but it balances speed, cost and coding well: it runs at more than twice the speed of the Sonnet models, scores 73% on SWE-bench Verified, close to much larger models, and reads images as well as text. It works as the default model for a wide range of everyday tasks.

Weak spots

LLM-jp-3.1: coding, math and context. It trails other labs' models on coding and math, which its developers name as an area still to improve. Its context window is short by current standards, and it can't read images. It isn't a frontier model for general work; you pick it for Japanese and humanities tasks.

Claude Haiku 4.5: closed, and not Japanese-first. No weights, so no self-hosting or fine-tuning of your own. It's not specialized for Japanese, so if fine-grained Japanese phrasing decides the quality of your output, compare the two on your own prompts.

Which to use where

• LLM-jp-3.1: Japanese generation and summarization, humanities knowledge tasks, and research or public-sector work where an open, domestically developed model is itself a requirement.
• Claude Haiku 4.5: general work including code, high-volume requests that need to be fast and cheap, and anything with image input.

Try both with one key

Japanese quality is best judged on your own prompts, not a leaderboard. On FastMetal both models are behind the same API key and the same balance; change model and nothing else.

# LLM-jp-3.1 8x13B instruct4
curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "llm-jp-3.1-8x13b-instruct4", "messages": [{"role": "user", "content": "こんにちは"}]}'

# Claude Haiku 4.5: same request, different model
curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "anthropic-claude-haiku-4-5", "messages": [{"role": "user", "content": "こんにちは"}]}'

The prompt is "Hello" in Japanese. The side-by-side comparison page has both models' answers to the same prompts.

FAQ

Which is better at Japanese? For Japanese writing and humanities tasks, LLM-jp-3.1 is the specialist. Haiku 4.5 is also perfectly usable in Japanese, and results vary by task, so compare them on your own prompts.

Which is better for coding? Claude Haiku 4.5, with 73% on SWE-bench Verified. Coding and math are LLM-jp-3.1's relative weak spots.

Is LLM-jp-3.1 an open model? Yes. NII publishes the weights; it's an open-weight model built for Japanese.