gpt-oss-120b vs Claude Haiku 4.5: open weights or a fast closed model

For high-volume work where cost and speed matter more than peak intelligence, there are two very different picks. gpt-oss-120b is OpenAI's own open-weight model, released under Apache 2.0. Claude Haiku 4.5 is Anthropic's fast, low-cost closed model.

The short version: gpt-oss-120b if you need the weights, to host it yourself or fine-tune it; Claude Haiku 4.5 if you want something fast and dependable you can call as-is.

Specs

Both are billed per million input and output tokens. Current prices are on each model page: gpt-oss-120b and Claude Haiku 4.5. gpt-oss-120b is much the cheaper of the two.

Both run in Japan on FastMetal: gpt-oss-120b on Sakura Internet's AI Engine, Haiku 4.5 on Amazon Bedrock in Tokyo.

Where each is stronger

gpt-oss-120b: the weights are yours. With only about 5.1B of its 116.8B parameters active per token, it is light enough that, quantized, it is reported to run on a single high-end workstation. Developers had it wired into their stacks within hours of release, and the early reception was broadly positive. It is considered relatively strong at reasoning, code synthesis and competitive-programming style problems. But the real draw is the license: you can host it on your own hardware, modify it and fine-tune it.

Claude Haiku 4.5: efficiency, close to finished. It is closed, but it runs at more than twice the speed of the Sonnet models at a much lower price, and it scores 73% on SWE-bench Verified, close to far larger models. Early users said the speed changed how they worked, and the GitHub Copilot team described it as matching Sonnet 4's quality, faster.

Weak spots

gpt-oss-120b overthinks. Reviews point to a habit of spending too long on exploratory reasoning, burning tokens and sometimes landing on a wrong answer anyway. It is weak at creative writing, its safety tuning trails the top closed models, and some reviewers found it underwhelming next to what they had hoped for.

Claude Haiku 4.5 is closed. No weights, so no self-hosting and no fine-tuning of your own. On the hardest tasks it gives way to Sonnet and Opus.

Which to use where

• gpt-oss-120b: you need to run on your own infrastructure, you want to modify or fine-tune the weights, or an open license is itself a requirement.
• Claude Haiku 4.5: you want the bulk of an everyday workload, say 70 to 90 percent of requests, handled fast and cheaply, with only the hard cases escalated to a larger model. That barbell setup is where Haiku fits best.

Try both with one key

On FastMetal both models are behind the same API key and the same balance; change model and nothing else.

# gpt-oss-120b
curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-oss-120b", "messages": [{"role": "user", "content": "Hello"}]}'

# 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": "Hello"}]}'

The side-by-side comparison page has both models' answers to the same prompts.

FAQ

Q. Is gpt-oss-120b an open model?

Yes. OpenAI publishes the weights under Apache 2.0, which allows commercial use, modification and fine-tuning.

Q. Which is better for coding?

Haiku 4.5 is the steadier coder, at 73% on SWE-bench Verified. gpt-oss is good at code synthesis but tends to spend tokens freely on reasoning, so the choice comes down to your task and what each answer costs.

Q. Which is cheaper?

gpt-oss-120b, by a wide margin on both input and output. Count the extra reasoning tokens it can spend before you compare cost per task. Current prices are on the gpt-oss-120b and Claude Haiku 4.5 pages.