Explore alternative candidates for the default model based on performance and cost

[neuromechanist]

We are currently using the OpenAI gpt-oss-120b parameters model. It offers good inference and a very fast throughput. However, more performant models or better performing models tend to converge to a higher quality annotation faster, making both the faster throughput and the cost even less than the gpt-oss-120b.

Since fine-tuning open-source models is not currently a priority, we want to explore how other models perform.

[neuromechanist]

Haiku's performance is impressive for the cost (almost on par with the full-fledged gpt-5.2):

However, performance/$ is another story and all open models crush the competition:

So I think this is the case for probably using a good model, probably Hiku by default, but at the same time working to fine tune an open model for our use case.

[neuromechanist]

Let's compare:

• GPT-5.2 (openai/gpt-5.2),
• GPT 5.1 Codex Mini (openai/gpt-5.1-codex-mini),
• GPT 4o Mini (openai/gpt-4o-mini),
• Gemini 3 Flash (google/gemini-3-flash-preview),
• Claude Haiku 4.5 (anthropic/claude-haiku-4.5),
• Mistral Small 3.2 24B (mistralai/mistral-small-3.2-24b-instruct),
• Nemotron 3 Nano 30B A3B (nvidia/nemotron-3-nano-30b-a3b)
  with specific, but not the same examples from HED-reousrces, and also couple of images and prompts.
  Still we should keep the base openai/gpt-oss-120b as our baseline, and qwen/qwen3-235b-a22b-2507 as our evaluation agent.