Choosing AI Models
Each tool — Activity Summarizer, Extractor, and ToolboxAI — has its own Provider and Model setting under Settings → Tool Settings, so you can pick a different tradeoff of cost, speed, and capability for each one. There's no single right answer — it depends on how much reasoning that tool actually needs to do.
The general tradeoff
Within both OpenAI's and Anthropic's current lineups, ToolboxAI offers a small/fast tier, a mid tier, and a large/most-capable tier:
| Provider | Smallest / cheapest / fastest | Mid-tier | Largest / most capable |
|---|---|---|---|
| OpenAI | GPT-5.6 Luna (Recommended default) | GPT-5.6 Terra | GPT-5.6 Sol |
| Anthropic | Claude Haiku 4.5 (Recommended default) | Claude Sonnet 5 | Claude Opus 5 |
As you move from left to right in either row: cost per request goes up, response time gets a little slower, and the model gets meaningfully better at nuanced reasoning, following complex instructions, and handling messy or ambiguous data. Exact pricing changes over time — check OpenAI's pricing page or Anthropic's pricing page for current rates before deciding at scale.
Picking a model per tool
Activity Summarizer — usually a good fit for the smallest tier (GPT-5.6 Luna or Claude Haiku 4.5). Summarizing a thread of activities is a fairly bounded task, and it runs often, so cost and speed usually matter more here than squeezing out marginal quality gains.
Extractor — depends on how many fields you're extracting and how well-structured your source data (activities, notes) is. If you're pulling a handful of well-defined fields, the smallest tier is usually enough. If you're asking it to infer fields from noisy, unstructured activity notes, step up to the mid tier (GPT-5.6 Terra or Claude Sonnet 5) for better accuracy.
ToolboxAI — this is the tool users interact with directly and ask open-ended questions, so it benefits the most from a stronger model. The mid tier (GPT-5.6 Terra or Claude Sonnet 5) is a solid default; consider the largest tier (GPT-5.6 Sol or Claude Opus 5) if your users are asking complex, multi-step questions or you're seeing ToolboxAI give shallow or incorrect answers on the mid tier.
Older model versions
Each provider's dropdown also lists older model lines you can pin to directly — for OpenAI, GPT-5.4 (nano/mini/full), GPT-5, and GPT-4.1 (mini/full); for Anthropic, Claude Sonnet 4.6, Sonnet 4.5, Opus 4.8, and Opus 4.7. These stay available for teams standardizing on a specific version they've already validated, but the current GPT-5.6 and Claude 5/4.5 lines above are the recommended starting point for new setups.
Custom models
If you want to use a specific model not listed in the dropdown (a newer release, or a specific dated snapshot), choose Custom model... and enter the exact model identifier (e.g. gpt-5.6-luna) in the field that appears. ToolboxAI will pass that identifier straight through to the provider, so it needs to match the provider's API exactly.