The Tools bucket in the context breakdown seems to vary a lot depending on which model is selected. In the same project, Fable 5 shows ~23K tokens of tool definitions while Grok shows ~10.5K.
I understand the models use different tokenizers, and I know Fable 5 uses the newer Anthropic tokenizer that produces noticeably more tokens for the same text. But that alone doesn’t seem to account for a 3x gap.
I’m curious what causes such a gap in tool definitions size and if I can reduce it for Fable.
Cursor version: 3.16.17 (user setup)
Mode: agents window, no specific mode
Hey, good question. I can see both screenshots showing the context breakdown, Fable 5 vs Grok.
The main reason for the difference isn’t the tokenizer, it’s the actual prompt content. Different models get different amounts of tool descriptions. For some models, the built-in tool descriptions are compact and single-line. For others, including Fable 5, they’re longer and include usage notes, examples, and limits. The tool set is the same, but the description text is much longer, and that creates most of the gap.
On top of that, the number in the breakdown is a calibrated estimate, not an exact tokenizer count. It’s scaled based on the total token usage reported by the provider, and Anthropic models usually report higher totals, which can inflate the Tools bucket even more.
What matters is that you can’t change this on your side. Tool description length is tied to the model and is picked to match what that model works best with. Also, MCP tools are counted in a separate bucket, so turning off MCP servers won’t affect Tool definitions.
So the Fable 5 vs Grok difference here is expected behavior, not something you need to fix on your side. Let me know if you’ve got more questions.
