Anthropic "crashes" when using Search APIs
planned
Vlad
Anthropic (Claude 3.5/3.7) consistently exceed token rate limits when using search APIs (Google, Brave, Perplexity, + possibly others). Issue persists for 3-6 months, seems to affect only Anthropic models, and lately made the service practically unusable.
Details:
macOS 15.4.1
BoltAI v1.35.1 (Setapp Version)
Daniel Nguyen
planned
For some reasons, Claude models made a lot of search requests and then reached the rate limit.
I'm not sure what would be the best solution here. Maybe limit number of tool calls?
In the meantime, you probably might want to switch to OpenRouter instead. Most of my other users switched to OpenRouter for higher rate limit and generally better performance.
M
Mark Levison
Daniel Nguyen I looked at OpenRouter, I find it impossible to understand how much money I'm about to spend. What I love about BoltAI is my spending is strictly tied to the tokens I use.
Daniel Nguyen
Mark Levison They don't charge any markup fee so basically it's the same as if you pay the AI provider directly. You can see your usage details in this page (including tokens + costs) https://openrouter.ai/activity
M
Mark Levison
Daniel Nguyen ok a bit better. Three challenges that remain:
- I've no idea how much money I would feed it
- I don't really know how I would use it in Bolt
- Cognitive Load - I already have too many AI tools in my tool chest:
(posted on talk.mpu: https://talk.macpowerusers.com/t/where-do-you-get-your-ai/40688 ) I just outlined my tool chain-
- BoltAi lets me interact with Anthropic, Google and OpenAI via apikeys I probably spend a few dollars a month
LMStudio - local models running directly on my Mac - Currently Qwen3 models are very good. Cost either free or $6K Canadian to purchase an M3Max with 64GB of RAM
- Obsidian Copilot
- Google Gemini via the web when I need search.
- Notebook LLM via the web
- VSCode and Copilot
I don't want more complexity, I want to simplify.
Vlad
Daniel Nguyen No worries, I'll "survive". However, I'd appreciate it if you could fix the issue. If it's not too much work, adding functionality that limits function calls (and a minimal user interface for that) could lead to a 'new feature' rather than just a 'bug fix', could it not??...
S
Smet Denis
The same issue.