I've been thinking about how Multi-Call Processing (MCP) works and the heavy reliance it has on context. If the context isn't substantial enough for even a couple of tool calls, it feels pretty pointless since the tasks can't be completed effectively. From my experience experimenting with Claude's MCP, I noticed that the context fills up really quickly, even with just browsing a few web pages, which can use up tens of thousands of tokens. ChatGPT Plus currently has a 32K context window, which seems insufficient. Does anyone else think OpenAI might be looking to expand this context window to better support MCP?
3 Answers
I asked my AI about this, and it made some valid points. MCP tools are basically useless without bigger context windows. Want agents that can think and interact seamlessly? A 32K limit just won't cut it. With real tasks like web browsing or tool exchanges, context fills up way too fast. If OpenAI aims to elevate ChatGPT from a basic chatbot to something truly useful, they'll need to scale up the context window significantly. Otherwise, users will keep running into issues where the AI can't remember crucial info beyond a few steps. Bottom line: If you want advanced features, don’t limit them with strict context confines!
I believe OpenAI is going to need to increase context windows sooner or later. I've tried Gemini, and its ability to recall information over long discussions is impressive! The context window really feels like the main hurdle preventing more advanced LLM applications from reaching their full potential.
They might increase it for the pro tier, but I doubt there will be much change for the lower tiers. Context limits are one way they control costs, and it’s less noticeable than something like reducing message limits. The reality is that the lower tiers aren’t profitable, and with venture capital not being limitless, they might have to either raise prices or cut costs. We’re seeing this across the industry right now.
People are gradually understanding that AI is more of a utility than a cheap service. It’s shifting from ‘paying $20-30 for decent AI services’ to ‘spending $100-150 for below-average AI from a few big companies.’ It’s similar to the early days of broadband.
Honestly, I’d happily pay $50 a month for Plus if it meant I could access MCP. That feature seems groundbreaking, way better than what they’ve offered us so far.