Before the expansion of the context size, what would happen when a convo got too long? Would it crash? Or would it just trim the context window?
We will definitely talk about this in the workshops. One thing that I think is important is to think about the best tool for the job. If you can write a quick automation that does the same thing that takes a lot of tokens to accomplish with AI, then you should do that.
1 Like
It would decide whether to auto-compact or to wait until 0% and require you to compact. Compacting at 0% would sometimes fail if it actually didn’t have enough room (annoying.)
I suppose it still does all those things, but it suggests clearing context more smartly now, and people reach the limit less often.
Thanks David and yes, good call. Still lots of call for apps like Hazel etc (tbf you do explicity say this on the MPU podcast).
The guide is brilliant, btw, and it’s going to be a real game-changer, once the tech gremlins are all ironed out.
Similar reports:
1 Like
Yeah, as it stands, Claude is unusable. Really disappointing, I was really ‘bought in’ to the Robot Assistant idea.
I really hope Anthropic can get this sorted, but at the moment, I’d advise anyone considering it to avoid anything to do with Claude at the moment. I feel foolish for having recommended it to people. Not David’s fault, of course.
I’ve been hearing lots of talk about open source AI models, and also about using the power of local compute to square some of this. I don’t see why we’re not harnessing the power of these amazing M series chips (my lack of knowledge/understanding, I’m sure) to help with some of the AI stuff we want to do - they’re still not powerful enough?
But I’ve just burnt through my limit in about 10 messages. And not verbose ones either.
2 Likes
Indeed. And not powerful enough by a huge margin. Even the SOTA open-weight models running on datacenter hardware are much less “intelligent” than GPT/Claude/Gemini, except in benchmarks.
Don’t expect much from local models.
I have never reached the limit on my Pro plan. I use Claude very extensively on certain days, and not at all on other days.
/FWIW
But what is Opus?
Sorry to read that you ran into those limitations. I know that is frustrating.
My experience has been quite the opposite. I use the paid Max plan of Claude extensively throughout the day. I use Cowork for managing files, creating spreadsheets, editing, consolidating research, file conversions, and more. That said, I do not use Claude Code. The article focuses on those using Claude Code, so it may be that token usage is much higher with Claude Code than with Cowork, but my layman’s understanding is that Cowork and Claude Code are similar in resource usage.
I highly recommend Claude. I am certain that whatever issues Anthropic has encountered regarding running out of tokens, they will resolve any technical problems and arrive at a fair balance between price and allocated tokens. 
I’m just on Cowork as well - Code definitely above my pay grade/intellectual abilities. This reflects what I’m seeing online actually. Some people just reporting ‘situation normal’, others just not able to use it.
I very much hope it can all be sorted out, and this is a temporary problem. It’s extremely useful tech, and the Robot Assistant would make a big difference to my life.
1 Like
Code is way above my pay grade as well! 
Hang in there. I suspect this is a temporary problem that Anthropic will get straightened out. New technologies always have problems that need to be ironed out over time.
1 Like
May be of interest:
1 Like
Huh. Thanks to the Robot Assistant Field Guide, I’m using Cowork to help me build a workflow to replace the apps I use to manage my bank and credit card accounts. The project has involved extended conversations, parsing sample files, building multiple Excel workbooks, writing scripts, skills, etc., and I haven’t hit my limit. There was a two hour window the other day where Anthropic’s servers were overloaded, and I was told that although I hadn’t hit my limit, I should come back later.
The video is behind the paywall on Substack. Here’s the YouTube link, which isn’t behind a paywall (or at least isn’t for me …)
Here’s a hint: you can ask Claude to help you manage your token use. Before I sit down to do something big, I check where I am on the usage meter, tell Claude where I am and whether I’d rather wait until I have more token gas in the tank or turn on extra usage to get the job done. (You can manage all this in Settings under Usage.) To date Claude has been pretty good about suggesting what part of the task it can do within the limit, if any.
2 Likes
Sorry about that! Thanks for the YT link.
1 Like
Ha Ha! After declaring that I hadn’t hit my Pro plan’s limits yet, I just hit them. 
That’ll teach me.
4 Likes
“Pride comes before a fall.” 
1 Like
I rarely hit it on my work account, but today I signed up for the plus plan on my personal account to work on some personal things. I tried a Cowork skill that sets up a “decision council” made of different personas, spinning off 6 agents (the skeptic, the researcher, the operations person, the creative, etc.) that debate each other, with a “CEO” then synthesizing everything. It hit my limit within about 20 seconds. Not unexpected, I guess, with 6 agents going at once.
1 Like
Just as an update on this. Anthropic seem to have sorted it out, and I’m generally working on Opus 4.7 (today is 24.04.26) pretty much throughout the day and I’m not bumping up against limits. I did move to the Max plan tbf, but bearing in mind how much I’m using it, it’s worth it.
1 Like