Claude Calendar Integration Fails Under Compute Limits
Key Points
- The new Claude feature that links calendar and email promised powerful daily briefings, but in practice it returned incomplete meeting and email lists, delivering a poor user experience.
- Anthropic’s core limitation is compute capacity, leading to aggressive rate‑limiting on API calls (e.g., only ~50 calls per month even on a $100 plan), which quickly exhausts limits when accessing multiple docs, calendars, or emails.
- This compute‑constrained environment forces Anthropic to roll out tools more conservatively than competitors like OpenAI, which recently introduced broader, behind‑the‑scenes tool integration for its models.
- The combination of limited calls and throttling prevented Claude from correcting its errors or rebuilding the briefing, highlighting how Anthropic’s resource constraints directly impact the reliability of its agentic tool features.
Sections
- Claude Calendar Integration Fails - The user discovered that Anthropic’s compute‑rate limiting caused the new Claude email‑and‑calendar feature to return only partial meetings and messages, resulting in a disappointing experience.
- Claude’s Token Limits vs ChatGPT’s Streaming - The speaker critiques Claude’s narrow output window and lack of token‑streaming architecture, contrasting it with ChatGPT’s design that masks token limits and delivers a smoother compute‑driven user experience.
- Capital Constraints Shaping AI Race - The speaker argues that limited funding and compute resources, not just open‑source advances, are dictating the pace and strategies of AI development, as seen in recent model rollouts.
Full Transcript
# Claude Calendar Integration Fails Under Compute Limits **Source:** [https://www.youtube.com/watch?v=TMnJDCNx06A](https://www.youtube.com/watch?v=TMnJDCNx06A) **Duration:** 00:08:35 ## Summary - The new Claude feature that links calendar and email promised powerful daily briefings, but in practice it returned incomplete meeting and email lists, delivering a poor user experience. - Anthropic’s core limitation is compute capacity, leading to aggressive rate‑limiting on API calls (e.g., only ~50 calls per month even on a $100 plan), which quickly exhausts limits when accessing multiple docs, calendars, or emails. - This compute‑constrained environment forces Anthropic to roll out tools more conservatively than competitors like OpenAI, which recently introduced broader, behind‑the‑scenes tool integration for its models. - The combination of limited calls and throttling prevented Claude from correcting its errors or rebuilding the briefing, highlighting how Anthropic’s resource constraints directly impact the reliability of its agentic tool features. ## Sections - [00:00:00](https://www.youtube.com/watch?v=TMnJDCNx06A&t=0s) **Claude Calendar Integration Fails** - The user discovered that Anthropic’s compute‑rate limiting caused the new Claude email‑and‑calendar feature to return only partial meetings and messages, resulting in a disappointing experience. - [00:03:35](https://www.youtube.com/watch?v=TMnJDCNx06A&t=215s) **Claude’s Token Limits vs ChatGPT’s Streaming** - The speaker critiques Claude’s narrow output window and lack of token‑streaming architecture, contrasting it with ChatGPT’s design that masks token limits and delivers a smoother compute‑driven user experience. - [00:07:57](https://www.youtube.com/watch?v=TMnJDCNx06A&t=477s) **Capital Constraints Shaping AI Race** - The speaker argues that limited funding and compute resources, not just open‑source advances, are dictating the pace and strategies of AI development, as seen in recent model rollouts. ## Full Transcript
I tried the new Claude feature that
connects your calendar and your email
into Claude. I was so hopeful. I was
like, I'm going to get insights from my
email I've never gotten before. I'm
going to finally hook an LLM into my
email account instead of using a third
party. I'm going to be able to get all
the power of a Frontier class model in
my inbox on my calendar preparing a
daily briefing for me. I had all these
ideas. No, none of it did not work.
And I dug into why and it's a super
interesting reason and it gets at one of
the core issues facing Anthropic right
now. Fundamentally, Anthropic is compute
constrained. That is shaping everything
they do and it's shaping this feature
roll out and making it much more
disappointing than it needs to be. I'm
sure Claude is a good model. My
experience with Claude has been great
overall. My experience with Claude on
this was terrible. Even though Claude
generally codes well, when I asked
Claude to code me a briefing React
artifact based on inputs from email and
from
calendar, it just did a lousy job. It
made one call. It dropped one list. It
got half of my meetings, not all my
meetings. I grant you I have a lot of
meetings, but still it pulled like the
first seven or eight. And it pulled some
of my emails, like the first five. And
when I dug into why, it turns out
Enthropic is probably rate liimiting the
call on the back end to save costs.
Which is why even on the max plan, if I
were paying a hundred bucks a month for
Anthropic, I would still only be getting
50 50 calls to calendar or docs or email
total.
that eats up real fast if you use it to
like look at three docs a day, if you
use it to look at your calendar twice a
day, if you use it to check your email
and like work on email responses. It's
gone like that. Now, they say they're
going to lift it eventually, but the
fact that even paying a hundred bucks a
month doesn't move that for me tells me
how compute constrained they are right
now. And I am concerned that we are at a
point where there is pressure on model
makers to show agentic tool use because
you notice like they dropped this the
same week that OpenAI drops 03 which has
I believe access to something like 600
tools that it can work with under the
surface like not where you can see it.
You don't get to pick it from a drop
down. Uh but it chooses the tools and
interacts with them. Well, fundamentally
they're just giving Claude more tools.
It's a gentic tool use wrapped inside a
chatbot poorly
described. The problem is anthropic is
much more compute constrained than chat
GPT. And at the end of the day, that is
showing up in the way these
rollouts actually go. And so I did I
jumped in, I grabbed the calendar,
grabbed the email, and I had a terrible
experience. And when I went back and
tried to get it to rebuild, I said,
"Claude, try again." Like, typically
this gets fixed when Claude fixes it
again because 3.7 and honestly 3.5 are
pretty good about fixing issues. Claude
could not go back and get even the basic
complete calendar drop on the second and
third try. Could not go back and get
like the last 15 emails. And when it
did, could not generate meaningful LLM
insights off of all of that data that it
had just ingested.
It may well be pulling those data
sources as separate
uh context ingests, so it can't look at
them as a merged file and view them
together. I don't know. But at the end
of the day, it's just not a great
customer
experience. I also think it's
increasingly problematic that Claude has
in theory a gigantic context window to
ingest, but a very very poor orderof of
magnitude context window output. In
practice, limiting your context window
output to something like 8K tokens on a
turn like it just feels so short and it
feels like Claude is deliberately
cutting corners even if you can pull the
entire dock in. And I think Chad GPT has
done a much better job masking that with
the way they've handled it. So their
theoretical token input limit doesn't
feel that way because they can stick
stuff on disk and like stream the tokens
in as they need them, which is something
that's sort of special to how they've
architected their LLM. And it basically
from a consumer perspective means that
you don't notice the token limits. And
so we spend time talking about token
limits as if they are the beall end all
of these experiences. But the reality is
compute limits are more interesting.
Open AAI is playing with their compute
because they have a lot of it. They're
doing things like um streaming tokens in
storing stuff on server making sure that
if they roll something out they have the
compute allocated so it's actually a
good experience for people offering
unlimited queries things like that. And
Claude is compute constrained
fundamentally and is basically only
going to produce the output tokens that
it can produce on any given turn. So it
looks really it feels chunky and is not
going to be able to actually use the
tools because the compute is constrained
on the tool use like you don't do 600
tools if you are open
AI unless you are confident you have the
compute to sustain it. And that's what
they have with 03. And so in a sense, I
do think what we are seeing in the roll
out for Claude with calendar and with
email. It's not fundamentally about the
knowhow of the team or the quality of
the model. It's not even about the token
limits, which are what you read on the
10 and I think are increasingly
deceptive. It's really about the capital
constraints in the space. Claude is less
well capitalized than OpenAI. Claude is
getting less GPUs netn net and Claude is
falling behind because of that hard
fact and that doesn't reflect at all on
how much I admire the team or what
they're trying to do with claude. Claude
is a model I love especially cloud 3.5.
It has hit a sweet spot. But at the end
of the day, the compute constraint means
that the customer experience which is a
layer removed from just output. Like we
don't just have a customer experience
that is input tokens, output tokens. We
have a whole layered customer experience
where we are looking at interface. We
are looking at how the conversation
feels and flows. It's those little
things like uh like I was saying
choosing to store a disc or chat GPT
choosing to show a little fragment of
train of thought and not the whole chain
of thought. Um choosing to prioritize a
particular experience on a particular
plan so it feels complete. Uh which is
why I think chat GPT has constrained 03
to only 50 queries per week on the plus
plan as of this date. um and much more
on the I think it's unlimited on the pro
plan. They they're doing that because
they want to provide complete
experiences to the people using the chat
as opposed to gating it and letting
everybody have it but giving everyone
kind of a bad
experience and I think that's a better
choice. Um, and I'm not saying that
Claude is not trying to gate a bit, but
I noticed the gating is not very
effective because
effectively they don't unlock the gate
at the pro plan. Like we talked at the
beginning of this, a h 100red bucks a
month and they are still not ungating
those 50 calls out to docs, to calendar,
to email. That is compute constraint.
That is capital constraint. And I think
we need to talk more about how capital
constraints are starting to shake out
the dynamics of this race. I know we
like to talk about open source. Deepseek
has done amazing things. That's another
video for another day. But capital
constraints are real. GPU constraints
are real. Compute continues to be a
driver of success in this space. That is
not going away anytime soon. And I think
that part of what we see with the roll
out of 03 this week, especially when you
compare it to what Claude needed to do
strategically to roll out an agentic
tool this week, you see the capital
constraints on display.
Tell me your thoughts.