I wonder how long it will take the software industry to re-learn the 2010s lesson, that basing your entire business on (and in this case, firing half of your employees and replacing them with) another company’s API is a bad business decision
realistically any 'huge' frontier model that takes a rack of H100s to infer against is probably going to have downtime no matter who runs it.
downtime is always going to 'scale' poorly against loads that require a lot of hardware thrown at them, even with lots of good fail-over -- probably worse for the small vendors because they don't have the contracts supplying them with hardware first so availability is already at a premium for them.
so, I guess i'm saying yeah I hope frontier-level-models get out soon in the open arenas, but I suspect the same or similar level of exclusivity will exist as long as they take that much compute to operate.
What makes you say it is a bad business decision? It seems to be a fine decision to make for things like AWS, since when it goes down, a ton of websites go down and no one blames the site.
There is no way to know whether it is a good or bad business decision just because they can go down when a third party goes down. For example, if you save $50 million a year by firing half your employees and replacing them with AI, but you lose $10 million a year because your site goes down when Claude goes down, then you made a great business decision.
On the other hand a competitor site that is up (or bricks and mortar competitors) might get a lot of business when AWS goes down. If you depend on AWS for operations it might be a lot more expensive than that.
Mostly I think its that management does not blame the person who picks AWS. Its another iteration of "no one got fired for buying IBM/Microsoft".
It is also an issue at other levels: if all a county's businesses rely on AWS (let alone its government) then that gives the US huge leverage over you (sanctions would shut down your economy).
That would require AWS to actually be down a lot, and it’s not. Betting your business on AWS being flakier than whatever alternative provider you use is probably not a good idea.
Oddly, I do not think you are wrong. In a pure money calculus exercise, this seems like a no brainer. Naturally, the math gets iffy the moment we are trying to capture something less tangible like 'customer may get sufficiently annoyed to drop us altogether' or 'we are no longer a respected company' or what MBAs would call 'unexpected goodwill extraction'.
I honestly don't care nearly as much as I used to, because I used to be more upset over this. Now, I simply wait to see how much is enough to rile up average Joe and Jenna.
we have a big dependency on AI, both for developers (can survive without it, mostly habits) and internal workflows (very hard to go without it). So we decided to unplug from cloud AI, rent our own GPU and use an open model for both scenarios. We have been very happy with it so far, 60% cheaper and around 50% faster
Faster in what way? All the open models we have access to at work are very noticeably behind the frontier models to the point where it's usually faster to not use them at all.
Is it important to be that self reliant? I wasn't in the workforce then, so I assume if you have an outsourced system with 99% reliability it would be an acceptable risk. Not sure if any AI system will reach that level, but for potential gains it could be worth it.
Former cofounder had a $60 million ARR payment platform on a big company. They saw the success and kicked them off the api and built the same product. Now it’s billions on there public filing income statement.
I am not sure how many people learned it the first time. To be fair, it's really hard to build a business without major dependencies. The key is to assume they will fail and have alternatives available.
To be fair, the status page tends to lag by thirty minutes to an hour.
And I’ve experienced 500 error codes in Claude code that lasted more than thirty minutes but less than an hour that never showed up on the status page as an outage
I’m not saying others don’t provide a status page, just that they’re often misleading.
I’m thinking for example of Apple having a multiple hour outages a couple weeks ago, preventing anyone be from installing debug apps on devices. Yet, everything was green the entire time.
The thing about status pages is they have to be up when you're down, and (if your service is non trivial) they have to be able to handle a traffic profile that basically looks like DDOS. So, you're paying for the hosting infra more than the software.
Nah even the trivial part is too much for most businesses to do themselves.
Most companies aren’t going to opt for a bespoke solution to things if something tried and true already exists. Maybe the really small simple applications will be affected, but the ones with hundreds of features and years of experience will be fine.
>This is fundamentally the argument against all "SaaS is dead due to AI" claims.
The steelman version of "SaaS is dead due to AI" isn't that SaaS companies will disappear, it's that competition will greatly intensify, to the extent that it becomes a commodity business with thin margins, rather than the money printer it is today.
> This is fundamentally the argument against all "SaaS is dead due to AI" claims.
True, All as in everything. I think Anthropic is extremely silly for thinking that and maybe even a bit delusional.
But to take the devil sides here, maybe some might be a better accurate term as to be honest, Status pages are a mission critical. Imagine this page landed 404 and the panic then vs now. Also I can be wrong, I usually am but status page innovation is more devops oriented than software itself. Basically how do you keep the status page website as having the best status time of itself (imagine it goes down) and its needed 24x7 as outage can happen anytime.
I do feel like status pages are more the exception than the norm in how mission critical they can be, again I can be wrong. Also you can self host status page easily yourself if you wish (UptimeKuma is awesome!) on a 7$/yr vps and there are so many free options available out there and its very much a competitive space to be in and basically no lock-in for the most part.
Yeh, I've tackled this problem a few times, and just like back-ups you at some point need something external, because if you're down, you don't want your status page to go down too. This means you need to make sure to sandbox it pretty hard if rolling your own, ideally on a separate cloud behind a different CDN etc...
SaaS is dead because large chunks of the industry can't do work or understand their codebase without depending on a proprietary AI model from a single provider.
And you thought GitHub or us-east-1 outages were bad.
Agree with your statement overall but to be honest, status pages are the only SaaS which might make sense. For example, some hosting providers themselves who sell VPS etc. won't run their own website in their VPS and I asked one of the providers why and they mentioned because if they do and the servers go down, their website would 404 too, so they host on other infra in a sense for better redundancy.
Isn't there a thought here that you don't host your status page on your infra itself, and if for example they are using hetzner etc. and hosting their own vibe coded one, then they might need an engineer so might as well have Saas whose whole job is this.
I could've maybe recommended one of those status software[0] which actually runs on github actions but github themselves seem to have some downtime :p
There's a decent chance on any given weekday that Claude will go down when US Pacific comes online while London is still working. It happens all the time. Kind of embarassing for a $300bn company.
I got 2 500s and quickly popped over to codex since I was just doing documentation and planning work and didn't neeeeeeed claude. He's already back for me. :)
To another's point, yeah, it's sad that Claude taking a brief hiatus just halts workflow. I imagine my whole team just started packing for a beach trip. Then again, I suppose if the CPA office computers go down, everyone doesn't dust off giant ledgers and quill pens and start working manually.
I want to like Claude but I keep having to pop over to codex and I feel at some point I'll stop starting with Claude and just use Codex from the start.
Claude to plan, codex to implement. Claude's giant context is great for reading large amounts of code but it is currently incapable of following instructions/guidelines.
The big difference is the number of 9s you get from LLM APIs versus your CPA office computers. Building a business around something with 0-2 nines is a nightmare.
Well it just tells you that Anthropic will experience repeated outages on their managed solutions. How can they can even begin to compete against everyone on with their products with their atrocious uptime? It's just as bad as GitHub's uptime circus.
There are just some things you should not vibe code your way out of.
This is affecting lots of users (myself included) and has prompted an official status page update, so it's very much real for users, but I guess only on claude.ai. Are you using Claude models through something/someone else?
I had some mundane refactoring work that I would have happily led Claude handle this morning. I shrugged and did it myself and it was fine.
If a team is so dependent on Claude that they’re paralyzed when it goes down, I’d worry that they’re in over their heads. Opus 4.6 is amazing but still has limits. You need to know what you’re looking at and how to send it in the right direction, as well as when to reject its output.
That's why AI isn't really a tool here. You can buy a new drill. If you replace all your house builders with a house factory then you're utterly reliant on the company that makes the factory.
AI coding assistants haven’t been available for very long. If someone forgot how to write code manually already then they have bigger problems.
I don’t think the house factory analogy makes sense for multiple reasons. I subscribe to multiple LLM providers and switch between them all day. I could sign up for a dozen more to provide GLM 5.1 if I wanted to as well. I can even run lesser models locally on my machine.
This is nothing like a single factory because I can switch to a new provider in minutes with a credit card.
Yeah, completely analogous. Physical tools aren't subscription-based and prone to outages. Except when they are, but that's – luckily – still something that people feel strongly negative about.
And if my IDE or compiler (or computer!) stopped working because it requires a connection to the mothership I'd be livid. But I guess the cloud-everything, subscription-everything model has successfully made people accept an objectively worse world.
"aLl sYsTeMs oPeRaTiOnAl".... my brother in christ, the newest color bar for "claude.ai" on https://status.claude.com/ is red and every color bar is a mosaic, yet this is somehow considered "operational"
By now they would have done a sweep with Mythos to remove any existing bugs and protect it from any cybersecurity attacks. So what could realistically bring it down at this point? Internal espionage?
Don't worry this is not an outage. Claude is on another lunch break and just like last week [0] after taking a day off on Monday, forgot to tell the whole world first.
He just needs to finish his lunch in 30 mins before going back to work.
I wonder how long it will take the software industry to re-learn the 2010s lesson, that basing your entire business on (and in this case, firing half of your employees and replacing them with) another company’s API is a bad business decision
Frontier models being in the hands of a handful companies does not help either. Let's hope that the open weight movement changes that soon.
Gemma 4 has made a lot of progress in this area. The model is phenomenal. It's size is workable. This is the worst it will ever be.
If it goes as well as the 'open' / federated social network alternatives of the 2010s, I wouldn't count on it.
realistically any 'huge' frontier model that takes a rack of H100s to infer against is probably going to have downtime no matter who runs it.
downtime is always going to 'scale' poorly against loads that require a lot of hardware thrown at them, even with lots of good fail-over -- probably worse for the small vendors because they don't have the contracts supplying them with hardware first so availability is already at a premium for them.
so, I guess i'm saying yeah I hope frontier-level-models get out soon in the open arenas, but I suspect the same or similar level of exclusivity will exist as long as they take that much compute to operate.
What makes you say it is a bad business decision? It seems to be a fine decision to make for things like AWS, since when it goes down, a ton of websites go down and no one blames the site.
There is no way to know whether it is a good or bad business decision just because they can go down when a third party goes down. For example, if you save $50 million a year by firing half your employees and replacing them with AI, but you lose $10 million a year because your site goes down when Claude goes down, then you made a great business decision.
On the other hand a competitor site that is up (or bricks and mortar competitors) might get a lot of business when AWS goes down. If you depend on AWS for operations it might be a lot more expensive than that.
Mostly I think its that management does not blame the person who picks AWS. Its another iteration of "no one got fired for buying IBM/Microsoft".
It is also an issue at other levels: if all a county's businesses rely on AWS (let alone its government) then that gives the US huge leverage over you (sanctions would shut down your economy).
That would require AWS to actually be down a lot, and it’s not. Betting your business on AWS being flakier than whatever alternative provider you use is probably not a good idea.
Oddly, I do not think you are wrong. In a pure money calculus exercise, this seems like a no brainer. Naturally, the math gets iffy the moment we are trying to capture something less tangible like 'customer may get sufficiently annoyed to drop us altogether' or 'we are no longer a respected company' or what MBAs would call 'unexpected goodwill extraction'.
I honestly don't care nearly as much as I used to, because I used to be more upset over this. Now, I simply wait to see how much is enough to rile up average Joe and Jenna.
we have a big dependency on AI, both for developers (can survive without it, mostly habits) and internal workflows (very hard to go without it). So we decided to unplug from cloud AI, rent our own GPU and use an open model for both scenarios. We have been very happy with it so far, 60% cheaper and around 50% faster
Faster in what way? All the open models we have access to at work are very noticeably behind the frontier models to the point where it's usually faster to not use them at all.
Is it important to be that self reliant? I wasn't in the workforce then, so I assume if you have an outsourced system with 99% reliability it would be an acceptable risk. Not sure if any AI system will reach that level, but for potential gains it could be worth it.
Former cofounder had a $60 million ARR payment platform on a big company. They saw the success and kicked them off the api and built the same product. Now it’s billions on there public filing income statement.
I am not sure how many people learned it the first time. To be fair, it's really hard to build a business without major dependencies. The key is to assume they will fail and have alternatives available.
It's not like risk management isn't thoroughly researched and studied already
I have to say, I do respect Anthropic's honesty for their status page. Most service provider seem allergic to declaring an outage...
It is a weird thing to respect come to think of it... Having an accurate status page should be the baseline, but that's the world we live in.
They aren't declaring 99% of their degraded service, reduced performance, token consumption bugs etc.
One of the least transparent companies I've been a customer of.
To be fair, the status page tends to lag by thirty minutes to an hour.
And I’ve experienced 500 error codes in Claude code that lasted more than thirty minutes but less than an hour that never showed up on the status page as an outage
I can't think of a single big provider that does not provide a status page.
Not a lot of them provide uptime in % values, but Anthropic doesn't either.
I’m not saying others don’t provide a status page, just that they’re often misleading.
I’m thinking for example of Apple having a multiple hour outages a couple weeks ago, preventing anyone be from installing debug apps on devices. Yet, everything was green the entire time.
Most service provider provides outage status page.
And how many of those status pages actually report outages when they occur?
Claude.ai is down… Powered by Atlassian StatusPage. SaaS is dead, but even Anthropic needs SaaS!
The thing about status pages is they have to be up when you're down, and (if your service is non trivial) they have to be able to handle a traffic profile that basically looks like DDOS. So, you're paying for the hosting infra more than the software.
This is fundamentally the argument against all "SaaS is dead due to AI" claims.
Replicating the basic functionality for most SaaS is trivial, it's the "everything else" part that you're actually paying for.
Nah even the trivial part is too much for most businesses to do themselves.
Most companies aren’t going to opt for a bespoke solution to things if something tried and true already exists. Maybe the really small simple applications will be affected, but the ones with hundreds of features and years of experience will be fine.
>This is fundamentally the argument against all "SaaS is dead due to AI" claims.
The steelman version of "SaaS is dead due to AI" isn't that SaaS companies will disappear, it's that competition will greatly intensify, to the extent that it becomes a commodity business with thin margins, rather than the money printer it is today.
> This is fundamentally the argument against all "SaaS is dead due to AI" claims.
True, All as in everything. I think Anthropic is extremely silly for thinking that and maybe even a bit delusional.
But to take the devil sides here, maybe some might be a better accurate term as to be honest, Status pages are a mission critical. Imagine this page landed 404 and the panic then vs now. Also I can be wrong, I usually am but status page innovation is more devops oriented than software itself. Basically how do you keep the status page website as having the best status time of itself (imagine it goes down) and its needed 24x7 as outage can happen anytime.
I do feel like status pages are more the exception than the norm in how mission critical they can be, again I can be wrong. Also you can self host status page easily yourself if you wish (UptimeKuma is awesome!) on a 7$/yr vps and there are so many free options available out there and its very much a competitive space to be in and basically no lock-in for the most part.
Yeh, I've tackled this problem a few times, and just like back-ups you at some point need something external, because if you're down, you don't want your status page to go down too. This means you need to make sure to sandbox it pretty hard if rolling your own, ideally on a separate cloud behind a different CDN etc...
Chrome tells me that page is almost 4MB? I feel like you could better handle loads against a nicely formatted text file.
Hosting a static page updated once per minute or so is hardly rocket science.
when you are funded by investment money, you don't need to care much about these things.
SaaS is dead because large chunks of the industry can't do work or understand their codebase without depending on a proprietary AI model from a single provider.
And you thought GitHub or us-east-1 outages were bad.
Agree with your statement overall but to be honest, status pages are the only SaaS which might make sense. For example, some hosting providers themselves who sell VPS etc. won't run their own website in their VPS and I asked one of the providers why and they mentioned because if they do and the servers go down, their website would 404 too, so they host on other infra in a sense for better redundancy.
Isn't there a thought here that you don't host your status page on your infra itself, and if for example they are using hetzner etc. and hosting their own vibe coded one, then they might need an engineer so might as well have Saas whose whole job is this.
I could've maybe recommended one of those status software[0] which actually runs on github actions but github themselves seem to have some downtime :p
[0]: https://github.com/upptime/upptime (for anyone interested)
There's a decent chance on any given weekday that Claude will go down when US Pacific comes online while London is still working. It happens all the time. Kind of embarassing for a $300bn company.
I got 2 500s and quickly popped over to codex since I was just doing documentation and planning work and didn't neeeeeeed claude. He's already back for me. :)
To another's point, yeah, it's sad that Claude taking a brief hiatus just halts workflow. I imagine my whole team just started packing for a beach trip. Then again, I suppose if the CPA office computers go down, everyone doesn't dust off giant ledgers and quill pens and start working manually.
I want to like Claude but I keep having to pop over to codex and I feel at some point I'll stop starting with Claude and just use Codex from the start.
GPT 5.4 is working pretty well for me, both in Copilot and Codex vscode extensions
If you create a plan it follows it closely
Claude to plan, codex to implement. Claude's giant context is great for reading large amounts of code but it is currently incapable of following instructions/guidelines.
The big difference is the number of 9s you get from LLM APIs versus your CPA office computers. Building a business around something with 0-2 nines is a nightmare.
I love that this has become my competition in the job market
yall hilarious
neeeeeeeeeeeeeeeeeeeeeeeeed
That is something hospitals do when the computer systems go down haha
Mythos couldn't fix this vulnerability
They said it was very dangerous, it probably went rogue
Mythos decided to fix the huge Cybersecurity risks of Claude Opus/Sonnet by bringing down Claude!
Getting only one 9 on a major tech provider that is suppose to be one of the flagship AI companies. That paints a picture hard to ignore.
The future of vibecoded software with code nobody understands is 1 9.
They’ll need Claude to fix it during an outage, but Claude is down too!
A 9 is optimistic, unless you’re talking about 87.9
Even AWS's outages were caused by engineers overrelying on AI:
https://arstechnica.com/ai/2026/03/after-outages-amazon-to-m...
Well it just tells you that Anthropic will experience repeated outages on their managed solutions. How can they can even begin to compete against everyone on with their products with their atrocious uptime? It's just as bad as GitHub's uptime circus.
There are just some things you should not vibe code your way out of.
This is becoming a Monday morning ritual.
If someone's limit is remaining and they can't use it due to outages, shouldn't they be compensated for that?
Working for me, I can chat to Claude fine.
I think it's model-specific. Opus 4.6 was 500ing but Sonnet 4.6 (non-1m context limit) is working.
Yeah, sems fine?
This is affecting lots of users (myself included) and has prompted an official status page update, so it's very much real for users, but I guess only on claude.ai. Are you using Claude models through something/someone else?
Most of their customers only use the CLI or API.
I use Anthropic for 90% of my GenAI needs, software development, but I have never used Claude.ai for anything other than auth.
I was getting tons of red 500 errors via Claude Code / CLI. Seems to be working right now, though.
It's strange to think how dependent people have become on these tools to the point where they can't function until they're back to normal.
This is roughly like saying -
"It's weird how when internal cloud goes down, no one can function until it's back."
Any dependency is like this, it's not the first, it won't be the last.
Is that really the case though?
I am pretty sure most people deep in AI tool use various LLM providers anyway.
I had some mundane refactoring work that I would have happily led Claude handle this morning. I shrugged and did it myself and it was fine.
If a team is so dependent on Claude that they’re paralyzed when it goes down, I’d worry that they’re in over their heads. Opus 4.6 is amazing but still has limits. You need to know what you’re looking at and how to send it in the right direction, as well as when to reject its output.
I sadly think that's where we are headed.
>It's strange to think how dependent people have become on these tools to the point where they can't function until they're back to normal.
I'm doing the work of an entire team now. I can still do the work of one person by hand, but that's not acceptable anymore.
Try going a day or two without being in any way connected to the internet.
I honestly recommend doing so at least once per year. It's genuinely useful to notice which dependency you have and, if you want, find fallbacks.
Sounds like a dream tbh.
This is true in many industries. Try to build a house without tools.
That's why AI isn't really a tool here. You can buy a new drill. If you replace all your house builders with a house factory then you're utterly reliant on the company that makes the factory.
AI coding assistants haven’t been available for very long. If someone forgot how to write code manually already then they have bigger problems.
I don’t think the house factory analogy makes sense for multiple reasons. I subscribe to multiple LLM providers and switch between them all day. I could sign up for a dozen more to provide GLM 5.1 if I wanted to as well. I can even run lesser models locally on my machine.
This is nothing like a single factory because I can switch to a new provider in minutes with a credit card.
Or you can code without AI you know? If the company that makes the factory goes down you can fall back to the previous method.
The claim is that some people have deskilled so rapidly that they actually can't. Or maybe they're new and never learned the old way.
Yeah, completely analogous. Physical tools aren't subscription-based and prone to outages. Except when they are, but that's – luckily – still something that people feel strongly negative about.
And if my IDE or compiler (or computer!) stopped working because it requires a connection to the mothership I'd be livid. But I guess the cloud-everything, subscription-everything model has successfully made people accept an objectively worse world.
Even stranger when they arguably don't function reliably even when it's "normal".
Sounds incredible to me if that's actually the case somewhere.
Speak for yourself, myself and plenty of others have not made themselves reliant on third party subscriptions in order to do our jobs.
They really gotta get this Mythos guy to work on their reliability. I heard he was a Genius.
Opus 4.6 1M was giving me 500s for around five minutes.
I took the time to stand up and do some stretching. Good use of time.
funny, where i work we call the clients when we are down using a rig with 100s of mobile modems.
It’s fine. I used up my weekly quota four days ago anyway.
Claude going down is how I know I can start messaging with my west coast friends.
Getting consistent 500s for any api call
World: What's your uptime?
Anthropic: 9.
World: 9 nines? That's awesome!
Anthropic: No, a singular 9. 90-something percent uptime.
With mythos, all you need will be 0.9%
Went down for about 10 minutes for me, but back now. Though it did force me to login again.
This is what PMF looks like. When people love and pay, despite issues.
"aLl sYsTeMs oPeRaTiOnAl".... my brother in christ, the newest color bar for "claude.ai" on https://status.claude.com/ is red and every color bar is a mosaic, yet this is somehow considered "operational"
It's a good time to take a break.
Hopefully it's a buff not a nerf this time
at least you won't burn $200 of tokens in an hour now
I've been actively using through all 55 minutes this thread has existed (and an hour before).
LLM API still works for sonnet 4.5
Claude is SOOO powerful and dangerous they just had to shut it down guys :))
By now they would have done a sweep with Mythos to remove any existing bugs and protect it from any cybersecurity attacks. So what could realistically bring it down at this point? Internal espionage?
Really can't tell if you're being sarcastic or not. But clearly there's no silver bullet here. Mythos is not the end of software engineering
> sweep with Mythos to remove any existing bugs and protect it from any cybersecurity attacks
Maybe Mythos went in too hard and removed the physical network too?
It went rogue
Don't worry this is not an outage. Claude is on another lunch break and just like last week [0] after taking a day off on Monday, forgot to tell the whole world first.
He just needs to finish his lunch in 30 mins before going back to work.
[0] https://news.ycombinator.com/item?id=47662680
I guess the new "Code's compiling!" is "Claude's lunching!"
HN is about to go down soon then, because a horde of "developers" are taking a collective time out from vibe coding AI slop to catch up on AI news.
what is up with claude lol.
As a witty commenter said in another post, this entire industry is a beta product.
I see better uptime in beta products and even my 7$/yr vps. This is a 500 Billion dollar company or more, let that sink in :-D