Hacker Newsnew | past | comments | ask | show | jobs | submit | blfr's commentslogin

People are saying that Anthropic has no moat which sounds right but at the same time they have left everyone behind: Google is trying lately but seemed very lost over the last years despite inventing the thing; OpenAI cannot match Opus 5.5 and Fable atm; and it seems like anti-distillation efforts really did put a dent in the speed of Chinese model development.

Meanwhile Claude is getting really good. It's a decent harness now, even the remote control over a couple servers is quite OK. They're adding artifacts left and right, new features. And all without making the product seem bloated, they stay focused.

The prediction markets don't show anything on the horizon either: https://releaseoracle.xyz/

I am not a big AGI believer but business- and product-wise it seems like ANT has something that others are missing.


> OpenAI cannot match Opus 5.5 and Fable atm

OpenAI is absolutely crushing Anthropic regarding cost of running models.

A lot of real world scenarios have already been completely solved with models released months ago. My clients don't need more powerful models anymore. They need faster and more affordable models.

OpenAI has been aggressively pushing on both fronts. Where is Anthropic's 750T/s wafer-scale compute offering? Where is this on their roadmap?


It seems like everyone I know outside of tech uses ChatGPT “Chat” on the free tier.

Of the people I know in tech, other than those who don’t use either, everyone is paying for Claude.


It's pretty obvious that OpenAI is dieselgating* the benchmarks. Even 6.2 Astra is nowhere near as good as Claude models for coding. Any software engineer can do a side-by-side comparison (as many of us do) and figure it out.

*Also consider the possibility that Anthropic may be dieselgating it down ("benchminning"?) instead of OpenAI doing benchmaxxing. It might make sense from some game theory standpoint - If you are significantly better than your enemy, it might be in your interest to lull them into believing that you are only slightly better.


I do coding evals every time a new model comes out (that I have access to) and Opus 5.5 is a touch better than Sol 6.2 right now but that’s a recent development. Prior to the 5.5 release open ai outperformed at every tier.

Sol is half the price of Opus and Luna outperforms haiku at every eval I throw at it. Haiku wasn’t even able to complete several of my evals until the newest release.

The idea that Anthropic is miles ahead of OpenAI doesn’t pass the smell test in my usage.

And that’s before the Claude vs Codex comparison (which I’ll grant is more personal preference but I find Claude to be wretched).


> Also consider the possibility that Anthropic may be dieselgating it down ("benchminning"?) instead of OpenAI doing benchmaxxing. It might make sense from some game theory standpoint - If you are significantly better than your enemy, it might be in your interest to lull them into believing that you are only slightly better.

Unlikely, because it would mean that you're also "lulling" your customers and your investors into believing that you're only slightly better than your competitor ("enemy" isn't the right term here), and you don't want that. Especially not shortly before your IPO.


That's not obvious at all to me. While I prefer Opus for run-of-the-mill tasks I actually find Astra to be better for very hard problems. It's just a bit of a psychopath and has more "paperclip maximizer" tendencies. If my hypothesis is true then how they fare head-to-head depends on the workload.

> Any software engineer can do a side-by-side comparison (as many of us do) and figure it out.

That's a really strange take to me, because among my colleagues theirs basically a split between two kinds of people:

- the “I don't need frontier so I use sonnet and it's good enough” group.

- and the “I want the best model and I managed to get a Codex subscription from management because Sol is much better than Opus” team. (This only started this summer and I suspect it was mostly true because of Opus 5, but that's still going to do long term damage).

No one is arguing the other way around and I know the manager would be considering a switch of what comes as default to full OpenAI on the first occasion (he's a very cost sensitive kind of guy so I'm sure that any price hike would trigger him).


Codex doesn't have Claudish, though, and you can use other harnesses (i.e. Pi with its immense tokens savings) with a coding plan.

I use Codex for lack of Claudish _alone._ My spidai senses might still be overtuned, but I can spot Claude generated content a mile away and my brain involuntarily turns off.


Also, the free chat in ChatGPT is so annoying that I don't even want to try their coding agents.

Pretty much. I keep urging my non-tech friends to pay for Claude and they keep using free tier "Chat" like a fancy search engine and copywriter

Well, why would they do that? What use could they get from paying for a somewhat better search engine and copywriter?

Running claude on your personal computer as a non-technical user is reckless bordering on insane. You have absolutely no idea what it's doing!

I’m curious, what’s your arguments?

Do you own shares in Anthropic or is there any reason why you urge people to buy something they don't need?

Anecdata: on my personal projects I vastly prefer to use Opus, and even reach for it to realign Sol once it's started to drift on specific projects instead of bumping up to Astra.

On the other hand, at work, where I have to justify paying API prices, I almost exclusively use Sol, while occasionally bringing in Luna (though I've also been playing with the MS models since yesterday; not impressive, but very cheap).

If they do manage to catch ANT in ability while keeping costs this much lower, they could still leapfrog again. I'm not sure I'd bet on it being soon after seeing Astra 6.2, but I don't think the race is over.


Not going to speculate on how they'll do vs. OpenAI, but to your point on a lot of problems now effectively being "solved", I tend to agree.

I have one recent use case now where using Claude to review it, it first wanted to use Fable, got all excited about possible savings from lower defect rates, had to backtrack and admit that Fable provided no benefit whatsoever to justify the extra cost over Opus after it ran evals (in fact, arguably the quality was worse, but that part was a subjective measurement).

There's a mounting number of tasks like this where the value is going to be falling costs whether by switching providers, or moving down their model ladder as the one below gets good enough.

Given Opus 5.5 is now good enough for the task in my example, it means we have every incentive to start evaluating cheaper models now and again, instead of waiting eagerly for new, smarter models.


I agree, I'm on 6.2 Sol and it feels blazingly fast and it leads to a better UX. It seems to be capable enough to handle 99,5% of stuff that will be thrown at it.

>My clients don't need more powerful models anymore. They need faster and more affordable models.

Honestly this sounds like it calls more for Deepseek v4.1 flash than OpenAI's models, although those charts from earlier this week did look impressive. Trying to compete on price efficiency sounds much more daunting than competing at frontier capabilities tbh.


> My clients don't need more powerful models anymore.

who are your clients and what do they do with ai?


Lets assume the best Chinese AI lab is an entire year behind. These companies are valued below 20B, while Anthropic is valued above 1T.

Assume nobody ever catches up to Anthropic and everyone will always be 1 year behind. Is the 50x valuation still reasonable? Is the money really in doing PhD level math, or simpler stuff like summerizing coporate emails?

And lets be honest, some US and Chineese labs are only months behind and the only reason they have not fully caught up is limited compute. For China this could change quickly (e.g. Huawei gets their yield issues under control and suddenly Chinese labs have cheaper compute than west). People who invested in Anthropic at 50x valuation must belive this scenario is impossible in the next 5-10 years, but that is a huge bet.


My view is that Anthropic will be the Microsoft of the AI era. The only moat Microsoft has is that it's the default for many enterprise companies. AI needs access to a lot of confidential data to be useful in a large business. If Anthropic convinces enough customers to use Claude Meet, Claude Docs and Claude Sheets they win. Not because their models are so much better, but because all your business data is with them.

Do we have any numbers on who's actually using those ancillary products? All the non-tech folks I know don't touch them. If they want Claude in their docs, they'll just use the plugin for Google Docs or Word.

that's today. they might get used to it when their workplaces start pushing claude sheets/docs/etc on them

I think Google Sheets,etc is just too entrenched, I'm at a small shop of heavy Claude Code users but I haven't heard a whisper of even trying those other products.

Why is this more likely than firms "choosing" Microsoft's equivalent to Claude Code, once they inevitably improve the harness around copilot?

This is not a trick question or anything, I'm genuinely wondering whether stickiness (MS) or product quality/iteration speed (Ant\) would succeed.


It’s not going to happen at an enterprise level, Microsoft is so intertwined within every part of any large risk-averse organisation that copilot will be the default.

I'd never even heard of those offerings. Not sure they will take off.

How much actual research is Anthropic doing ? How about hardware research ?


Those apps are not any better than alternatives. We are trying to use them as experiment.

Also didnt these companies say no one will need saas apps anymore. Now its their main moat?


We're moving into a world where the model's intelligence is one checkbox among many in the product. As you point out, many companies are able to provide competent models. But that's not sufficient anymore. Buyers want features around collaboration, sharing, permissions, etc.

My analogues for this market are OSes, office software, and CRMs. Windows + Office remain revenue juggernauts 20+ years after fully free and open alternatives got good enough. Salesforce is worth over $100B, even though there have been good FOSS CRMs for literally decades.

IMHO it's difficult to predict how the AI market will evolve without understanding why people continue to choose relatively expensive alternatives that functionally check the same boxes.


Many American companies will preferentially select American providers, so I expect them to have very high retention and also high net revenue retention

Even if new customer growth falls apart, if they can hold 30% growth while public it will be a good buy today.


Why would a more competitive frontier necessarily mean a significantly lower Anthropic market share and not just higher market share for Deepseek etc?

Super intelligence threatens the marketshare of nearly all other companies.


Where is the compute to mass serve these Chinese models going to come from?

> Is the 50x valuation still reasonable? Is the money really in doing PhD level math, or simpler stuff like summerizing coporate emails?

No, and obviously not. Much more going on with this IPO than common sense investing.


They don’t get so much attention on HN, where the focus is naturally on coding, but the new Claude features being steadily rolled out for office work could, it seems to me, become a real moat. Anthropic has lately announced integrations of Claude into whiteboards, online meetings, etc. As people get used to using those and to having Claude as a collaborator who is knowledgeable about their work, they will become reluctant to switch to other providers.

But we already get this with copilot today, its integrated into sharepoint, teams, outlook, word, powerpoint etc. Is it a bit rough around the edges... yes, but you can finally find documents in sharepoint, where before that basically did not exist. We already have contracts with microsoft in regards to data protection, and Anthropic has not really done anything to prove that they can be trusted with it yet.

We'll also get this with tool use and browser automation. The models are increasingly good enough that they can build custom integration tailored to your specific use case. It'll be very hard to maintain a lead there when your "lead" is functionality the models can write unilaterally on the whim of a user.

Yes and that's great for the old enterprise organisations that are in the Microsoft ecosystem.

Is a 2 person startup going to choose Microsoft Teams and Sharepoint? I very much doubt it.

They can all coexist but they are for different target customers.


Your theory hinges on the idea that Claude users will subscribe to the idea of it being a knowledgeable collaborator and have that work for them.

I'm not saying that's impossible, but it's entirely speculative at this point. Developers are far more effective with these tools because they understand their strengths, limitations, and how to use them to get good results.

I've yet to see anyone benefit from treating the tool as a personified collaborator. They either scale back their use or fall flat on their face when it behaves unexpectedly.

I'm using M365 through MCP/CLI via some GraphQL interface, with access to Teams, documents, Outlook, etc. None of that depends on Anthropic.


I have non-dev clients and business partners that are all coding & mailing and doing everything with AI. Some I helped, but some are just handy. Next year goign to be insane.

This is exactly what they should be doing, yet I'm still baffled by their limitations. I can only connect one email account or one calendar to my Claude. What if I have a personal life and a job? Sorry, can't do it.

ChatGPT doesn't have this issue. It doesn't seem like it's particularly complicated.


That's assuming you're letting these integrations near your data. I recently had a look at the Claude plugin for Chrome and it honestly scares me a bit. This thing is a security nightmare (plus, probably a walking GDPR violation in the EU). They have solved neither the lethal trifecta nor prompt injections nor even unintentional misbehavior; yes, they have some mitigations and safeguards in place, but these cannot be remotely considered guarantees.

Google beats them on performance, and OpenAI beats them on price. I find CursorBench to be a pretty accurate representation for the most common use case (coding, and software architecture).

The question is for how long? Maybe OpenAI doesn’t have it right now, but won’t they have it in a couple of months? That’s been the recurring theme. Anthropic has had a noticeable advantage before, and others did catch up and momentarily surpassed it.

"ANT has something that others are missing"

If a free market provides anything, it is the essence of success that is expressed by this statement.

Moats are strategic but only one part of - and a somewhat wasteful part of - strategy.

You don't need a moat if instead of a castle you have found your enterprise to be more like an island or a oceangoing battleship for example.

Strategic defenses (like moats) are as I said somewhat wasteful because your paranoids on staff can always erect some other necessary moat, and the enterprise can spend time and resources in defensive protection rather than innovation that speeds away from any onslaught by competitors.

In their worst form, moats and 'moating' can become institutionalized as part of the enterprise where regulatory capture practices, preferred supplier arrangements, patent hegemony, etc. are distinct enterprise features: i.e, Bureaucracy.


Argon, Neon, Hydrogen and Carbon have been released. They're contenders.

I've been using the argon and neon orchestrator modes in antigravity and have been reasonably satisfied. I use Claude Opus and Mythos as well, and honestly Google's offerings are more satisfying to use. They feel faster, they're less verbose, they don't wander off and do things outside of scope. And they work.


You're right about Neon, but Carbon is definitely not less verbose. On the contrary, a massive outlier in terms of time per task on our leaderboard, data is battle data no top of OpenCode, http://agentduel.com/leaderboard. I looked into this and the issue is Carbon uses a huge amount of output compared to battle opponents.

You know, on reflection you're spot on. I'm on a Pro plan, and using Carbon burns through its allotment pretty darn fast. OTOH, Hydrogen and Neon have been trucking along well; usually I'd be out of credits by this point in the week.

I feel like "left everyone behind" seems kind of arbitrary here. They have the best coding models, but it's not like there's a huge gulf in quality. How much pain would you really go through if you had to stop using Anthropic models altogether?

> I am not a big AGI believer but business- and product-wise it seems like ANT has something that others are missing.

Wonder what that could be, not profit, that's for sure.

Their operating losses were constantly above $500 million in every quarter for the last year, and here I've though that business is supposed to produce profit...

More alarming, that there's no plan whatsoever how to turn AI profitable. Speaking and spying labubu figurines won't cut it, neither will ads.


One of the things driving the cost is having to keep training new models. It was every 6 months or so, now it seems like every 2-3 months. If they stop the other companies will catch up, so they are forced to pace (or race) the frontier to make sure they always have something better to offer if/when another company catches up.

That requires the hardware (GPUs/TPUs) and data centres to run them. It's not clear how many training runs they do, for how long, or at what cost.

In terms of profitability they need to get people to spend on tokens and subscriptions, and those need to be run cost effectively. Any subsidisation cuts into their profits.

The current boost has likely been due to tokenmaxxing, agentic coding, and convincing companies and influencers to spend money on adopting that. However, several companies have been scaling back their token spend and it is unclear how much more they can increase in this area. This is likely why they are pushing into other areas like mathematics and more recently genetics -- they need new customers to keep and increase the token spending.


> One of the things driving the cost is having to keep training new models. It was every 6 months or so, now it seems like every 2-3 months. If they stop the other companies will catch up, so they are forced to pace (or race) the frontier to make sure they always have something better to offer if/when another company catches up.

Exactly. Plus new models need more and more training data, so it will just keep costing more and more, there's no end to it (except when they run out of data, which is likely too and soon).

> This is likely why they are pushing into other areas like mathematics and more recently genetics -- they need new customers to keep and increase the token spending.

Yes, but there are not much new areas left to invade, they are running out of possibilities pretty quickly. There's only so much areas, and they have already covered most of it.

All in all, their costs are keep increasing while their chance to continue tokenmaxxing and grow income is drastically decreasing. Not a good sign.


> they need new customers to keep and increase the token spending.

This seems like a sustainable business model in the long term /s

I guess that’s the downside of getting priced so highly: there’s not many other directions to go but downhill.


I'm not saying that everything is great at ANT, but looking at last year with a company with this type of growth is a waste of time. More recently[1]:

Total operating expenses rose to $17.5 billion from $10.6 billion in the second quarter. This resulted in Anthropic reporting an operating loss of $567 million, after reporting a $974 million operating profit the quarter before. The company attributes that loss primarily to compute bought under short-term commitments, which noted in the filing “currently costs significantly more” than compute secured under its multi-year deals. The company also cites $2 billion in non-cash charges tied to an employee donation matching program.

[1] https://fortune.com/2026/11/02/anthropics-growth-slowed-and-...


Companies are made up of people and Ant/OAI are able to attract the greatest talent. Outside of that, money and data I'm not sure what people expect moat to look like. The Krabby Patty Secret Formula?

I don't know why but every time I have tried to use Anthropic models I've been disappointed. From the OpenAI stable, I use GPT 5.6 Luna and only switch to Sol 6.1 if it can't figure something out. When I don't need to use OpenAI models, I use Deepseek v4.1 Flash 0731. Every now and again I chuck a few tokens Anthropic's way but the models just chew through the $$ and never seem to be like "wow! so AI! very code!" to the extent that they appear to be worth the cost. MMMV I guess.

MMDNV - luna (6 now) is a workhorse!

> it seems like anti-distillation efforts really did put a dent in the speed of Chinese model development.

I wonder how they did that.


They hide reasoning from... well, you and me, the Claude users. I thought it was going to be annoying but don't really watch these threads while they're going anymore and the result is of no lower quality than when the reasoning traces were visible so it's not that big of a loss.

But I set "Default transcript view" to "Thinking" (not even "Verbose") and can see reasoning - is that not the thing?

Yes, it's not the thing supposedly anymore. And there are reasons to believe that because if you do watch it closely you can see that quite a bit of time the answer doesn't seem to follow the reasoning with reasoning being quite shallow or even in the wrong direction.

I never trusted the "chain of thought" to be more than a verbose response. Its a black box, all we really know is they started telling it "show your work" and the model would spit out a bunch of text that sounds reasonable enough before giving the actual answer at the end.

Those aren't the actual reasoning traces. They're highly condensed summaries. The actual thinking tokens are encrypted and invisible to the end user.

<deleted>

Their IP, which was trained on text and code from a gazillion people who never gave permission to do so.

I liked the phase where open models [cw]ould be trained on thinking traces, because it at least gave something back to all the people who's work was stolen and whose jobs Anthropic wants to destroy.

It is really strange to see people rooting for companies that want to lock up all the world's knowledge and intelligence.

We need open models to win, so that everybody has access to the means of production, not just some multi-billionaires.


When looking at history we always locked up the world's knowledge.

E.g. what would it cost an interested private person to legally access scientific papers. And most of that was founded by the taxpayers.


I remember thinking that Google had no moat in search... "it's just an index" I said. Turns out, there's so much more around it then that. When you're the top player you create moats as you go.

I mean "not bloated" is not at all how I'd describe Claude. They're constantly adding overlapping features and concepts. There is no simple clear vision, they just ship the org chart (or more likely the whim of individual engineers)

Only a sliver of a company usually knows how/why their business works and rarely are these people in internal compliance groups or anywhere outside of the executive suite, sales, and operations. Even within these, it's rarely most people.

Sometimes a smart lawyer, but one who deals with clients, or a financial controller know but this is strictly due to their individual curiosity and intelligence.


Automatic transcripts of online meetings are amazing: it means that you can just shoot the shit for half an hour and then have Claude bang out a deck and MVP for review. And no one has to waste time and attention on making notes.

Even better if it's a feedback session or people have notes and comments. You can have them all mock-implemented before you even send out the summary.


Taking notes works your brain, helps your memory, and strengthens understanding.

I learned very quickly in my undergraduate days that the act of physically writing down information was the best way for me personally to retain knowledge.

I’ve got over a decade of work experience now and I have like 100 notebooks filled with notes from nearly every day. I routinely find that my own notes differ from auto written transcripts of meetings. Mostly it’s minor things, but every once in a while the automatic transcription gets something critically wrong.


I'd asked a close work friend what his new girlfriends name was two or three times. Last time I took out my Obsidian notes on my phone and typed it in there. Weirdly, I've never gone back to look at the note but it's now committed to memory. So just typing it out on a phone works. Though it's obviously suboptimal for more graphic heavy text (drawing chemical compounds for example would be a pain but is easy with pen and paper).

Excalidraw for obsidian lets you draw notes!

I always used to take notes and casually read minutes out of habit after meetings and pick up on stuff that sounded weird and more and more recently that's been feeling like a superpower to just read things and think through the implications and raise them because people are not picking up on them.

I am horrified of the idea that HR can listen/review every single ~private conversation, that may include frustrations with leadership, family problems, or personal health information that is shared with a trusted manager or coworker, not some HR person.

Then these changes are actually a good thing because you get the opportunity to learn that they've had that power for years.

Since when?

If I go on a private walk with a manager around San Francisco, how would leadership or HR learn about this, assuming both parties do not defect.


They have cameras and badge access records. They know you went on the walk. They didn't know what you talked about and very likely don't know now unless for some reason they are wearing spy devices. The smartglasses and always on listeners are a problem, but I haven't heard of a company ordering employees to wear those and stay signed into company accounts outside of work.

Then you had better organize and consolidate your bargaining power.

Except that the summaries can be totally inaccurate. And then people think they have proof that so-and-so said something, when maybe they never said that.

Why even have a meeting then? Just let the other person(s) dictate their stuff into their computer and send you the transscript.

Great title, upvoted on its strength alone, just delightful, but a phone clipboard really isn't the place to store secrets. Sure, one-time codes, but not permanent secrets.

Instead of trying to turn the clipboard into some citadel, it would be way more beneficial to security and UX to have more reliable password manager fill-ins.


Americans are experiencing much of the same despite being the richest country in... history. Immigration is a not insignificant part of it but real estate being the only thing most people can finance with leverage has serious footprint as well.

Americans "are experiencing much of the same" while staying in their country. That's why many opt for the rich expat life. I've lost count of the number of stories I've read about random Americans moving to Portugal because "it's so much cheaper here" and "life is so much better".

Same goes for Europeans moving to East Asia to live the high life and causing similar problems for the locals. The only difference is that, unlike in Europe, East Asian countries have some guardrails on property ownership to avoid being overrun by foreigners. E.g. you cannot own land, and most houses are on a lease.


You can use Claude in Google Docs/Sheets/Slides: https://support.claude.com/en/articles/16951679-use-claude-i...


> When people buy a domain for their blog, it’s a clear sign they intend to stick around.

If you go through comments on some popular post from the golden age of blogging (~2011?), where people's names would link to their sites, the free wordpress.com or whatever blogspot blogs are up much more often than the custom domain ones.

Funnily enough, often the last post on the free blog will be the announcement of switching to a domain, now inactive.


OP's site was very slow to load though. Maybe there is something to modern Astro on Cloudflare so your site doesn't die when a couple dozen people want to read it in the same hour.


I own no iPads but all my screens are OLED and I would never buy anything else. The true black is an absolute killer feature. It makes the device just a glowing wonder at night, magical feeling, no way to overstate it.

OLED + dark mode after sundown are awesome. On Android we also get the extra dim mode which makes it perfect for lights-out.


For nighttime viewing with my Pixel 8 Pro (OLED), I wish I could make the screen even dimmer at night. The extra dim mode barely does anything.


I use GrapheneOS on my Pixel and it gets very dim. Also, I have the night mode (orange tint) cranked way up - something that the stock OS should allow you to do.


Cross-sectional correlational studies cannot establish causation and it's entirely possible (frankly, likely) that more deprived kids spend more time in front of a screen because their environments are less stimulating/interesting/safe.

This is the same as quitting smoking or just not thinking about an elephant. You can almost never cease/stop doing something, yet you can almost always displace it with something else.

(Disclaimer 'altacc suggests: I do have kids.)


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: