> increasingly commoditized
Is it increasingly though? As of 5.5 I stopped using other models all together (and I used to regularly swap, indifferently). My $20 sub no longer runs out, even though I do _much_ more than ever. I use it for every single task at work, even trivial ones, and my spend has gone down. I was given _free_ Codex and realized... it still wasn't worth it. My rational brain says Anthropic has no moat and this is temporary. The progress of all the other models is amazing. I certainly wouldn't bet on them.
But if I look at the actions of myself and the most competent coders around me, they are all doing this same thing - using 5.5 and not really looking around any longer. And we know its not even their best model. OpenAI may respond, Chinese labs may catch up. But for the moment, Anthropic seem to have added separation, not lost it, based on current vs prior _behavior_ of coders I know at home and work.
A couple of points.
First, it does not really reflect my experience. I find modern models to be largely interchangeable when it comes to common software engineering tasks, including some of the more recent open weight ones. The strongest models become interesting for Ph.D. level of work for me, which is not something I deal with regularly anymore.
None of them are perfect; they all have their strengths and shortcomings, but I can reasonably work with all of them and it honestly doesn't matter much anymore. I do sometimes wonder what exactly people are doing to see such a difference; at a minimum, their workflows must be very different from mine.
Second, you're arguing based on subscription pricing. But Anthropic makes the bulk of its revenue (and has its biggest margins) from enterprise/API pricing, where they are the most expensive provider by far. Anthropic isn't going to make a lot of money off your $20 subscription. With a subscription, you currently get more usage per dollar out of Anthropic's than OpenAI's plans, which approximately offsets the difference in API pricing.
> I find modern models to be largely interchangeable when it comes to common software engineering tasks
If you think that, it means you're not demanding enough from your models.
Opus 5.5 still makes a lot of mistakes. Even if it can technically do all tasks you throw at it, there's no way it can do it at 100% reliability.
To be clear, AI writes 100% of my code since last year. But it still produces bugs, needs multiple prompts, and has less intuition than a human.
If I'm not "demanding enough" from models, it's precisely because they are imperfect. Having them do long, complex tasks autonomously usually comes with more cleanup work afterwards IME.
My usual workflow (not counting throwaway code) is to use agents as virtual pair programmers, which gives me a good balance of quality and velocity.
Each interaction with the agent remains individually comprehensible and makes progress towards one of the following goals:
- Understanding the problem/solution space.
- Implementing a solution.
- Understanding the solution, especially any agent-written parts and interaction with existing code.
- Evaluating the solution (tests, code analysis, what-if scenarios, consider edge cases).
This does not just involve the model generating code, but also explaining or analyzing it as needed.
This process allows me to retain a mental model in lockstep with developing the written code and have some assurance both of its functionality and non-functional requirements. It also fits in nicely with most existing team-level workflows.
But it's also something that I can do with pretty much any of the modern models.
Well that's kind of my point. Models still have a lot to improve. They are still imperfect enough that you don't trust them and have to use them as pair programmers.
Yeah, but that's also a large part of why they remain interchangeable, I would say. I feel like we may be talking past each other?
They are interchangeable maybe in your case. In my case for my tasks and for my product, Opus and GPT are at a level where I can vibe code with them and trust them, whereas the others aren't really there. When models get better they'll reach whatever point you need to be able to trust them.
You said above that "Opus 5.5 still makes a lot of mistakes". This seems to contradict what you are saying here?
Sorry just woke up, I wasn't being clear.
There's two thresholds.
1. Tasks models can do. 2. Tasks models can do with enough reliability that you can YOLO merge without looking at the code.
Most models can do most of 1. That's no longer the interesting part.
2. is quickly expand. I can trust opus 5.5 enough with easy to medium tasks that I can YOLO merge them. For open models, I can only trust them with easy tasks if I want to YOLO merge.
I'd love to have a stronger model that I can trust to YOLO merge hard tasks.
I think this is where we may have to agree to disagree, because this not something that I'd do outside of toy projects. Not necessarily because I'm worried about the correctness of a single MR (in fact, it may very well be more likely to be correct than a human-authored and -reviewed one), but because I'm concerned about the accrual of technical debt. Even if each individual MR is correct, the trajectory for code quality can trend downwards.
The recent update to SlopCodeBench that includes Astra and Opus 5.5 (including the insight that higher reasoning levels may not improve the situation) seems to match my intuition. [1]
[1] https://github.com/humanlayer/advanced-context-engineering-f...
I am genuinely curious, at what point did you find yourself comfortable “YOLO merging without looking at the code”?
I struggle to find myself ever feeling comfortable doing this in a production environment at my day job.
Probably around GPT 5.6 Sol.
These days your average Opus 5.5 PR with multiple passes of review by Opus 5.5 and Sol 6.1 agents is much less likely to contain bugs than your pre-AI PR.
But sounds like Opus and GPT are interchangeble for your case.
Both have good models. My workflow these days is usually Opus plans and codes, then Opus and Sol review, finally Opus verifies the review and applies fixes.
Do you have example, open source you make using non-interchangeable model, that you can not make using other model?
That wasn't my point, my point was that open models can technically do everything but they make mistakes more often. As an example imagine Opus makes a mistake every 5 prompts and Deepseek every 20. Because you still have a human in the loop, you tell it to fix that mistake and it does, but it's time wasted or bugs sneaking up on you.
Having said that, I was working on an AI for a certain video game, and I had Astra 6 and Sol 6/6.1 work for 30 hours straight, with some human steering to get it unstuck every now and then. It could not get it to work right, and the code it created was a mess.
When Opus 5.5 came out in like 4 hours it cracked it.
I didn't try but I'm very confident if Astra failed, the open models would've failed too.
Less time wasted or bugs sneaking are also good example, can you share video game AI example?
Sorry it's proprietary but imagine a bot for a multiplayer videogame that can replace real players who disconnect mid game.
To clarify, i use both subscription pricing (personal use) and api pricing (at work). My spend going down was referring to api pricing, while using it more (for "everything").
Personal efficiency is noted as never hitting my limit despite it doing huge amounts of work at a time (ex: rewrite my notes app in swift for ios and install on my phone, 2 hours later its opened on my phone, looks nice, is working (base set of features). No follow up prompts, around 2 hours of work. 45% of my 5-hour limit.
Work efficiency is spend, task complexity, code quality (how often i have to steer - rare now), and wiring it up to more stuff (ex having it setup metrics and dashboards etc etc).
I used to sing the same tune as you are now- "models are interchangeable". For today, all other models seems significantly dumber and waste more of my time than Opus 5.5.
tl;dr: You should try Opus 5.5 and report back.
I have been using Opus 5.5 (as well as pretty much all frontier models). I am not sure why you think I haven't.
It hinges on what you consider "common engineering tasks", this definition may not be shared amongst the audience.
The history of tech is one where “good enough” beats “best in class” nearly every time. Anthropic’s whole business plan is that for reasons they can’t explain this time “best” is what matters.
All indications are the models are increasingly worth less and less and rapidly approaching zero. Even worse, everyone is rapidly shifting to “router” architecture where they switch models constantly. LLMs are perhaps the least sticky product in the history of tech.
The only money to be made is selling compute to run models on. The only real out I see for Anthropic is if somehow compute becomes so constrained that they can flip to sell all their committed compute at a markup to run commodity models. Absent something like that it would appear they don’t have a viable business plan. As many have been pointing out, they casually left Q3 off their charts. Put that back in and suddenly the charts look very worrisome.
You're admitting the problem here, anthropic needs revenue to keep growing and growing, yet you're already settled into a willingness to pay of only $20/month as probably a fairly ideal customer, bigger better models just aren't valuable to most people at the price they need to be to make the money back.
It is estimated that only 2% of the global workforce is paying for AI right now. If that grows to double digits - especially in the 30%+ range - and Anthropic captures a significant proportion of it, their valuation is justified. This is assuming they don’t offer virtual employees and these are 100% in the pipeline.
Sort of. If they can rely on differentiation and loyalty now, pre-IPO, from their large-but-cheap customer base, they can focus their growth on serving enterprise/other markets, and lean on their existing base as a low-churn-risk (compared to the rest of the AI ecosystem) customer and marketing surplus. Focus is good.
Anthropic does not make money if everyone is like you, spending $20/m on their SOTA model and are happy. That sounds very commoditized to me.
An example is I like Claude a lot, but for the same $20 spent with Google to get Argon 4, I also get ad-free on YouTube, a ton more storage, and it can access all my Google Drive files natively. I’m not going to pay for both. And I already stopped paying ChatGPT last Nov when Claude got good.
Make sure you test egress so you're not stuck paying a premium price for storage if you ever want to switch off. Having too much in iCloud was a major pain point when I wanted to downgrade my subscription.
Ngl, I've been on Opus 5.5 but had to switch to open for a while. DSv4/Minimax 3.1 were pretty damn good/comparable. Definitely not the same but completely servicable.
If some later version of Deepseek had the capability as Opus 5.5, would you use Opus?
What many of us experienced this year is that Fable 5 and Opus 5.5 are good enough and future model improvements may have diminishing returns even if benchmarks improve.
I’m very excited for smarter models but what I really want now is not capability but speed, cost and open weights.
There will be a version of Deepseek with the capabilities of opus 5.5, but by that time there will be Opus 6 and it will likely cost the same or cheaper than Opus 5.5 and have nearly unlimited usage for subscribers.
Every model release people say the model is good enough, yet Opus and Sol are still more popular than Deepseek. The frontier is shortening, but there are still considerable room for models to improve. SlopCodeBench is new and not close to being saturated.
this conversation happens over and over every 6 weeks. It is absolutely a commoditized product. The models are getting better but they're _all_ getting better. GPT beats Claude beats Gemini. Then open models distill them at cheaper prices, etc.
This is great for us, but that behavior is the behavior of a commoditized product.
Right, the thing is, for most tasks, there’s a finish line. Once a model can do a given task right, future improvements may be of little to no benefit beyond cost reduction.
Right now OpenAI and Claude (and maybe Gemini) are over the finish line with reasonable prices for many tasks whereas the rest are still in the race.
But what happens when the open weight models cross the line and it just comes down to price?
I do a similar thing with 6.1 Sol. I find it pretty comparable to Opus 5.5 in many ways.
> My $20 sub no longer runs out, even though I do _much_ more than ever.
Somewhere in there is a definition of commoditization.
Eh, we have cursor at work with most of the top models and I switch regularly. I went from Claude to sol 5.6 and preferred its prose. Then my work took away the frontier models because they want to see if we can get by on cheaper models. I using Gemini Astra now because it’s discounted and it’s 90% of sol level and completely functional for planning and coding. Some coworkers are using muse coding, some Kimi and glm. Everyone is still producing code. I don’t think it’s a real moat as in you don’t have other options or switching models is actually a difficult process
> I using Gemini Astra now
Probably meant Gemini Argon
At $20/month you're the reason they're 80B in the red lol
"Yahoo! is the best search engine, and I cant imagine anyone making anything better."
Wouldn't a better analogy be Google, which would actually contradict the point you are trying to make?
If you'd ever worked for Google, you'd know how insanely you're missing the point. There have been probably over 1000 self disrupting changes to google search as it continued to stay on top.
I take it you have - i guess can you provide some examples and how they would materially differ from what Anthropic would do then? It feels very similar on the surface.
They could. Let’s see if they do. I’m just saying it ytd if they are yahoo or google.