If Anthropic indeed nerfed Opus 5.5 on November 3, it could be because they wanted to be prepared for inference cost going up due to the data center moratoriums and the election results. I really hope this is false, but I don't see how inference cost can go down in the near future.
1) it was certain for a long time before the election that the unwholesome party is going to get slaughtered
2) the wholesome party is not in power yet
3) even when they are, the orange man will veto everything they might attempt to do
4) even when the orange man is replaced with someone wholesome, and that wholesome person enacts a bunch of wholesome legislations, data centers will simply move literally anywhere else on Earth.
Regarding 4, I don't think anyone in the US, regardless of how they voted, would want a bunch of SOTA GPUs being in any other country than the US. It's either they are all here in the US, or they are not anywhere. As a result, if data centers cannot be built in the US, the GPUs will not be produced only to be exported elsewhere.
>In Ohio alone, at least 14 municipalities passed bans or restrictions on datacenter development, including in Sunbury, where Amazon has proposed a $2bn datacenter project. In Maine, eight municipalities passed temporary datacenter moratoriums. In Emporia, Kansas, where a 1,000-acre gigawatt-scale datacenter was recently proposed, a ban on high-impact datacenters was approved by voters.
and now they will be built elsewhere.
I think there's a rather obvious precedent for that inevitable outcome.
I mean, sure, they'll be built somewhere else (and I think the approach of "stopping" rather than "protecting residents from price changes" is terrible), but buying land, permitting, etc take years and tons of money. It's going to slow down the industry a bit.
They are looking to lower it further to $100 [0]. This is definitely concerning for frontier labs since Accenture also caps their employees' usage to $29 a month [1].
Yes. The frontier models are good, indeed, but they are just far too expensive to use much. Recently I was showing my executive leader that we could simply run a certain set of tasks on the top frontier model and we'd be good to go. Oh but it would cost over 15M/year. We're a relatively small company, that's over two magnitudes more than what makes sense for us. So, never going to happen.
What remains then is an optimization exercise to bring the cost down. For months now we've been optimizing usage so we can run nearly all tasks on very cheap models to good success. It's fun to optimize and keeps the CFO from having a heart attack.
But not good news for the AI companies relying in a firehose of cash coming in from their frontier models.
It's kinda like commuting in the autobahn with a Porsche.
Some agentic madness is like doing donuts in the parking lot drinking beer impressing the local baddies.
So, so much fun. But at the end of the day we used to get so much more done with much, much less computer power, energy, mental drain.
It's always more software tooling, more moving parts, more apps, more more more. Fun. Looks neat. Used sooo much energy, I feel so drained. It was fun.
But at the end of the day, for the long run what I want in my life is to win at rally on a dependable machine with a nice navigator I'm friends with. Or have a shed full of tractors making money for my family on the farm everyday.
LLMs are fun. But I want to use my mind to write great software that adds value in as little lines of code as possible designing systems that people can understand readily and maintain well 2 or 3 handovers from me.
$100 day should be more than enough. To exceed that you have to coming up with creative ways to waste it. You couldn’t hit that in normal development using AI all day.
For the edge cases that might need more like running a security audit, it would make more sense to submit that as a project proposal and have the expense approved.
I don’t agree. You should be empowering your engineers. A higher budget allows them to:
- freely experiment and understand the limitations of AI
- create things outside of what they are officially expected to
I’m not advocating for people who burn tokens for the sake of it. But I do think a single engineer has more potential for impact today than ever before and a high budget unlocks that.
It also isn’t _that_ much more when a fully loaded mid-level engineer is costing you $300k+
I find it hard to use that much in a day. Especially if you delegate to a cheaper model for anything context hungry like reading large amounts of data etc. if you’re doing tasks that cost $1000 a day I feel like there’s some software engineering or script intermediate states to be done to optimize that work than letting the model just go ham
Token-maxxing aside, what's the business value for something like this? Slide-deck generators have been around for a long time already. And at least for me, the LLM-authored decks I've created have always needed tweaking / editing to really make sure the content is useful and correct.
At my company at least, I think our developers waste a ton of labor and money on getting LLMs to produce tools like this. Then they use tool a few times and ultimately forget about it or throw it away. At the end, they've spent a few hundred dollars on frontier tokens for the effort.
More time gets spent tweaking/iterating with the LLM than it would take to make the decks by hand. And sharing takes effort, so our devs mostly roll their own tools with minimal reuse. At the end there's no business value to show for it other than the slide decks that could have been made the old-fashioned way.
> what's the business value for something like this?
We already generated slide decks from a template. The decks generated were consistently missing data, incorrectly formatted, and required many manual edits.
This led to hiring contractors to fix up the decks before being sent internally before being presented to clients.
So, thousands of contractor hours every year going towards fixing up decks in additional to the engineering effort from ~5 bugs each month due to bugs in the deck or feature gaps.
> Then they use tool a few times and ultimately forget about it or throw it away.
This is a running production service
> More time gets spent tweaking/iterating with the LLM than it would take to make the decks by hand.
I think you misunderstood. The LLM produced deterministic deck templates in code that is then templated from a payload and the deck dynamically changes depending on what is in the payload. The user can then further modify the deck as they please since the deck is just code. Then they can export it to a PPTX and make any further adjustment in their favorite tool since it emits a "normal" PPTX.
Techie types seem particularly susceptible to spending massive effort re-inventing things that have already existed for a while.
See also, Matt Levine’s recurring joke about the crypto community constantly re-inventing/re-discovering the need for financial regulations that the larger finance industry have had for a while.
We don't know how low it will get yet. The first article I linked also has this: "[m]any groups outside of engineering already face a $100 daily cap, but Apple has considered reducing that limit to as little as $25 for some teams, the people said."
They could be tightening the purse strings further until they hit some magical ROI numbers.
> $100 day should be more than enough. To exceed that you have to coming up with creative ways to waste it.
You do not need much creativity for that. Just forget to put build artifacts or node_modules in your agent's ignore list and it will happily burn through the cap in thirty minutes
Why is it "concerning"? Personally, I get a lot of stuff done on the $200/month subs. +$100/day per employee seems like a fairly significant amount of money and if a lot of companies do that (even a lot less than that) that seems entirely fine for frontier labs, given all the other uses as well.
That is not to say that you can't spend a lot more. And I would be surprised if Apple did not allow for that too. Note that the text says "for some of its own engineers", which could amount to a total nothing-burger, given the overall vagueness.
Depends on the sub you're using, but the anthropic one is giving you ~$12k/month [0], or $400/day. So you're already working at 4x the $100/day you consider significant.
Accenture lol, in that case, considering the quality of the output previously released by their humans, maybe local LLMs are already more than enough for that use case.
For smaller companies they can get Claude team plan for $20/month/seat and the plans give a lot of usage. We rarely hit limits. Codex plans used to have more usage and those were the go-to. Before that, GitHub Copilot CLI was the service with insane usage limits for $10-$20/month.
So it's about staying nimble enough so you can easily switch over. Don't tie yourself down to a single product/harness.
The Claude Team plan is subsidized, and pricing for them is significantly less than what enterprises pay with API pricing. For example, your $20/month plan would cost the equivalent of $400/month if they charged you using API pricing.
That's why those plans are only available to small companies with 150 employees or less.
> That's why those plans are only available to small companies with 150 employees or less.
As far as I can tell, that's not the limit. A team plan may not have more than 150 members but the size of the company doesn't matter. I you have 150 people who need to use it inside an organization of 200, 500, 5,000, whatever, that's fine.
Accenture isn't a giant software company though. For the sort of things their employees do, that might be a reasonable average (think chat, PowerPoint, a bit of research).
Yes, considering their CEO said that AI usage was a requirement for promotion earlier this year [0]. Companies are backpedaling now when the AI bills come out and bite them.
I switch harnesses regularly depending on the model providers, and I only use skills which are universally usable in any model and harness combination. I don't want to get tied to a particular model or harness, and skills provide me the freedom to do so.
I can include any custom programs I need in a skill, and the SKILL.md file is just a doc of how to use those. It helps me save tokens and ensures determinism with repeated workflows since redoing a task is just calling a skill's program.
I also load only skills that are necessary for a project. I don't want to pollute the context window with unneeded skills.
In a few years we'll be able to look back and compare the technical debt in code written since 2026 vs code written in the 40 years prior. We'll then have another metric to quantify the impact and usefulness of AI.
If AI's code is sloppy but bug-free, is it better than the alternative?
I'm not sure if I missed something, but the benchmarks [0] look outdated since they use Claude Opus 5 (instead of Opus 5.5), GPT 5.6 Sol (instead of 6.1 Sol), and GLM 5.2 (instead of DeepSeek V4.1 Flash 1017, which is the most popular model on OpenRouter currently [1]). I also cannot find any information regarding thinking efforts for each.
Since these benchmarks were done in July, they were probably using older versions of Claude Code, Codex, and OpenCode as well.
Would appreciate if the benchmarks are updated, and the thinking efforts and the versions of Claude Code, Codex, and OpenCode are specified.
I remember when Opus 5.5 came out, we had a discussion about whether it was nerfed a week later [1]. It's not possible to do so with Gemini 4 models since none has been made available the day it is announced.
Without us being able to access Carbon when it is announced, there's no way to know the Carbon we get a few weeks later is the same thing as what has been benchmarked here. I think benchmarks should verify what is benchmarked is exactly the same as what's been released.
I used to land on SO questions and answers through Google. Nowadays I ask ChatGPT or Claude instead of Google, and then go to the links they give me to verify. The links are either to the official docs or some blog posts. Never have I seen LLMs recommend any SO links.
True. And if Carbon (6) is better than Argon (18), wouldn't that make Hydrogen (1) the ultimate Gemini model, or will the next model jump to another random element?
The problem is that these names are random and never-ending. They are not categorized like Claude and GPT, both of which have four classes of models each. That's eight classes that I can remember.
There is no way I will be able to (nor do I want to) remember more than four names just for Google or any other single provider, and there's no value in doing so for me. There's nothing unique or distinguished that Gemini does that neither Claude nor GPT can.
reply