They seem to be pushing users away from using OpenAI and Anthropic models. On March 16th, they released pricing changes to stop subsidizing the state of the art models for team/enterprise users on legacy pricing models.
GPT 5.4 and Claude Opus/Sonnet 4.5/4.6 are now billed at API rates for all users, even enterprise customers. Previously, they were subsidizing these models around a factor of 10x, billing per request and not per token. Composer 2 bills for $0.08 per request on the fast model, and $0.04 per request on the slower model - no matter the tokens used.
It seems like they are targeting Enterprise above all else, relying on a enterprises signing up a bunch of paying users that rarely touch Cursor to subsidize the power users using an excessive amount of Composer tokens. It's a fair strategy, as cursor seems to increase output by 20-30% so the price is well worth it for Enterprise customers.
To be fair these frontier models have been seriously increasing their pricing as of late. Opus 4.6 requests regularly cost over $5 now, with average requests costing ~$1-2. If Composer is benchmarking better than Opus and costs $0.08 per request that's a win for everyone.
I know people like to hate Composer but competition is a benefit to all of us, and I don't doubt Composer will take it's own chunk of the consumer market.
I work at a large company that is contracted to build warehouses that automate the movement of goods with conveyors, retrieval systems, etc.
This is a key candidates to use AI as we have built hundreds of warehouses in the past. We have a standard product that spans over a hundred thousand lines of code to build upon. Still, we rely on copying code from previous projects if features have been implemented before. We have stopped investing in the product to migrate everything to microservices, for some reason, so this code copying is increasingly common as projects keep getting more complex.
Teams to implement warehouses are generally around eight developers. We are given a design spec to implement, which usually spans a few hundred pages.
AI has over doubled the speed at which I can write backend code. We've done the same task so many times before with previous warehouses, that we have a gold mine of patterns that AI can pick up on if we have a folder of previous projects that it can read. I also feel that the code I write is higher quality, though I have to think more about the design as previously I would realize something wouldn't work whilst writing the code. At GWT though, it's hopeless as there's almost no public GWT projects to train an AI on. It's also very helpful in tracing logs and debugging.
We use Cursor. I was able to use $1,300 tokens worth of Claude Opus 4.6 for a cost of $100 to the company. Sadly, Cursor discontinued it's legacy pricing model due to it being unsustainable, so only the non-frontier models are priced low enough to consistently use. I'm not sure what I'm going to do when this new pricing model takes affect tomorrow, I guess I will have to go back to writing code by hand or figure out how to use models like Gemini 3.1. GPT models also write decent code, but they are always so paranoid and strictly follow prompts to their own detriment. Gemini just feels unstable and inconsistent, though it does write higher quality code.
I'm not being paid any more for doubling my output, so it's not the end of the world if I have to go back to writing code by hand.
I got so tired of cursor that I started writing down every bug I encountered. The list is currently at 30 entries, some of them major bugs such as pressing "apply" on changes not actually applying changes or models getting stuck in infinite loops and burning 50 million tokens.
I've spent $17.64 on on-demand usage in cursor with an estimated API cost of $350, mostly using Claude Opus 4.5. Some of this is skewed since subagents use a cheaper model, but even with subagents, the costs are 10x off the public API costs. Either the enterprise on-demand usage gets subsidized, API costs are 10x higher, or cursor is only billing their 10% surplus to cover their costs of indexing and such.
edit: My $40/month subscription used $662 worth of API credits.
oh, I figured out the costs for the enterprise plan. It's $0.04 per request, I'm not charged per token at all. The billing is completely different for enterprise users than regular users.
There was a promo from Straight Talk wireless, that a bunch of users on slickdeals used, for a $360 iphone 16e (+1 month service) with the intention of buying and unlocking to be used on another carrier. The FCC guidelines were explicit about being 60 days after activation without indication of fraud, with no mention of active service.
After the first few initial customers put in tickets to unlock their phones after 60 days passed, Straight Talk changed their policy from 60 days since activation to 60 days of active service, breaking the FCC guidelines knowing that no one would sue them in a federal court over a small amount. They forced users to buy a second month of service to unlock the phone. One user even successfully won in small claims court for breach of contract since Straight Talk refused to activate their phone, since you can't just change the contract after the sale is complete. You sadly can't sue for breaching FCC policies in small claims, that requires hundreds of thousands of dollars worth of lawyers. I put in a FCC complaint over this, but the FCC more or less ignored it.
Verizon is just doing this as a pretext. It's a continuation of them ignoring this policy after users were buying cheap phones to use on other carriers and waiting 60 days. It just looks better to claim you are defeating criminals.
The pyramid being upside down with grain on the bottom and fats and oils being on top is directly from south park.
The only difference is that meat, fat, dairy, fruits and vegetables are grouped together with this new pyramid with grains on the bottom. while south park puts fats -> meat and dairy -> fruits and vegetables -> grains as the order.
I find Cursor's tab completion to be distracting enough with multi-line changes that I just disabled it, while I use IntelliJ's tab completion regularly.
Cursor's tab completion is better, but it doesn't seem to have a concept of not trying to tab complete. IntelliJ is correct half the time for completing the rest of the line and only suggests when it is somewhat confident in its answer.
I agree about the multi-line blocks Cursor proposes. Like it gets the first two lines right and then after that it's nonsense. I'd rather it stuck with a single line change at a time, and let me press enter before it predicts again.
from the author of this patch on discord - the level 9 for compression isn't practical and is too slow for a real production server but it does show the effectiveness of zstd with a shared dictionary.
So you start off with a 755.2 MiB world (in this test, it is a section of an existing DEFLATE-compressed world that has been lived in for a while). If you recreate its regions it will compact it down to 695.1 MiB
You set region-file-compression=lz4 and run --recreateRegionFiles and it turns into a 998.9 MiB world. Makes sense, worse compression ratios but less CPU is what mojang documented in the changelog. Neat, but I'm confused as to what the benefits are as I/O increasingly becomes the more constrained thing nowadays. This is just a brief detour from what I'm really trying to test
You set region-file-compression=none and it turns into a 3583.0 MiB world. The largest region file in this sample was 57 MiB
Now, you take this world, and compress each of the region files individually using zstd -9, so that the region files are now .mca.zst files. And you get a world that is 390.2 MiB
Author here -- the solution I discussed in that message isn't quite the same solution as the one linked. The `paper-zstd` repository is the one using dictionary compression on individual chunks. But in the `.mca.zst` solution I'm not using dictionaries at all. It's more like a glorified LinearPaper -- just take the region file, decompress the individual chunks, and recompress the entire 1024 chunk container together. It breaks random access to individual chunks, but it's great for archival or cloud storage offloading of infrequently visited parts of a MC world, which is what I'm using it for.
I don't remember the exact compression ratios for the dictionary solution in that repo, but it wasn't quite as impressive (IIRC around a 5% reduction compared to non-dictionary zstd at the same level). And the padding inherent to the region format takes away a lot of the ratio benefit right off the bat, though it may have worked better in conjunction with the PaperMC SectorFile proposal, which has less padding, or by rewriting the storage using some sort of LSM tree library that performs well at compactly storing blobs of varying size. I've dropped the dictionary idea for now, but it definitely could be useful. More research is needed.
> You set region-file-compression=lz4 and run --recreateRegionFiles and it turns into a 998.9 MiB world. Makes sense, worse compression ratios but less CPU is what mojang documented in the changelog. Neat, but I'm confused as to what the benefits are as I/O increasingly becomes the more constrained thing nowadays. This is just a brief detour from what I'm really trying to test
Might make sense if the region files are on a fast SSD and the server is more CPU-constrained? I assume the server reads from and writes to the region files during activity, a 3.5x increase in IO throughput at very little CPU cost (both ways) is pretty attractive. IIRC at lower compression levels deflate is about an order of magnitude more expensive than lz4.
zstd --fast is also quite attractive, but I'm always confused as to what the level of parallelism is in benchmarks, as zstd is multithreaded by default and benchmarks tend to show wallclock rather than CPU seconds.
> Might make sense if the region files are on a fast SSD and the server is more CPU-constrained?
I wrote that when the feature had just come out. Now it's been a bit since Minecraft started natively supporting the LZ4 chunk compression option. It seems safe to say that this tradeoff does in fact make sense, even when the CPU is quite powerful. Several servers have adopted it and have seen decent improvements.
Note that each region file contains 1024 chunks that are designed to be (but probably aren't) accessed at random, so compressing a region file is like a solid archive with a solid block size of 1024 files.
GPT 5.4 and Claude Opus/Sonnet 4.5/4.6 are now billed at API rates for all users, even enterprise customers. Previously, they were subsidizing these models around a factor of 10x, billing per request and not per token. Composer 2 bills for $0.08 per request on the fast model, and $0.04 per request on the slower model - no matter the tokens used.
It seems like they are targeting Enterprise above all else, relying on a enterprises signing up a bunch of paying users that rarely touch Cursor to subsidize the power users using an excessive amount of Composer tokens. It's a fair strategy, as cursor seems to increase output by 20-30% so the price is well worth it for Enterprise customers.
reply