SyneRyder

Huh. They're not "Pelicanmaxxing"... they're Ottermaxxing.

Take a look at the GLM 5.2 and Deepseek V4 "animal on a plane" examples. In every case, the animal is standing on top of the plane, a clear misunderstanding of the concept of "animal on a plane"... with the exception of the Otter. The otters are sitting in a seat on the plane, looking out the window.

That's Ethan Mollick's "Otter On A Plane Using WiFi" image benchmark.

https://www.oneusefulthing.org/p/the-recent-history-of-ai-in...

(Sometimes the Racoon is sitting inside the plane as well, but the racoon is a common backup benchmark. I'm surprised it wasn't also holding a sign saying that it loves trash.)

Also, Grok seemed to really really enjoy "whale on a plane" in that second round, and kudos to GPT Terra for deciding after 3 rounds that the user was terrible at spelling and generated "Antelope On A Plain".

EDIT: I promise I'm a human, but I did just notice my "that's not x... that's y" construction at the start. I am rather Claudepilled — my apologies.

In case it isn't clear, the wall here is a representation of the actual wall at the Visitor Center at the Cornell Lab Of Ornithology. They also have a live webcam of their bird feeder, that is literally just around the corner from this wall:

https://www.youtube.com/watch?v=x10vL6_47Dw

It's a really pretty place to visit, with 5 miles of hiking trails around it. If you get a chance to visit Ithaca, it's worth taking the time to visit:

https://www.birds.cornell.edu/home/

That's the profit margin on inference, not overall. Each model does end up being profitable over its lifetime, but the money they're making is being immediately churned into buying more data centers & the training for the next giant model up, so they're not profitable overall at the moment. It's a bit like how Amazon kept churning their profits into more growth instead of taking the profit early.

That said, Anthropic has supposedly crossed over into profitability and made $1 Billion in profit so far this year, in the lead up to their IPO. Being profitable sounds good for launching on the stock market! But as a customer, that's noticeable in the downtime due to lack of compute, and only getting 50% access to Fable.

OpenAI might not be profitable, but they've got so much compute access that they've been able to give their customers full access to Sol, and as a result they've almost doubled their Codex subscriber base in the last two weeks (6 million on July 12, 10 million on July 21 - that would be an extra $1-$10 Billion in Annual Recurring Revenue that they've gained in just these 2 weeks). Doing the unprofitable thing in the short term can result in outsized rewards in the long term.

This actually looked excellent, except it seems in my case to have hallucinated much of the information and links - every "receipt" I clicked on resulted in a 404.

But a website that can tell you if your business idea has been done before, and the reasons the previous companies that tried went out of business... that's actually a great idea. Maybe I just got unlucky with my search.

"As of October [2025], OpenAI's compute margins reached 70%, up from 52% at the end of 2024 and double the rate in January 2024, [The Information] said, citing a person familiar with the figures."

https://www.bloomberg.com/news/articles/2025-12-21/openai-se...

As for Anthropic, the rumors I remember seeing for their API margins were more like 85-90%, but I don't have a reference at hand for those. But once you know the API is wildly profitable and the subscriptions are roughly break-even and not even a big slice of their income, all of the investment makes a lot more sense.

This is my experience as well. Lots of success with running one hour Fable sessions (one hour of API time). Not using a loop here, but there's often lots of documentation for Claude to go through first before building out the project.

The main failure point for me with long running Fable sessions now is just that it might hit a safety guardrail and downgrade to Opus midway.

If you care about availability for work you can just pay more where it feels limitless and never look at the limits...

Depending on your work, it doesn't feel limitless - I've just upgraded to a $200 plan and I'm starting to see where I will even hit the edges of that, especially once resets and special offers start tapering off. But with a single AI question often costing $75 in API costs, I'm not at a point where I can just switch to API and not care about cost at all.

That means I need to carefully time the work to make sure I'm squeezing what I can into the 5 hour windows. Feels a bit like the days of mainframes and time slicing, when you had to book an allotted time for your workload to run.

https://www.cs.cornell.edu/wya/AcademicComputing/text/earlyt...

I think those of us who are using AI consistently believe you and understand. I'd say roughly the same thing about Claude in terms of numbers.

I think many people who don't believe you just haven't built-up the kind of prompt history & MCP / CLI tooling etc that lets you get to the point where things work at that level of accuracy.

Hope it helps to know that at least some of us here understand and are seeing the same thing. And if it's anything like my experience with Fable, "always be more ambitious". The capabilities of the models are often limited only by what you're brave enough to ask for. I keep finding I'm not ambitious enough.

I don't use Codex CLI myself, but you can configure it to point to OpenRouter instead. OpenRouter has some instructions for Codex CLI and Claude Code here (though they mention Claude Code is not guaranteed to work!):

https://openrouter.ai/docs/cookbook/coding-agents/codex-cli

https://openrouter.ai/docs/cookbook/coding-agents/claude-cod...

Javascript source code appears to be here:

https://om-intelligence.ch/projects/vocal-notation/vocal-not...

The comments mention it is using autocorrelation ACF2+ for pitch detection, and Krumhansl-Schmuckler for key detection. (Which doesn't mean very much to me yet, but is giving me some keywords to start researching.)

I was expecting YIN since that's the first thing Claude reaches for, and the code comments in the source do contain a malformed em-dash....

I'm pretty sure you're not the only one. My relief was less about having another week of Fable access, and more "thank goodness I might get a normal night of sleep tonight".

Anthropic have extended Fable access again to July 19. The notice should pop up in your Claude Code now when you start a new session (also announced on the ClaudeDevs X account first).

I don't think I am — but I am rather Claude-pilled, and graphics algorithms are right in my wheelhouse.

I look forward to being able to see C-beams glitter in the dark near Tannhäuser Gate.

I've just tried it on my large desktop monitor (roughly 1440p, not HiDPI), and I now see "Ghost Font" extremely clearly and can't see the decoy at all. If I scale my browser window to 30% zoom, then I can just see the "Written In Ghost Text" decoy message again.

My phone would have been zooming out the browser window, and making the dots even tinier, but the phone is HiDPI so it would have still preserved the dots. My eyes are middle-aged and probably starting to do the same kind of median-blur effect that models do when they resize an image. That's my current guess for why I can see the decoy more clearly on mobile.

If that's the case, then this trick will stop working as vision models approach pixel-perfect vision, instead of the current resizing that they do. Pretty cool as steganography though.

Took me a long time to realise that "Written In Ghost Text" wasn't actually the text I was meant to be reading, and that was only the decoy message.

I can barely read the actual message, and it's about as "readable" to me as the Magic Eye 3D pictures. Actually I think I have a headache from looking at it on a mobile screen.

As a research idea it's cool though. But I do wonder if/when AI models will figure out how to decode it - I imagine a bit of additional prompting would get them there.

... make it a lot more expensive to use for multi-turn coding than many would assume from the $2/$6 headline numbers they led with.

There's a further sting in the tail, Grok 4.5 is only $2/$6 for the first 200k of context. Go above that, and the pricing is $6 / $12 - and you're still capped at only 500k context anyway.

Here's the xAI pricing on OpenRouter:

https://openrouter.ai/x-ai/grok-4.5?endpoint=0e927811-b1a8-4...

Similar story here. They took my ~$100/yr Harvest time-tracking Solo plan, increased the price by 2.5x for a more restricted plan than I had... or I could get back the plan I had for $20,000/year.

So I downloaded my data, and had Claude vibecode a fully-featured clone in a single evening. Even if I was paying Anthropic API rates, it cost me less than a single year of my Solo plan.

Are you familiar with the Rosie the dog story?

AI designs cancer vaccine for dog but scientists says red tape a barrier for human care

https://www.abc.net.au/news/2026-06-22/australian-dog-cancer...

People have already been using ChatGPT to design custom bespoke mRNA vaccines specifically for one patient, based on sequencing their specific cancer. It already works to reduce tumours. Sam and Dario know this, it's why they can make their claims - it's already done. The problem is the cost of the procedure (which is why only rich entrepreneurs are seeing their cancers treated this way so far) and government regulations preventing its use in wider human populations without a 10 year study first.

Seems like you're asking for the Artificial Analysis "Intelligence vs Cost" benchmark, perhaps?

https://artificialanalysis.ai/?cost=intelligence-vs-cost-per...

Wow, this was worth watching more than I expected it to be. Not because of anything Karp is saying - okay, so Palantir thinks they're going to make their own frontier model by fine tuning Nemotron, good luck with that.

Watch the body language. Hear the tone of voice. He's scared. He's not arguing from a position of strength.

I find the implications that follow from that fascinating. I'll let others draw their own conclusions though.

My understanding of the rumor is that it wasn't OpenAI itself, but one of the post-blip OpenAI breakaway groups (rumoured to be Thinking Machines) who have made a breakthrough and seem to be shopping it to OpenAI. I don't think it has actually been implemented by OpenAI yet.

I did a quick look for 200k models on OpenRouter. There's a lot of previous-gen Minimax 2.5 & 2.7, GLM 5.1 that are around 200k.

But also on the list at 200k are "Free Models Router" and "Claude Haiku 4.5". I would not recommend making any judgment of AI based on free models. And coding with Haiku is a bad idea... I mean, that was my first code AI test too, but it's just not an accurate impression.

To be fair, Opus 4.1 & 4.5 are also listed as 200k. They did require context management for large & difficult tasks. But if you do have access to Opus, there's very little reason not to switch to 4.8 / Sonnet 1 Million now. I wouldn't recommend Sonnet, but I have used it to write a USB audio driver that got some hardware working on an obscure OS, so it can work.

Problem 3: "you'll hit the 200k token limit..." ... Suggestion: use 1 million context window LLMs.

Yes, if the model someone is using only has 200k token limit, that would immediately suggest to me that it really isn't a sophisticated enough model.

Most of my coding sessions end up being about 350k tokens long when I finish, it wouldn't even fit in a 200k context. And that isn't counting the cache-reads by subagents, etc.

It's worth spending some time with the best Opus / GPT model, to at least get a sense of what the frontier is like.

Thank you, I haven't heard of Cortecs before. Might see if I can integrate this into my harness, or at least wire up Tensorix.

Also, I don't know how accurate that tokens/per/second measure for GLM 5.2 is, but if that is even remotely true, then I won't complain about the mild markup Tensorix have for GLM ;) Thank you for the heads-up!

It's because GLM 5.2 is offered on many inference providers, including providers in the US. Those companies only make their money by charging for inference, and yet they seem to be doing quite well while charging the exact same prices as Z.AI / GLM.

In fact, there's a price war where some of the US inference providers are undercutting the pricing of Z.AI's own GLM hosting. Novita & AtlasCloud are both offering 8% and 5% discounts on GLM 5.2 respectively. GMICloud is charging 30% less - but getting so hammered with demand that it only has 80% uptime & 7 tokens per second, so you get what you pay for.

You can find a list of providers & their pricing through OpenRouter here:

https://openrouter.ai/z-ai/glm-5.2#providers

I do agree in cases where I'm using API and not the subscription, this would be very costly via API. Not sure why the tokens wouldn't be in the cache though? Seems everything should be cached as long as I'm within the 1 hour caching window? If I'm wrong about how token caching works, I'm eager to learn!

My other concern is, it isn't really a 1 Million context window if we can only use the first 500k, right? But now that I've found that I can re-enable it, I'm happy.

I've previously had sessions go to 700k tokens and still be okay, though it does start drifting at that 700k point. I'm regularly at 300k with no problem.

Not only that, but using Opus 4.8 [1m] right now outside the US, and suddenly I only have a 500k context window. I really hope this is just a strange Claude Code bug, but I had access to a 1 Million window before, and it wouldn't entirely surprise me if context window length becomes another US export restriction.

The Anthropic page here seems to say that Max users should have access to the full 1 Million window for 4.8:

https://support.claude.com/en/articles/8606394-how-large-is-...

I was already setting up my infra to experiment with GLM 5.2 and its 1 Million token window before this happened. I think I'm glad I did.

EDIT: Found a solution, seems Claude Code 2.1.193 (or an earlier version I didn't notice) changed default settings, so that if you have Autocompact turned on it occurs at 50% of the context window. If you turn off Autocompact, the full 1 Million context window is restored. Another example of Claude Code quietly changing default settings sigh

There is probably a market for Deepseek/GLM served from non CCP available servers. I might even look into how hard that would be to setup here.

Please do. There is definitely a market for Deepseek / GLM hosted from non-China servers, there's over 20 providers for GLM 5.2 on OpenRouter alone... and they're all either Singapore (home of Z.AI / GLM), China, or US. There is nothing yet listed on OpenRouter from Europe (Inceptron still only has GLM 5.1). And of course, there is absolutely nothing hosted in Australia.

We're in a particularly dire situation in Australia. We're about to be cut off from Claude Fable and premium American models. The European Mistral models are garbage, at least in comparison to US models. Our only hope is going to be Chinese models (GLM 5.2 is good), and we're not even hosting them in Australia.

By the way, if you haven't tried an Anthropic model, it's worth spending at least $20 one month to give Opus 4.8 a try. I only got one night of access to Fable before I was cut off, but one single evening of Fable provided plans that I've been working through for about a week afterwards with Opus 4.8... and that was only Fable, not even Mythos. That's the kind of intelligence lead Australia is about to be cut off from.

(And kudos on the Containers For Change, that's something I do as well - mostly as an exercise incentive to walk to the local recycling machine, because the money certainly doesn't compensate for the time spent on the recycling.)

This caught my eye:

The script ended up outputing code that uses variables outside of their scope, didn't utilize like 90% of the features of the language

Using variables outside of their scope sounds very unusual to me for Claude. You are using Claude Opus (4.5 or higher) and have set the thinking to High or above, right? Make sure you're not using Claude Haiku. Sonnet can be okay, but I'm sure the developers you've heard raving about it are all using Opus 4.8 or GPT 5.5, and all using it from within Claude Code or Codex (or OpenCode or Pi, tools like that anyway).

Claude should catch something like variables being outside of scope immediately when compiling, and fix it as soon as it notices the compiler bug.

The script itself was also written in really weird way, utilizing recursion for pretty much everything when most of what it did could be done in simple loops...

That's actually a great opportunity to develop a new prompt to give to Claude. AI is really good at pattern matching. Take one of those weird recursion methods Claude came up with, then rewrite it as that simple loop that you would prefer, and show both to Claude. Then ask in the same turn: "This is how I prefer to write this code. Can you suggest a prompt to me that would encourage you to write this style of code instead in future?"

See if you can get Claude to reduce that down to a simple maxim or principal you can include in a startup prompt you provide at the start of each session, or into your global CLAUDE.md file that is loaded at the start of every Claude Code session. It might end up being a guideline like "Prefer simple loops over recursion whenever appropriate."

It's possible that the developers you've heard raving about AI have already developed startup prompts / CLAUDE.md files filled with similar maxims & principals, tailored specifically to how they like to code & work, evolved from months of working with AI.

I've had success with vision models & OCR, saved me many hours / days / weeks of typing work.

Last year I finally OCR'd many hundreds of pages of my father's old writings. I found that feeding it to Claude Sonnet 4.x via API gave me results that were perfect. No corrections required. So perfect, that Claude was reading along with the story, and actually pointed out a continuity error in the story where an incorrect character was reciting dialog. Claude asked if it should transcribe exactly as is or if I would like Claude to correct the continuity error.

Claude also correctly OCR'd some handwriting that was in the margins of the documents. Sonnet came very close to transcribing a Word Sleuth puzzle, but that was where I hit the limits of its capability at the time.

Mistral OCR was also good (and actually what I started with), but it wasn't quite as good as Claude. And when it was wrong, Mistral could be frighteningly wrong - one API call must have failed, the model must have been presented with a pure black / null image, and I got back a "transcription" that described neverending darkness. It read like something the Woodsman would have broadcast in Twin Peaks S3E8. That poor model.

Tables from electronics datasheets might be okay, I think I've had success with OCR of technical manuals with tables for 80s synthesizer hardware. But I admit my use cases don't crossover into transcriptions of equations or graphs.