WASDx

It's like fast food vs healthy food.

Sounds like a hypothesis to me.

I think the top models on that one have minimum goodwill and rejects refunds.

AA published their article on Carbon now at the same time so it's absolutely coordinated.

AA updated their index version yesterday. For sure Google waited for that update so they could be #1.

1. "Building on Gemini 4 Argon". 2. Same price. 3. Relatively short time between releases.

Makes me think it's further post-training of the same model.

It's broken for me in Firefox, worked with Chrome.

Their benchmark result table is broken in Firefox.

Great benchmark!

I believe effort can mean both "solve harder problems" and "produce more stuff". I mostly don't want more stuff than I asked for, so a capable model at lower effort can indeed be the wiser choice. Faster response too.

I believe it's Astra quantized. It was briefly named astra-lite.

When you see these absurd numbers, they are re-counting cached input for every turn. So a simple tool call when your context size is at 500k counts as another 500k to the sum.

Total Opus 5.5 token usage on OpenRouter last week is 5000B, presumably not counting cached input.

How do you validate the functions are correct? I did something similar, letting it (mostly deepseek 4.1) translate from assembly to C but it commonly made mistakes, some really hard to discover and fix.

Can you provide any examples for us interested?

And "system one".

Specifically, it happened over a year ago. Jev just hyped it up.

https://laya.convaiinnovations.com/

The new OpenAI Decisions API calls it "predicate". Also calling the API "decisions" rather than "system one". Usually I don't like inventing new standards but I hope the OpenAI schema takes over. We don't need this hype terminology.

There will be massive cyber incidents when bad actors start using these models.

The model is only 0.3B params so probably not much.

I tried calculating historical "intelligence per cost" recently but stopped when I realized intelligence is not linear. For any meaningful "x per y" you can just double "y" if you have a half as efficient system to get the same result but so-called intelligence doesn't work like that.

It works if you neglect code quality and understanding. I've done a few vibe coded projects on my free time but I would feel shame for submitting that kind of code at work where I use LLMs more responsibly.

Likely because it uses fewer thinking tokens (that you don't see anyways).

Give a task you have to 3 different models and see what actually works for you. There are no good benchmarks.

This is like saying mass production is "cheating" against handcraft.

Once it figures out a puzzle it could probably be instructed to design a specialized harness for Luna to be able to solve other instances of the same puzzle. Minimum wage workers are not solving novel problems.

They explain it here: https://openai.com/index/how-two-settings-tripled-our-arc-ag...

TLDR: The official ARC harness throws away old context and reasoning. No real-world harness is this bad, the model has to re-learn the game repeatedly. OpenAI basically just added standard compaction. Their harness is still "general".

With the contributor pricing being more than 10x cheaper than the standard, that would make it best and cheapest on the DeepSWE leaderboard! It feels fast in my experience too. LLMs keep improving at an insane pace.

I'm party using 1.2 to reverse engineer and re-implement an old game binary and it has been quite good and fast. The contributor pricing is very attractive, excited to try 1.3 and see if I feel a difference. 1.2 can get stuck outputting similar sounding thought summaries with no apparent progress when asked to solve bugs. Then I've switched to GLM-5.3-Flash which for this use case has been clearly better at finding suspected causes and following tracks.

3.7 high and 3.8 medium are essentially the same on AA intelligence and cost. Output tokens on DeepSWE gives the same picture. So there might be something to it but they have done other things as well. At least the tokens are really fast.

If their platform allows any kind of processing as it sounds then anyone could just dump the data. So I don't see why they would not allow downloads for local processing.

Then use other providers hosting open models. Companies and individuals already put their whole code base on the cloud. I'm genuinely interested in privacy-oriented use cases where ZDR is not enough.