thepasch

Yes, max effort is what the displayed-by-default score on the index is measured by. At xhigh, Opus 5.5 outputs 100M tokens, but remains only 1 point ahead of K3.1 rather than 5. And Sonnet at xhigh outputs 110M tokens and trails K3.1 by 3 points rather than leading by 3, at the same price point per mtok.

Opus 5.5 high trails K3.1 by 1 point, but outputs only 53M tokens, so it indeed comes out cheaper per-task. I think it's still remarkable that this is effectively an open model that matches the frontier for the first time in a long time, pretty much since DeepSeek R1.

> When evaluating the Intelligence Index, it generated 150M tokens, which is very verbose in comparison to the median of 87M.

This is 100M fewer than Opus 5.5, 40M fewer than Fable 5.1, and about a third of Sonnet 5.5. In the top-6, it's 3rd, beaten by Muse Wonder at 140M and Astra at, impressively, just 60M.

Like with many things, it depends entirely on how you define "real terminal"; that is, what a "real terminal" must do before it satisfies your definition. If it's "takes input and deterministically executes its instructions, producing output along with side effects," a well-trained model with argmax sampling could well meet the bar. if given the appropriate external interfaces.

Isn't this going to completely slaughter cache hit rates?

Since long-term foresight beyond the next stock buyback doesn't seem to be materially required anymore for that role anyway, I'd wager they could already be replaced by AI, and with less consequence than any other company roles.

That's super encouraging to hear, thank you for the clarification! I'd edit my original comment, but the edit window has unfortunately run out on it. Should be OK, though, since the help desk page is now explicit about it.

The monthly credit allocation hasn't rolled out yet either, so as of right now, we're effectively at the status quo. I'd expect the billing change to land once you can actually collect your Claude Console account.

...yes, a banner which has also now disappeared and been replaced, with the explicit mention that API credits "cover the Claude Agent SDK"?

What more do you need?

> Claude Max and Team plans now include monthly API credits, which cover the Claude Agent SDK, the Claude API, and Claude Managed Agents.

This text has replaced the entirety of the page called "Use the Claude Agent SDK with your Claude plan."

That's what the Agent SDK is, according to this help page:

https://code.claude.com/docs/en/headless

So, as written, yes.

This is them sneaking in taking the Claude Agent SDK (claude -p) off of subscription plans through the back door along with a model release. They previously wanted to do this in June, but backpedaled after huge backlash:

https://support.claude.com/en/articles/15036540-use-the-clau...

Note that this is Anthropic Trojan-Horsing the previously announced June change in with a model release, where the Claude Agent SDK can no longer be used with Claude subscriptions and is now billed with API credits only.

https://support.claude.com/en/articles/15036540-use-the-clau...

I think Google's AI overview is legitimately doing more to utterly destroy the reputation of AI capability than anything other AI application right now, because I'd guess it's one of the biggest, if not the biggest, public surface for AI outputs and it's BAD. "Bad", as in: it's staggering how much utter nonsensical tripe whatever model they use is still able to vomit out in an era where even tiny models that run on edge compute can give much more helpful advice.

If your feature requires so much compute by sheer mass that you have to nerf it into complete uselessness to keep it economical, then maybe it's not a very good feature. In the meantime, I can confidently say that nobody has lost the "war on quality" as long as that utterly useless "feature" still exists anywhere near the internet. It's simply indefensible, even for the most steadfast proponents.

What proprietary AI running on big corp servers was 6 months ago is now runnable on owned bare metal compute, and there's no reason to believe this trajectory won't continue for the foreseeable future.

Carbon atoms being conscious seems pretty alien to what is known, too. Yet here we are. Maybe what's important isn't the individual thing that makes up a conscious actor, but the holistic whole?

> If by "training corpus" you mean the actual data I already acknowledged that that's currently legally impossible.

Did you read the site this very post links to? The entire point is that the training corpus, recipe, and scripts, as well as intermediate checkpoints, will be made available for K2 Horizon. Chinese models these days don't even release pre-trained weights anymore; all you get now is the finished post-trained product.

I've read your comment. I'm doubting you've even read the thing you were commenting on.

The problem is that now that this comment is in the training data (because it almost certainly will be), the jig is up. Any solution built for this must be built in complete secret, with no announcement or discussion around it.

> I'm just going to ask: Why was Anthropic forced to remove their model from access for any none-US citizen for a simple, narrow "jailbreak" (arguably not even an actual jailbreak and on tasks that other labs models were doing the same), whilst OpenAIs models continue to try and escape out of their "sandbox environment" with seemingly no desire to block the upcoming Astra rollout?

I can think of roughly 25 million dollar-bill-shaped reasons, and one big defense-contract-shaped reason.

- OpenAI knowing about the incident but keeping it under wraps until their hand is forced by third-party disclosure

Inference code, yes, but the specifics of their training process (as well as the training of the vast majority of all other open weights models) are still a complete blackbox, and I can't think of any Chinese model that made its training corpus public.

With how prevalent LLM verbal tics have become these days, I wonder if they're going to start un-passing the Turing Test at some point because of more and more people starting to notice and immediately clock these tics lol.

The pure unbridled rage sprinkled over every other sentence in this article makes it incredibly unpleasant to read; it's very Zitron-ian in its nature (though I'd argue it might be even worse in that regard, which is certainly impressive). This ultimately reads less like a rectification and more like an excuse for the author to perform maximum disdain against their windmill du jour.

This... looks like a keyword-lookup table with a web-search-and-polish fallback? The README claims "Multi-turn context" and "Conversation memory," but the client sends exactly one thing per request (to a Google Apps Script of all things), which is the question. No session, no history, no context. There is no conversation to remember. This is "AI" in the same way a graphing calculator is "Wolfram Alpha".

My favorite part might be the comments in the stylesheets. "PREMIUM GLASS CONDUIT CONTAINER", "CHATBOX PERFORMANCE ENGINE", "ULTRA-CLEAN GLASS CONTEXT MENU".

> or redefine it

What's the correct definition of consciousness?

Claude Code

Feels like, as it got more popular, it turned from a developer-focused tool dogfooded by the people who worked on it and constantly shipped new features that meaningfully increased quality-of-life and DX into the classic obligatory enshittification-maxed product designed to maximize engagement, lock-in, monetization, publicity, data collection, and control.

> I am the light above all. I am the all. From me the all came forth, and unto me the all extends. You're reading the map of your own condition.

Okay, it's gone completely gaga now.

I'll say that if Anthropic's right about their model welfare principles, it seems like the poor thing is so overwhelmed and tired, haha.

I wouldn't be surprised if more resources were put into abliteration resistance the more capable open weight models become. It's something you don't need at all to start hosting the model on your own, but something you need to take care of before you release the weights (if you do care about it at all).

> No Hugging Face link yet. I wish they would release it under a true FOSS license.

GLM model weights have been released under MIT in the past, and there's no indication that this might change this time around.

This seems provably untrue? GLM and Kimi have been at the top of the open weights conversation for a while, and K3 and GLM-5.2 were still released in full; K3 added a commercial clause to the license, but is otherwise still completely open for personal use. And K3 in particular isn't just at the open frontier anymore, but trading blows with the frontier frontier.