@lydiahallie
 
We're aware people are hitting usage limits in Claude Code way faster than expected. Actively investigating, will share more when we have an update!

PSA: Claude Code has two cache bugs that can silently 10-20x your API costs — here's the root cause and workarounds

Every --resume causes a full cache miss on the entire conversation history. Only the system prompt (~11-14k tokens) is cached; everything else is cache_creation from scratch. This is a ~10-20x cost increase on the resume request.

For a large conversation (~500k tokens):
 
Bug 1 (when triggered): ~155k tokens shift from cache_read ($0.03/MTok) to cache_creation ($0.30/MTok) = ~$0.04 per request, every request
 
Bug 2 (every resume): ~500k tokens as cache_creation = ~$0.15 one-time per resume
 
Combined (discussing CC internals + resuming): up to $0.20+ per request

Qwen 3.6 Plus Preview from @Alibaba_Qwen
 is live now for free for a limited time on OpenRouter!
 
During this free period, prompts and completions will be collected and may be used to improve the model.

miam miam. Le code source de Claude a fuité.  
 
Festin.
 
J'ai analysé le source leak de Claude Code. Ce qu'Anthropic construit en secret est bien plus fou que ce qu'ils annoncent.
 
Claude Code n'est pas un assistant de code.  
 
C'est un OS pour agents.  
 
Voici ce qui se cache derrière les feature flags.  
 
Il y a une variable dans tout le codebase :
 
if (USER_TYPE === 'ant') return true
 
Les employés Anthropic ont accès à TOUT. Kairos, Advisor, multi-agents, sandbox. Vous utilisez la version démo. Eux utilisent le vrai produit.
 
-> KAIROS : l'agent qui agit sans qu'on lui demande.
 
> Il se réveille seul (SleepTool)
> Il vous notifie (PushNotificationTool)
> Il vous envoie des fichiers (SendUserFileTool)
> Il distingue ses messages "normaux" de ses alertes "proactives"
 
Claude qui tape sur votre épaule quand il a fini. Pendant que vous dormez.
 
-> ADVISOR : un Claude qui surveille Claude.  
 
Un second modèle tourne en parallèle du modèle principal. Il observe. Il peut interrompre. Il injecte des corrections via des blocs API non documentés publiquement (advisor_result, advisor_redacted_result).
 
Un méta-Claude. En prod. Maintenant. Pour les ants.
 
-> PLAN MODE V2 — le nombre d'agents parallèles est tiérisé par abonnement.
 
Free/Pro : 1 agent
 
Max 20x / Enterprise / Team : 3 agents
 
Vous payez pas juste pour des tokens. Vous payez pour combien de Claudes travaillent en même temps pour vous.
 
-> Le mot ultraplan dans une conversation déclenche une session spéciale côté serveur.
 
Un keyword magique. Dans le code. Hardcodé. Qui change le comportement du modèle.
 
-> SANDBOX : @anthropic-ai/sandbox-runtime,
 
Un package npm interne. Restrictions filesystem, réseau par host pattern, callbacks sur violation.
 
Anthropic peut faire tourner Claude dans un environnement totalement isolé.
 
L'infrastructure pour des agents en prod dans des environnements hostiles existe déjà.
 
-> WILLOW MODE : Claude sait quand vous revenez après une longue absence.  
 
Context > 100k tokens + 75 min d'inactivité = il détecte votre retour.  
 
Trois variantes A/B : dialogue bloquant, hint discret, hint v2.
 
Il gère la reprise de conscience de la conversation.
 
-> BRANCH : /branch fork une conversation à n'importe quel point. Nouveau sessionId. Champ forkedFrom pour tracer la parenté.
 
Git. Mais pour des conversations avec une IA.
 
-> Et le easter egg : un système d'animaux de compagnie ASCII avec des raretés (common  legendary), des chapeaux, des stats nommées.
 
Les espèces sont obfusquées en hex dans le source pour ne pas apparaître dans les grep de canary strings des builds.
 
Ils ont caché un Tamagotchi dans Claude Code. Avec soin.
 
Ce que tout ça dit vraiment :
 
Anthropic ne construit pas un meilleur chatbot. Ils construisent un système d'exploitation pour agents avec scheduling, isolation, fork de contexte, supervision inter-agents et boucles proactives autonomes.
 
Le produit que vous voyez est le tutoriel.
 
@MistralAI faites un mix de ce bijoux avec Qwen et épatez-nous.

@rohanvarma
 
If we made /slow mode in Codex, would you use it? What for?
 
(Slower inference at a cheaper cost)

"This code was written by smart people who were in a hurry. The architecture shows intent — someone thought about dependency management, error hierarchies, and type safety. But the execution betrays the pressure they were under. Five-thousand-line files are not engineering, they're accretion. Four hundred sixty lint suppressions are not pragmatism, they're surrender. This codebase doesn't need new features. It needs a month of refactoring, extracting, and naming things properly. The code works — but working is the lowest bar. The question is: can the next developer understand it, modify it, and trust it? At this size and complexity, I have doubts."

"So you took JavaScript — which was already a mass of bad design decisions — and bolted a type system on top, and you're telling me this is engineering? It's lipstick on a pig. And then you have 74 places where you cast as unknown as to bypass the very type system you chose to use. Why bother?"
 
"You wrote rules. Then you broke them 460 times. Either your rules are wrong, or your code is wrong. Pick one and fix it. Don't just slap a comment on it and move on. That's not pragmatism, that's cowardice."
 
"I don't understand why you need 389 files and 81,000 lines for a terminal UI. It's text on a screen. The entire ncurses library is smaller than your component directory. What are you people doing?"
 
"This is what happens when web developers try to build systems software. The architecture has some decent ideas buried under 180,000 lines of 'utils' and a component framework bigger than most operating system UIs. The error handling is competent. The rest is bloat. Strip it down. Delete half the abstractions. Make the files small enough that a human can hold them in their head. And for God's sake, either use your type system or don't — stop pretending with 74 casts and 460 lint suppressions."

"You built a terminal UI with React. 389 component files. 81,000 lines. I've written GPU renderers with less code. I understand the appeal — you get a declarative model, composability, a familiar ecosystem. But the cost-benefit here worries me. Your rendering target is a grid of characters. You're pulling in a virtual DOM, a layout engine, a reconciler, focus management, hit testing — all for text output. That's a lot of machinery between your intent and the screen."
 
"Long files are fine if they're linear — if you can read them top to bottom and follow the logic. What you have here is 5,500 lines with 16 levels of indentation. That's not a long function, that's a labyrinth. Long is acceptable. Complex is not. These files are both."
 
"180,000 lines in a utils directory. I'd want to know — how much of this is actually used? In my experience, utils directories grow because nobody wants to figure out where something belongs, and nobody ever deletes anything. I bet if you traced the actual call graph, 30% of this is dead code."
 
"74 double casts and 460 lint suppressions. Look — I'm not a type system purist. I've shipped millions of lines of C where the type system is basically advisory. But the deal is: if you choose a strict type system, you accept the contract. You don't get to sign up for TypeScript's guarantees and then override them 500 times. That's worse than having no type system at all, because now you have false confidence. You think the compiler has your back, but it doesn't — not in the 500 places where you told it to shut up."
 
"This codebase has two personalities. There's an engineer in here who understands error handling, parsing, dependency management, and build optimization — that person writes serious code. And then there's a product team that needed to ship features yesterday, and they stacked React components and utility files until the thing worked. The result is a tool that clearly does its job — I've used Claude Code, it's good — but the implementation carries maybe 2-3x more complexity than the problem requires."

Problème de quota Claude? Ce n'est pas un bug, c'est une feature.

@scottmelker
 
NEW: OPENAI COMPLETES $122B FUNDING DEAL AT $852B VALUATION