My takeaways:

• Input costs compound fast: Even though output tokens cost more individually, input volume compounds on every turn, which is why long chats quietly drain your budget. Plus, longer prompts gradually degrade the model's accuracy even on simple tasks.

• Language isn't billed equally: Token counts for the exact same text can differ by up to fifteen times across languages, meaning speakers of some languages get less usable context space and higher costs.

• Tools trigger total restarts: When a chatbot uses a tool (like searching the web), it isn't actually browsing - it's just writing a request. The surrounding app catches it, runs it, and feeds the result back as a brand new input that repeats the whole pipeline from the beginning.

My takeaways:

• When AI models answer hard questions, they generate a full reasoning process that is never displayed, using an encrypted AEAD envelope sent to the client to maintain stateless conversation continuity.
• Because authenticated fields often lack user or session bindings, blocks can suffer from cross-session, cross-user, or cross-model replay, allowing blocks from one context to be accepted in another.
• Encrypted reasoning blocks from strong models can be fed into weaker models in the same family as prior context and transcribed into plaintext, entirely bypassing the original model's refusal training and output filters.
• Because sanitization only acts on plaintext, public reasoning traces frequently leak sensitive data like API keys, passwords, and private tokens that were stripped from the visible chat output.