AI got cheaper. Your token bill didn't.

Subject: AI got cheaper. Your token bill didn’t.

Preview text: A 30% token tax hiding in the small print, and how to get AI answer engines to cite you by name.

Dear Reader,

Two fresh posts from me this week — one about who gets to use the frontier models (and who profits from them), and one about making sure AI answer engines credit you rather than a competitor. They pair rather well, as you will see.

AI Weekly Signals: Tokenizer Tax, Cache Rules, and Who Owns AI’s Upside https://daehnhardt.com/blog/2026/07/03/sonnet-5-tokenizer-tax-ai-weekly-signals/

My weekly AI signals round-up opens with Anthropic releasing Claude Sonnet 5 — near-Opus-class work at a lower price, now the default on Free and Pro plans. Lovely, except for one detail buried in the docs: a new tokenizer means the same text now costs roughly 30% more tokens. It is a bit like your supermarket cutting the price per pack while shrinking the pack — recount before you trust last month’s budget. Meanwhile, Google folded computer use straight into Gemini 3.5 Flash, making “my agent can click things” a default feature rather than a premium one.

On the exact same day Sonnet 5 launched, the US government lifted the export-control order that had forced Anthropic to suspend Claude Fable 5 for more than two weeks — a useful reminder that infrastructure you don’t control can be switched off by people who aren’t your vendor. Not to be outdone, OpenAI previewed its GPT-5.6 family to the public while gating it to roughly 20 government-vetted organisations — then reportedly floated the idea of handing the US government a 5% equity stake — early-stage discussions, worth some $42.6 billion at OpenAI’s reported valuation only if they ever conclude. The week’s question shifted from who gets the model to who owns the upside. I close with Anthropic’s Claude Science launch and a finding worth bookmarking: in one narrow domain, viral sequence queries, a single deterministic retrieval tool took models from unreliable to near-perfect accuracy.

What Is GEO? How I Optimised My Blog for AI Search https://daehnhardt.com/blog/2026/07/02/what-is-geo-generative-engine-optimization/

Which brings me to the second post, because if AI engines are now doing the reading, you want them to cite you. GEO (Generative Engine Optimization) is structuring your content so ChatGPT, Perplexity, and Google AI Overviews are more likely to cite it and name you as the source — no guarantees, just better odds. My mental model: SEO gets your book onto the right library shelf; GEO is the paragraph the librarian reads aloud when someone asks a question. I walk through exactly what I changed across roughly 95 posts — one-sentence definitions, question-shaped headings, answer-first paragraphs — and share the first honest results, including what I cannot yet claim credit for. There is a checklist you can steal at the end.

Take care, and do hit reply if either piece sparked a thought.

Elena