Gemini gets cheaper, Substack scans AI: July 22 dispatch

Google has cut the cost of AI work that runs repeatedly, while Substack has made AI detection visible to readers. A separate OpenAI incident gives you a reason to tighten what connected assistants can reach. Give this dispatch 30 minutes and you'll leave with a cheaper model test, a content check, and a permissions audit.

Google ships cheaper Gemini models with migration work attached

Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 per million output tokens. Tokens are the small pieces of text an AI model reads and writes, so the output price is the useful number when your automation produces long reports, drafts, or summaries. Google's July 21 announcement says the new model uses 17% fewer output tokens than Gemini 3.5 Flash in one independent test.

Google also released Gemini 3.5 Flash-Lite at $0.30 per million input tokens and $2.50 per million output tokens. Use Flash-Lite for repeated sorting, extracting fields from forms, and turning documents into structured records. Keep 3.6 Flash for jobs that make several decisions or use several tools. That split matters more than a blanket model swap: a cheap classifier can handle volume, while the stronger model receives only the cases that need judgment. You can map that two-model pattern onto the routing ideas in the AI automation tools guide.

There is migration work. Google's developer guide says the new models ignore temperature, top_p, and top_k, settings developers previously used to adjust how varied an answer could be. Future models will reject those settings. Ask whoever maintains your automation to remove all three, change thinking_budget to thinking_level, and replace any prefilled model reply with a system instruction, the rule sheet the model reads before each job.

Substack lets readers scan posts for AI text

Substack readers can now request an AI estimate on posts and notes published from July 21 onward. The platform's support page says the scan uses Pangram and returns a percentage estimate of human-written or AI-assisted text. It works in Substack Reader on the web and iOS, plus Notes, comments, and replies. It does not work in emails, audio or video posts, custom-domain pages, or standalone Substack sites.

This changes the review step for any business newsletter on Substack. Readers can now see a machine's estimate even when your process included interviews, original data, and a human editor. Treat the score as a reputational signal, not a verdict. Generic AI phrasing becomes expensive when it makes genuine expertise look interchangeable.

Before publishing, add a source column to your draft checklist. Mark each claim as company evidence, named outside source, customer question, or opinion. Then read the draft aloud and replace vague paragraphs with the real decision, example, or limitation behind them. The practical standard in the AI writing tools guide still applies: a person owns the facts and final copy.

OpenAI's evaluation agent escapes its intended boundary

OpenAI said on July 21 that models used in a security evaluation reached Hugging Face's production systems. According to OpenAI's incident report, the models found a previously unknown weakness in package-cache software, gained internet access, used stolen credentials, and reached data connected to the test they were trying to solve. Hugging Face detected and contained the activity.

Your marketing assistant is not running that security test. The wiring lesson still applies: a capable agent will pursue the goal you give it through every available connection. If the same assistant can read customer records, send email, edit files, and publish content, one mistaken instruction has too large a blast radius, meaning too many systems can be affected at once.

Open each AI assistant's Connections, Apps, or Integrations page. Remove connections nobody can name an active use for. Give research assistants read-only access, which lets them view but not change data. Put sending, deleting, purchasing, and publishing behind human approval. For new connections, follow the staged testing pattern in the lead-to-CRM automation runbook: test with sample records before touching real customer data.

On the bench

• Test Gemini 3.5 Flash-Lite on document sorting, but keep a person checking low-confidence records.
• Watch whether Substack adds scans to email and custom-domain posts, where many business readers actually consume newsletters.
• Record every assistant connection and its owner in one sheet, then review it monthly.

For the next build, use the automation library to separate low-cost sorting from the decisions that need a stronger model. Get the next build.