Google cuts AI costs, Substack scans copy: July 22 brief

Google has made high-volume AI work cheaper, which matters to any business paying to process support tickets, product copy, or customer research at scale. OpenAI is also courting smaller companies directly. Meanwhile, Substack is giving readers a way to question whether a post involved AI. Today is about cost on one side and trust on the other.

Google lowers the bill for fast AI work

Google released three Gemini models on July 21, led by Gemini 3.6 Flash at $1.50 per million input tokens and $7.50 per million output tokens. A token is a small piece of text that an AI system reads or writes. Google says the new model uses 17% fewer tokens than Gemini 3.5 Flash on one independent test, so some jobs should need less paid processing as well as carrying a lower listed price.

That matters when AI runs repeatedly in the background. A staff member drafting one email will barely notice a per-million-token change. A company sorting thousands of messages, rewriting a large product catalog, or producing first drafts every day will. This is the part of AI pricing that gets missed: the cost of one request looks tiny, while a repeated process turns it into a line item.

Google also introduced 3.5 Flash-Lite, its faster low-cost option, and 3.5 Flash Cyber, a security model limited to governments and selected partners. Most owners can ignore the security release. The useful comparison is between the model already inside your process and 3.6 Flash on the same real work.

This also sharpens the case for choosing AI writing tools by the job they perform, rather than paying for the newest name on the label.

Substack lets readers scan writing for AI

Substack began rolling out a Pangram-powered detector on July 21 that can scan posts, notes, replies, and comments longer than 100 words. Readers on the web and iPhone app can open the three-dot menu and ask for an estimate of how much text may involve AI. Android support is due later.

The platform is also adding a “How I make this” statement, where writers can explain their process. That second feature may be more important. AI detectors make estimates, not proof, and Substack itself says the scan cannot judge how much human care went into a piece.

For a business publishing expert advice, the trust risk now sits in plain view. A reader may see an AI flag even when a person researched, checked, and rewrote the material. Generic copy becomes harder to defend because the audience has a button for challenging it.

Our view is simple: businesses using AI to draft should keep the human evidence visible. Add named sources, original examples you can support, and a short process note when the subject affects money or safety. If your newsletter is a serious channel, our newsletter platform comparison explains the broader trade-offs around owning that reader relationship. Strong editing matters more than trying to beat a detector.

OpenAI courts small businesses directly

OpenAI introduced its ChatGPT for small business program on July 21. The announcement is aimed at smaller companies adopting ChatGPT at work, rather than developers building their own software around an AI model.

The business consequence is bigger than one program. OpenAI is separating everyday company use from the wider consumer product, which signals more attention to setup, staff use, and business controls. Owners should expect the buying conversation to shift from “does anyone here use ChatGPT?” to “which work belongs in it, and under whose account?”

Do the housekeeping before chasing any program benefit. List the recurring tasks where staff already use ChatGPT, remove customer secrets from casual prompts, and assign one person to approve shared instructions. Then measure whether the work saves paid time. A special label does not repair a loose process.

For businesses that want to be found inside assistants as well as use them internally, our guide to how AI systems recommend companies covers the customer-facing side of the shift. The two jobs need different plans.

Worth watching

• OpenAI and Hugging Face disclosed a security incident during model testing. The immediate lesson is narrow: pre-release AI systems still require tightly limited access.
• OpenAI’s new advertising page appeared one day after we covered ChatGPT’s published ad pricing. Watch for clearer eligibility and measurement details before treating it as a normal ad channel.
• Google says Gemini 3.5 Pro is still testing with partners. Businesses do not need to delay a current cost review while waiting for it.