Gemini 3.8 lands, Meta splits data pricing: Sept. 3

Gemini 3.8 can spend more words thinking even though its listed price has not risen, Meta now makes data use part of the model price, and a Perplexity audit found thin buying guides inside its evidence. Give these changes 30 minutes. You'll set one cost limit, check one privacy choice, and tighten how your team verifies AI recommendations.

Google ships Gemini 3.8 Flash with adjustable effort

Gemini 3.8 Flash arrived on September 2 at $0.75 per million input tokens and $3.75 per million output tokens. A token is a fragment of text counted for billing as the AI processes or produces language. Google's announcement says those rates match the introductory price of Gemini 3.7 Flash, but the new model can use more tokens on difficult work because it takes extra reasoning steps and calls tools repeatedly.

That distinction can raise your bill without a price increase. A customer-research routine, for example, may read more material and make more searches before returning the same one-page summary. Businesses using Gemini only through the consumer app can watch output quality. Anyone paying for usage needs to watch total tokens per completed job.

Open the duplicate or test version of your automation. Replace its current model with gemini-3.8-flash, set the effort control to a lower level first, and run ten normal jobs through both versions. Record the total input tokens, output tokens, completion time, and whether a person approved the result. Keep 3.7 Flash if the new route costs more without improving approvals. The AI content approval workflow gives you a simple human review stop.

The part that breaks is comparing the price card instead of the finished job. Your useful number is cost per approved result. If the workflow also runs through an automation service, use the automation platform cost guide to count those charges separately.

Meta makes training use a model-pricing decision

Meta published Muse Spark 1.3 on September 2 with two sharply different prices. The contributor version, whose data is used to improve Meta's products, costs $0.10 per million input tokens and $0.20 per million output tokens. The version marked as not used for improvement costs $1.25 for input and $4.25 for output. Both accept up to one million tokens of context, meaning the amount of material the model can consider at once. The figures appear on Meta's model page.

For a business owner, this is a data-classification choice before it is a model choice. Public product descriptions and dummy test records may fit the cheaper route. Customer messages, unreleased offers, internal pricing, and staff documents should not be sent merely because the contributor model costs less.

Add a field named data_class before the model step in your automation. Allow only public or synthetic, meaning made-up test data, to reach the contributor model. Send confidential work to the non-contributor version or keep it on your approved provider. Log the selected model beside each job. For the review fields and permission check, adapt the content approval runbook.

Ignore Muse Spark if you do not run usage-based AI models inside automations. If you do, make the privacy route explicit now. The expensive failure is letting a staff member choose the cheaper model without seeing what that choice permits.

Perplexity audit finds manufactured buying guides in citations

7,534 citations from a September 2 test included 215,128 generated “best software” pages across three apparently connected sites. The published audit and dataset asked Perplexity's Sonar and Sonar Pro models for five products in each of 380 software categories. Of the citations returned, 59.8% pointed to domains ranked below the top 100,000 sites, while 23.4% pointed outside the top million.

This does not prove that every recommendation was wrong. The test covered Perplexity only, and website popularity is not the same as accuracy. It does show why an AI answer should not be your purchasing shortlist by itself. A confident answer can rest on mass-produced pages with inconsistent rankings.

Change your buying prompt to require the vendor's official feature page, official pricing page, and one source that explains its testing method. Then open every cited page. Reject any comparison that hides its owner, evidence, test date, or selection method. If you are working on your own visibility in AI answers, pair that check with the AI recommendation visibility setup, which focuses on clean business data rather than bulk list pages.

The part that breaks is treating a citation as proof. A link only tells you where the model looked. Your team still has to decide whether that page earned trust.

On the bench

• Gemini 3.8 Flash Cyber is limited to trusted defenders, so most small businesses should wait for ordinary security tools to adopt its work.
• Meta says Muse Spark 1.3 can inspect images, video, and documents. Test only with synthetic records until your data route is documented.
• Re-run the same software question in two AI services next. Compare the sources, not just the product names.

Get the next build from the AI tools runbooks.