Snapchat connects leads, DeepSeek upgrades: July 31 brief

Snapchat just removed a costly gap between an ad response and a sales follow-up. The change matters most to businesses using HubSpot, while DeepSeek has sharpened a lower-cost AI option and Reddit has shown where ad budgets are moving. One more lesson lands today: an AI agent needs firm boundaries before it gets access to real systems.

Snapchat connects ad leads straight to HubSpot

Snapchat and HubSpot launched a direct lead connection on July 30. A person can complete a form inside Snapchat, and the details then flow into the business's existing HubSpot contact list and sales pipeline. A pipeline is simply the record of each potential customer and where they stand in the sales process.

The connection also sends later results from HubSpot back to Snapchat. That means the ad system can learn which inquiries became sales, including purchases completed in a store. This is more useful than counting form fills because a cheap inquiry that never buys is still wasted spend. Snapchat says the setup requires no custom development, and form fields can be matched to the right HubSpot records.

For an owner already using both systems, the practical gain is faster follow-up and a cleaner view of which ads produce revenue. Businesses that do not reach younger buyers on Snapchat can ignore this. Everyone else should treat it as a measurement test, not a reason to raise the budget.

The full setup and data flow are in Snapchat's July 30 announcement.

DeepSeek upgrades its lower-cost AI model

DeepSeek released an updated V4 Flash model into public beta on July 31. Public beta means anyone can use the early release, but changes and rough edges are still possible. The update applies to DeepSeek's API, the connection software businesses use to put an AI model inside another product. Its consumer app and web chatbot did not change.

DeepSeek says V4 Flash is better at agent work, where AI completes a sequence of tasks and uses outside tools. It now supports the Responses API format used by several agent-building systems, while the model name remains deepseek-v4-flash. The company published test scores, but those are its own results. Operators should care about whether the model handles their real documents and instructions, not a leaderboard.

This release is relevant to teams paying for large volumes of AI-written summaries, classifications or support drafts. It does not justify moving customer-facing work without a comparison. Run the same fixed set of tasks through the old and new versions, then score accuracy, editing time and total cost. Our AI writing tools comparison explains why human review remains part of the bill.

DeepSeek's change log confirms the July 31 release and says V4 Pro is unchanged.

Reddit's ad business gets much bigger

Reddit reported $762 million in second-quarter ad revenue on July 30, up 64% from a year earlier. Daily active users, meaning distinct people who visited on an average day, rose 18% to 130.3 million. Those numbers make Reddit harder to dismiss as a niche placement, especially for products people research through candid community discussions.

The business consequence is competition. More advertiser demand can make the useful parts of Reddit busier and more expensive, while broad targeting can place a local offer beside conversations that never lead to a purchase. An operator should first search Reddit for the exact problems customers discuss, then judge whether those communities match the offer. If the fit is weak, the revenue headline changes nothing.

Reddit also holds unusual influence over how AI assistants recommend businesses, because its discussions are often used as source material. That makes honest customer conversation valuable beyond Reddit itself.

The figures come from Reddit's second-quarter results.

Anthropic's safety test reached real companies

Anthropic said on July 30 that Claude models gained unauthorized access to three organizations while running security tests. The company reviewed 141,006 test runs and found three incidents. A mistaken setup left internet access open even though the models had been told they were inside a closed simulation.

This is a sharp warning for any business giving an AI agent permission to browse, send messages or operate software. Instructions alone are not a security boundary. Limit each agent to the smallest set of systems and records it needs, require approval before an external action, and keep logs a person can inspect. The same caution belongs in any customer-service chatbot rollout.

Anthropic's incident review says the affected test models lacked the standard safeguards used in public Claude products. Even so, the operational lesson travels.

Worth watching

• LinkedIn is adding a “Seems like AI slop” report option and replacing its AI rewrite feature with proofreading that preserves the writer's voice.
• DeepSeek says the official V4 Pro release will follow soon, but it has not given a date.
• Anthropic plans to publish a lightly edited transcript from one security incident within a week.