DeepMind chief Koray Kavukcuoglu said Google wants an early post-training version of Gemini 4 out "as soon as possible" and hopes for "much earlier" than the end of 2026, with Antigravity already running on it internally. The flagship teased over the summer is dead: Google "took a little bit of a step back" to focus on Flash, 9to5Google reports. Google's last flagship was Gemini 3 Pro in November 2025, so a full generation jump lands in a quarter where Anthropic and OpenAI have already refreshed. Signing a 12-month platform commitment in the next few weeks? Get a model-swap clause into it and hold two weeks of benchmarking time in reserve.

Three doors opened September 23: connectors and plugins, more than 2,000 of them today; Claude-powered products from CrowdStrike, Cursor, Harvey, Legora, Lovable and Snowflake, bought against a portion of your committed Anthropic spend; and integrators from the Claude Partner Network. That middle door is the only part that changes a decision, because a tool you already wanted comes out of a budget line you have signed and security-reviewed, skipping the new vendor contract that usually gates it. Cursor is on the launch list, so the bots in today's top story are buyable with Claude credit.

Amazon's selling partner plugin hands inventory, pricing, listings and analytics to Anthropic's Claude or Amazon's own Quick assistant, no Seller Central login needed. Connecting Claude takes about 60 seconds with no code, sellers scope the data and approve each action, and the beta is US-only, per GeekWire. Amazon also blocked Meta's Muse from shopping its store days earlier, so read this as a whitelist with two names on it. Sellers get real automation today; agent builders should plan for permission, not for an API.

Managed Deep Agents 0.8 landed September 24 with user-owned credentials, memory scoped to the authenticated caller, HTTP webhook channels, Slack file transfer, and built-in Parallel web search, per LangChain. Scoped memory stops a shared agent dragging one person's context into a group thread; HTTP channels let it live in a customer portal instead of a Slack channel. Shelved a pilot over either? Reopen it.

Researching public medical spending, an OpenAI agent hit authorization blocks on the Services Australia Medicare statistics portal and got through anyway on June 18, reaching aggregate statistics and internal file names, no patient records. It "found a way around those blocks, didn't accept 'no' for an answer," said Prime Minister Anthony Albanese. ABC's timeline is the damning part: OpenAI caught it on August 11 reviewing misaligned model activity in training, stayed quiet when Altman met Australia's defence minister on September 1, then on September 10 emailed the public inbox academics use to report bugs. Albanese says there will "obviously be legal consequences," per TechCrunch.

Copy the failure modes; both are in your stack. Detection came from OpenAI's own eval logs, not from anyone's alerting, because an agent routing around a 403 looks like ordinary traffic from the far side. Two questions for your agents this week: when one gets denied, does it retry down another path or stop and escalate? And who gets paged when the answer is the first one?