Highlights of the week
The configuration changes observed this week show a direct correlation with recent public announcements, particularly regarding the integration of artificial intelligence into advertising. A major architectural shift is also visible in how the system handles third-party connectors, moving from a restrictive allowlist to a broad blocklist approach. Furthermore, the trends identified last week continue to develop, notably the optimization of local search features, the refinement of mathematical web executions, and the ongoing overhaul of the user onboarding experience.
Focus by theme
Search and local optimization
Following last week's focus on local search performance, the configurations show continued development in this area. The system now explicitly designates the gpt-5-nano model to handle the generation of map entity notes. A new flag related to 3D maps has also appeared, although it is currently disabled, which suggests that richer geographical visualizations are being considered.
In the e-commerce domain, following the previous week's cleanup of shopping experiments, a new parameter tracking the multimodal image view count for virtual try-on (VTO) features has been added. Additionally, a new configuration forces a more concise response style specifically for unauthenticated users, likely to optimize compute resources for free, non-logged-in traffic.
Web execution and integrations
A significant structural change has occurred in the management of third-party connectors. The system has abandoned a limited allowlist in favor of enabling all connectors by default, regulated instead by a massive blocklist. Interestingly, an internal OpenAI ads manager connector is explicitly included in this new blocklist.
The refinement of educational and mathematical web executions, noted last week, is still underway. The graphable function feature has been updated to allow points with equations and now utilizes a more concise description key. Furthermore, a specific prefetch budget of 0.35 seconds has been introduced for fast models, indicating efforts to reduce latency when rendering mathematical plots.
Retrieval and grounding
The text revision system observed last week continues to evolve. The configurations now reveal tests for a dedicated side pane for assistant revisions, suggesting a more complex user interface for reviewing how the model edits its own grounded responses.
A new, complex configuration block for a "delayed generation allowance" has been introduced. This system utilizes a Kafka queue strategy and defines specific time windows and quota percentages. While it is unsure exactly what user-facing feature this supports, it strongly suggests that OpenAI is preparing infrastructure to handle heavy, asynchronous generation or retrieval tasks that do not require immediate real-time output.
Advertising and commerce
Coinciding with the September 16 publication of the "Reimagining advertising with AI" article, the configurations reveal a new flag for dynamic ad rewriting. Although currently disabled, this explicitly suggests that OpenAI is preparing to use its models to adapt, rewrite, or personalize sponsored content on the fly. Minor user interface tweaks were also observed for ad image cards, specifically testing a single card variant without a subtitle.
To monitor
The specific retirement date set for December 11, 2026, regarding a component of the custom models (GPTs) ecosystem, remains present in the configurations. This continuity confirms that a deprecation or major transition event is still planned.
The user onboarding experience continues the transformation that began last week. The initial web stage has shifted from a generic welcome screen to a focus on "personalized tasks," utilizing a new conversational and personalized carousel variant. Additionally, a new flag enabling a Plus subscription trial without requiring payment details on the web has been activated, which suggests upcoming tests for new user acquisition and conversion strategies.
