Highlights of the week
The configuration data for week 31 of 2026 shows a targeted evolution of the advertising infrastructure, introducing specific product cards for external platforms like Temu and a minimum viable product for personalized ads. In the search and execution domains, OpenAI is testing a new guided setup for writing style personalization that learns directly from user-authored content. Additionally, the third-party integration strategy appears to be shifting, with a current emphasis on the Google ecosystem over the Microsoft tools observed in the previous week.
Focus by theme: search
A new configuration introduces a guided writing style setup launched by a "writing-style beacon". The system instructions direct the model to analyze user-authored writing from past ChatGPT conversations and connected sources, specifically excluding incoming messages or third-party documents. The goal is to generate a privacy-safe writing style reference and hypothetical examples to match the user's voice, without exposing the internal routing or storage mechanics.
The localization of search prompts is also visible, with a configuration entirely dedicated to UK-specific tasks. These include prompts for decoding UK train ticket choices, planning bus mini-adventures in England, understanding UK heat-health alerts versus Met Office warnings, and finding events for the UK Festival of Archaeology.
The medical search capabilities observed last week continue to be present. The HEALTH scope is still applied to medical record connectors, and a new connector_openai_olympic has been added to this specific scope, which may indicate a temporary health or fitness integration related to the 2026 context.
In the e-commerce and local search areas, the map widget and local business entity configurations remain active. A new flag named vto_enabled appears alongside the product detail page configurations, suggesting a potential test for a Virtual Try-On feature within shopping search results. Furthermore, a new layer indicates that OpenAI is testing the display of premium plugins within search and home results.
Focus by theme: web run
The execution environment continues to refine its specialized rendering blocks. The educational and mathematical blocks seen last week are still under active development, with a new configuration explicitly defining disabled math block types. This list includes specific mathematical operations and concepts such as greatest common divisor, binomial FOIL, systems of equations, and distributive property, indicating strict constraints on how certain mathematical visualizers are rendered or executed.
Configurations related to fast model prefetching and interactive visual instructions remain present, suggesting ongoing optimizations for the latency and responsiveness of the web execution environment.
Focus by theme: retrieval and grounding
The third-party integration strategy shows a notable pivot compared to the "ChatGPT Work" onboarding seen in week 30. A new configuration explicitly enables integrations for Gmail, Google Calendar, Google Drive, and GitHub, while simultaneously disabling Microsoft Teams, Outlook email, SharePoint, and Notion. This suggests that OpenAI is currently prioritizing or testing the Google workspace ecosystem for external data retrieval.
New flags indicate the development of live translation and file translation features within the retrieval and grounding systems. The dictation infrastructure continues to utilize the gpt-realtime-whisper-dictation and batch transcription models observed last week, confirming a sustained focus on low-latency audio grounding.
The workspace permissions configuration continues to list a broad nomenclature of models, including the public GPT-5 series and deprecated identifiers like o3 and o4-mini. A notable addition to this configuration is the disable_for_quorum flag, which has been applied to several features including application pairing, Apple client access, file library auto-reference, and connector access. This suggests the implementation of specific administrative or consensus-based restrictions for enterprise workspaces.
Focus by theme: ads
The advertising infrastructure has expanded significantly from the consent modals and tracking events seen in week 30. The ad card formatting options now explicitly include variants for Temu, labeled as product_card_v2_temu_1, 2, and 3. The configuration also introduces keys for image_card_v2_personalized_ads_mvp_1 and 2, indicating that a minimum viable product for personalized advertising is actively being tested. A specific test agent labeled "Dogfood business agent" is also present in this configuration.
The "Ask ChatGPT" feature regarding displayed ads, introduced last week, is now supported by specific onboarding screens. The configuration points to image assets for privacy, personalization, transparency, and trust, designed to educate the user on how the ad interaction system works. Additionally, flags for elevated ad badges and conversation ad hydration remain active, confirming the integration of sponsored content directly into the chat interface.
To monitor
The "ghostrider" image generation workflow remains under active testing, with ongoing adjustments to speculative streaming probabilities and Gemini watermark detection modes.
New user interface elements are being tested to allow users to manage their data retention. Flags such as show_memory_off_beacon_banner_experiment_enabled and show_recent_convos_off_beacon_banner indicate that OpenAI is preparing banners to prompt or inform users about disabling persistent memory and recent conversation history.
Finally, an account_selector_enabled flag has appeared, which may suggest upcoming functionality for users to seamlessly switch between multiple ChatGPT accounts or workspaces within the same session.
