Retirement of custom GPTs and proactive ChatGPT voice mode

Highlights of the week

The configuration data reveals a major architectural shift in how real-time voice mode operates, introducing a proactive backend agent designed to investigate requests before the user finishes speaking. This aligns with the launch of GPT-6.1 Terra this week, which is optimized for speed and precise computer use. Additionally, the data indicates preparation for a significant product transition: the potential retirement of custom GPTs by the end of the year, alongside the introduction of a new, highly granular subscription tier system priced dynamically by region.

Focus by theme

Search

The map and local search functionalities continue to be refined. New flags enable stateful map widgets and map entity notes, suggesting an upcoming feature where ChatGPT can directly annotate or attach reasoning to specific map locations. The system is also testing a strict limit for local search results and experimenting with localized home feed suggestions specifically for the Japanese and Indian markets, although these are currently disabled.

URL routing configurations show new endpoints for an embedded Codex application and direct installation scripts for Codex, reflecting the doubling of usage limits for Codex announced this week.

Web run

The role-based automation tools observed last week (targeting engineering, human resources, legal, and other professions) are evolving. The configurations now append a "personalized" tag to these tools, such as personalized inbox summarization and project plan reviews. A new specific capability has been added for engineering and data science roles to automatically triage GitHub pull requests.

Integration with third-party services continues to expand, with the addition of a specialized connector explicitly named for the legal sector.

Retrieval and grounding

The system prompt for real-time voice mode has been completely rewritten to support a dual-model architecture. The executing model is now instructed to act as a backend agent that receives ongoing transcripts from a frontend voice model. Crucially, it is directed to begin investigating—searching code, files, or connected apps—proactively while the user is still exploring a topic, rather than waiting for a final prompt. It must send commentary messages back to the frontend to ground the ongoing conversation in concrete evidence, reducing latency.

The "Pages" (or Canvas) feature is expanding its document templates. New configurations include a product management feature backlog and an educational learning plan. The system is explicitly instructed to populate these templates by retrieving context from the user's past conversations and connected applications.

The homework helper widget, which saw fluctuating configurations last week, is now explicitly routed to use a specific GPT-6 model variant (chat-luna-ev3-gpt6-v1), leveraging the GPT-6 Luna model that is already available on the API.

For image generation, new layers indicate preparation for C2PA edit provenance, which will likely embed metadata and watermarks into images that have been edited or inpainted by the model.

Ads and commerce

A comprehensive global pricing matrix has been deployed in the configuration, confirming a new structure of subscription tiers: go, prolite, plus, pro, promax, and pro5x. The pricing is dynamically adjusted based on whether a country is categorized as having a low or high average revenue per user.

Regarding advertising, the platform is testing a new consent modal titled "Make ads relevant for you," alongside adjustments to how long late-loading ad events are kept visible in the interface.

To monitor

A highly significant configuration change indicates that OpenAI is preparing to disable the creation of custom GPTs. The flags explicitly mention disabling GPT creation and set a retirement date for December 11, 2026. This suggests a major upcoming shift in how personalized AI agents are handled on the platform.

The claude_import tool remains active in the role-based configurations, indicating that a feature allowing users to migrate data from Anthropic's ecosystem is still under active development.

Finally, background configurations for native presentation and slide creation continue to be adjusted, though several user-facing introduction modals for this feature have been temporarily disabled.