Weekly summary · 2026-W43

ChatGPT experiments tracker › Summary archive

ChatGPT experiments summary 2026-W43

Daily archive of ChatGPT feature flags and experiments (public Statsig payload): new entries, timeline, weekly summary focused on search, web run, retrieval and ads.

2026-W43 · 2026-10-19 → 2026-10-25

2348
feature gates
490
dynamic configs
880
layers
2026-11-17
last update
Powered by RESONEO
2

The summary

Generated every Sunday evening from the week's new entries.

Highlights of the week

The configuration data from this week indicates a strong focus on workspace automation, specialized scientific environments, and expanded document creation capabilities. The introduction of the Rosalind Workbench aligns with the recent availability of the GPT-Rosalind models. Additionally, extensive routing logic for subagents and new document types suggest an expansion of ChatGPT's role as a comprehensive enterprise work environment.

Focus by theme

Search

A detailed configuration outlines a persistent task coordinator utilizing the gpt-5.6-terra model. It defines routing logic for different types of work, explicitly delegating asynchronous tasks requiring browser control or computer use to subagents. It also establishes rules for project-bound work, using Codex threads for repository edits, continuous integration diagnosis, and pull requests.

A new handoff tool has been configured to redirect users from ChatGPT to specialized execution environments when a request involves local coding, complex financial modeling, or computer-use automation.

URL routing configurations reveal endpoints for features such as deep research, workspace agents, WhatsApp and Slack integrations, and parental controls.

Workspace permission layers now include a broad list of model identifiers, such as GPT-4.1, GPT-5.1, GPT-5.2, o3, o4-mini, and gpt-lifesciences. While their presence in the configuration does not confirm their active deployment, it indicates the structural groundwork for granular access control across a wider range of model generations.

Web run

Task suggestions and unread automations are being prioritized within the user interface for both chat and work surfaces.

Configurations for third-party connectors have been expanded to include a comprehensive list of enterprise domains, such as Google Workspace, Microsoft 365, Salesforce, Notion, Slack, and Atlassian.

Retrieval and grounding

Supporting the new Rosalind Workbench, specific tags have been introduced to filter scientific literature retrieval, including systematic reviews, randomized controlled trials, and observational studies. The workbench itself is designed to connect scientific tools, databases, and sequence viewers.

The platform is testing native support for new document formats identified internally as granola_workbook, granola_presentation, and granola_whiteboard. This is accompanied by a suite of presentation templates, such as Product Strategy, Weekly Business Review, and Annual Impact Report.

Following last week's configurations for goal tracking, a new system prompt instructs the model on how to interactively gather necessary details from users, such as check-in dates, before saving an automated goal. The model is instructed to offer concrete choices and wait for user answers before executing the automation.

The gpt-4o-mini model is now explicitly configured for batch dictation transcription, alongside speculative language hints to improve streaming dictation.

Ads and commerce

Configurations for elevated ad cards and conversation ad hydration remain active, indicating continued testing of sponsored content within the chat interface.

Following last week's introduction of auto-reload mechanisms for credit balances, new layers show a transition in checkout providers and the implementation of embedded credit checkouts.

A complex pricing matrix has been deployed, defining costs for various subscription tiers (go, prolite, pro, plus, promax) across multiple countries. These countries are categorized by low and high average revenue per user to adjust regional pricing dynamically.

To monitor

The flashcard widget, which was actively being refined last week, has had several of its specific feature flags disabled, though the onboarding flip remains active. This suggests a potential pivot or consolidation of the educational tool.

OpenAI is testing role-based capabilities that map specific professions, such as engineering, human resources, legal, and life sciences, to tailored tools. These tools include inbox summarization, data science charting, and a specific claude_import function, which suggests preparation for a feature allowing users to migrate data or conversations from Anthropic's platform.

The image generation pipeline continues to show adjustments, with the strict inpaint prompt being enabled and temporal prompt context caching being activated to optimize generation workflows.

3

Summary archive

Every archived week, immutable, in French and English, on its own page.

Price alerts for shopping and new map interfaces on ChatGPT This week's data reveals the preparation of shopping price alerts, new interactive map hover cards, and dynamic layout updates for the Pages feature. 2026-W46 2026-11-09 → 2026-11-15 66 items with substance ChatGPT Business workspaces and new dashboard format for Pages OpenAI is preparing a new Business offering with ad credits for workspaces. The Pages feature also gains a dashboard format and direct-save options. 2026-W45 2026-11-02 → 2026-11-08 55 items with substance Retirement of custom GPTs and proactive ChatGPT voice mode The code reveals the planned retirement of custom GPTs in December. ChatGPT is also testing a proactive voice agent and new subscription tiers. 2026-W44 2026-10-26 → 2026-11-01 44 items with substance ChatGPT experiments summary 2026-W43 2026-W43 2026-10-19 → 2026-10-25 97 items with substance Launch of ChatGPT Sites and new background tracking agents The code reveals the rollout of ChatGPT Sites for web artifacts, the addition of background agents, and predictive ad preloading based on click rates. 2026-W42 2026-10-12 → 2026-10-18 33 items with substance ChatGPT tests computer control and deep research plugins Hidden configurations reveal new ChatGPT plugins for computer use and deep research. An internal payment system is also tested to replace Stripe. 2026-W41 2026-10-05 → 2026-10-11 33 items with substance Dots autonomous agents and expanded ChatGPT logged-out limits The configuration reveals instructions for the new Dots autonomous agents and removes the wallet plugin. ChatGPT now allows fifteen logged-out messages. 2026-W40 2026-09-28 → 2026-10-04 9 items with substance Pinned composer ads and background agent controls in ChatGPT ChatGPT tests pinned ads above the composer, frames silent background automations, and introduces watermark detection for images generated by Gemini. 2026-W39 2026-09-21 → 2026-09-27 37 items with substance ChatGPT overhauls connectors and introduces an ads manager OpenAI shifts ChatGPT connectors toward an advertising manager, tests a payment-free Plus trial, and prepares a delayed generation system via Kafka. 2026-W38 2026-09-14 → 2026-09-20 37 items with substance Custom models retirement date and local caching in ChatGPT This weekly summary covers new local search caching features, the addition of a Lowe's advertising agent, and a retirement date for custom models. 2026-W37 2026-09-07 → 2026-09-13 44 items with substance ChatGPT adds autonomous web browsing and artifact creation ChatGPT system instructions now include autonomous browser actions and artifact generation. The GPT-5.6 Luna model is also deployed for text editing. 2026-W34 2026-08-17 → 2026-08-23 42 items with substance Financial data for ChatGPT Work and hidden GPT-5.6 models The configuration reveals financial data access for workspaces, the hiding of recent GPT-5.6 models, and the addition of a Wayfair advertising agent. 2026-W33 2026-08-10 → 2026-08-16 56 items with substance Expedia business agents and local search testing in ChatGPT The code reveals dedicated business agents for Expedia and the integration of Yelp for local search. Rich text editing features are also in preparation. 2026-W32 2026-08-03 → 2026-08-09 43 items with substance ChatGPT tests a personalized ad agent and Google integrations The latest configuration data reveals a minimum viable product for personalized ads, new Google Workspace integrations, and writing style personalization. 2026-W31 2026-07-27 → 2026-08-02 55 items with substance ChatGPT exposes future models and integrates medical data Configuration files reveal unreleased models like GPT-5-pro and o4-mini. ChatGPT also integrates official healthcare databases and new ad tracking tools. 2026-W30 2026-07-20 → 2026-07-26 58 items with substance Upcoming ChatGPT Work interface and enterprise integrations This summary details the upcoming ChatGPT Work interface connected to enterprise tools, native document generation, and a global billing infrastructure. 2026-W29 2026-07-13 → 2026-07-19 9 items with substance

Other sources and studies

Nov 2026 What Google sees in an image: the two vocabularies of its vision models, Knowledge Graph entities and the dimensions and values of Google Shopping filters, then the record, the 86 filter boxes and the signals that rank each image Nov 2026 Google’s 300 ranking signals, and the fifty that apply to your page: the 305 sets Google Search records for each result and each query, the 48 that apply to a web page, 31 of which score it, the rankers in a row and the 34 already serving AI Overviews, AI Mode and Web Guide Nov 2026 How Google spots good publishers, even small ones, from the site section score, topic authority and citations Google attaches to every blue link, all the way to the SERP and into AI answers, to the “human voices” pipeline behind the hidden gems Nov 2026 What Google personalizes, from the SERP to AI Mode: three questions at every AI Mode query, 392 kinds of profile data including profiles written for the models, Gmail read by AI Mode and the confirmation email that writes a brand into its customer’s profile Nov 2026 Google Shopping, from the SERP to AI Mode: what a product must prove to be shown, then bought, from four catalog scores (two of them drawn from clicks) and 82 filters to the five sheets the AI model reads and the universal cart Nov 2026 Twiddlers: what still moves a result after ranking, from the fan-out search plan, a Bard filter and a relevance oracle to E-E-A-T expertise, topic authority, spam and DMCA, with 120 twiddlers Oct 2026 Knowledge Graph: how Google feeds it, maintains it and makes it grow, from the passport every fact carries to the 194 access requirements that name its suppliers and the 8,679 collections that decide which requests an entity can fill Oct 2026 AI Overviews and AI Mode: what Google’s AI reads before it answers, from 138 result types to Knowledge Graph facts that reach the answer with no URL, with 9 internal schemas to browse Oct 2026 How Google matches pages to search intent: Intent Brain measures twenty broad intents from clicks for every query, language and country, Orbit sorts every page into 86 categories, with 15 internal schemas to browse Oct 2026 KScorer: how Google decides what a search is after, with the 56 signals annotated by its engineers (clicks that predict the rater score, the weight of the top 10, entities, context) and 15 internal schemas to browse More stuffs...
RESONEO