Summary of "GPT-5.5: 9 Features You're Not Using (Full Walkthrough)"

Overview

The video argues that GPT-5.5 is more than a “smarter chatbot.” Its biggest value comes from enabling advanced features that let it:

What it does: Automatically browses the open web, pulls documents, reads them, and returns a full report with clickable citations.
How to enable/use: In ChatGPT, open the mode dropdown → “Deep research”, then ask a question (e.g., “Summarize the latest developments in renewable energy research with references”).
Timing: Not instant—~5 to 30 minutes. You can leave and return later.
Limits: Compute-heavy. Quotas depend on plan (approx: Pro ~250 runs/month, Plus/Team ~25, Free ~5 then downgraded). Treat each run as valuable.
If it stalls: Switch to agent mode with a similar prompt.

What it does: Goes beyond reading—uses a tool toolkit (text/visual browser, code terminal, and app connectors) to complete tasks step-by-step.
How to enable/use: Select “agent mode” from the model switcher; give it a goal and it narrates actions.
Example task: Plan a themed dinner party with recipes, a shopping list, and a schedule.
Safety behavior: Won’t perform sensitive actions (purchases, transfers, other high-consequence actions) without user confirmation.
Control commands: Use “stop” or “undo last action.”
Setup requirement: You must link connectors in settings first, or the agent can’t access external apps.

What it does: Generates images with legible text, retains fine detail, and supports multi-language output.
How to enable/use: Enter an image prompt or use the image button. On Plus/Pro, enable beta “images with thinking” in settings to improve composition planning—especially for complex graphics like infographics.
Practical tip: Provide detailed visual direction (style, lighting, and the exact text). If a label is wrong, correct it explicitly, e.g., “That sign is wrong. Fix it.” (it regenerates with corrected text).

What it does: Reaches the live web for current details (news/prices/etc.) and can include citations.
How to prompt: “What are the most recent developments in AI chips? Include links to sources.”
User habit: Always check the sources list, since browsing can be slower and may misread pages.

What it does: Lets you upload files (CSV/Excel-like) and ask questions. It writes/executes Python and returns results like charts and explanations.
Common uses: Trend analysis, data cleaning, visualizations, math, quick statistical work.
Error handling: If code errors occur, it attempts fixes automatically.
Practical limit note: For large jobs, trim the dataset to avoid timeouts.

What it does: Stores details from past conversations and uses them to tailor future answers.
How to enable/use: Turn on memory in settings; save facts such as diet, profession, training goals, etc.
New transparency concept: A “memory sources” panel shows which stored memory the model used, and users can remove incorrect/outdated items.

What it does: On Plus/Pro, connects to real apps (e.g., Gmail, Google Calendar, GitHub, Slack) so the agent can fetch and act on information.
How to make it reliable:
- Authorize connectors in settings first.
- Start requests with an action verb (e.g., “find” or “fetch”) to prompt “lookup” behavior.
Examples: Check Friday schedule, find emails about a project.

What it does: Accepts uploaded images (charts, screenshots, diagrams) and answers questions about them.
Capabilities emphasized: Better reading of text within images and interpreting graphs.
Boundary: Works with images, but not audio/video.

What it includes: Heavy red teaming and refusals for genuinely harmful requests.
Agent risk control: Requires confirmation before high-consequence actions (e.g., moving money).
Practical reminder: It can still be wrong—review outputs before acting.

Deep research: “Analyze the impact of electric vehicles on energy consumption with references.”
Agent mode: “Organize my travel itinerary for Paris, including flights, hotel, and a daily schedule.”
Images 2.0: “Generate an image of a fantasy dragon in a sci-fi city at night, highly detailed.”
Code interpreter: “Feed code interpreter a sales file and ask it to calculate the total and plot the trend.”