# D3 · ChatGPT Surfaces and Features
Choosing the right ChatGPT model and feature – models from Luna to GPT-6 Pro, Projects, canvas, search, deep research, study mode, memory, files, data analysis, images, voice, GPTs, scheduled tasks, sites, workspace agents and company knowledge.
import { Accordions, AccordionItem, Tabs, TabItem } from '@prosefly/astro-components';
This domain is worth **14% of the mock – roughly 8 of 60 items**. It tests one skill above all: given a task, pick the **right model and the right feature** instead of doing everything in a plain chat with the default model. The exam does not reward memorising a feature list; it rewards matching a capability to a need – search for current facts, deep research for a sourced report, a project for recurring context, data analysis for real computation, and the cheapest capable model for the job.
## What you need to know
ChatGPT is not one thing; it is a model picker plus a toolbox. The model choices in September 2026 range from the low-cost **GPT-5.6 Luna** through **Terra** and **Sol** (and **Sol Pro**), a lightweight **GPT-5 Thinking Mini**, up to **GPT-6 Pro** (powered by **GPT-6 Astra**) on higher-tier plans. On top of the model you choose features: Projects, canvas, Search, deep research, study mode, memory, file uploads, data analysis, image generation, voice, GPTs, scheduled tasks, Sites, workspace agents and company knowledge. Choosing well is the difference between a fast, reliable result and a slow, ungrounded one.
## Learning objectives
By the end of this page you should be able to:
1. Select an appropriate **ChatGPT model** for a task by capability and cost.
2. Distinguish **Search** from **deep research** and know when each applies.
3. Choose between a plain chat, a **Project**, **canvas**, and **memory** for holding context.
4. Route computation to **data analysis** and use **file uploads** correctly.
5. Pick the right feature for images, voice, recurring tasks and internal-data tasks.
6. Apply a single decision table to select a surface under time pressure.
---
## 3.1 Choosing a ChatGPT model
The models trade capability, speed and cost. Pick the least powerful one that clears the task.
| Model | Position | Reach for it when |
| --- | --- | --- |
| **GPT-5.6 Luna** | Lowest cost, fast | Clear, repeatable, high-volume tasks: extraction, classification, simple drafts |
| **GPT-5.6 Terra** | Pragmatic all-rounder | The everyday default for most knowledge work |
| **GPT-5.6 Sol / Sol Pro** | Higher capability | Complex, open-ended, higher-value work needing stronger reasoning |
| **GPT-5 Thinking Mini** | Lightweight reasoning | Quick tasks that still benefit from a little step-by-step thinking |
| **GPT-6 Pro (GPT-6 Astra)** | Top capability, higher-tier plans | The hardest end-to-end work: sustained reasoning, judgment, multi-tool |
Legacy models remain available, but for new work choose from the current lineup. The instinct to reserve is "use the most powerful model for everything" – it is slow and expensive for tasks Terra or Luna handle perfectly.
:::tip[Assessment signal]
"High-volume, repetitive extraction/classification" → Luna. "The hardest reasoning/judgment task" → GPT-6 Pro / Astra. "Everyday all-round work" → Terra. If a stem says `MOST cost-effective`, it wants the *cheapest model that still meets the requirement*, not the cheapest model outright.
:::
## 3.2 Search versus deep research
Both reach beyond the model's knowledge cutoff, but they answer different needs.
| | **Search** | **Deep research** |
| --- | --- | --- |
| Purpose | Quick, current answer with links | A structured, multi-source report |
| Time | Seconds | Minutes (it browses many sources) |
| Output | A short answer plus citations | A long, organised synthesis with citations |
| Use when | "What is the latest X?", a fact to confirm | "Compile a briefing on the competitive landscape" |
```text
need current info?
│
├─ a quick fact / a few links ─────────► Search
└─ a thorough, sourced report ─────────► deep research
```
Both are the answer to the D1/D2 knowledge-cutoff problem: when a task needs current information, add a retrieval feature rather than trusting model memory.
## 3.3 Holding context: chat vs Project vs canvas vs memory
A frequent D3 item asks *where* a piece of context should live.
| Surface | Best for | Lifetime |
| --- | --- | --- |
| **Plain chat** | One-off tasks; no reuse | The conversation only |
| **Project** | Recurring work sharing the same instructions and reference files | Persists across many chats in the project |
| **Canvas** | Iterating on one document or code file side by side with editing | The document you are building |
| **Memory** | Small, durable facts about you or your preferences to apply everywhere | Persists across all chats until you edit/delete it |
- Put **stable, reusable** material (a style guide, product facts, custom instructions) in a **Project**.
- Use **canvas** when you are drafting and revising a single artefact and want to edit inline rather than re-paste.
- Use **memory** for lightweight preferences ("I work in EU timezone", "prefer British spelling"), not for bulk documents. Memory is expanding, and memory-with-past-chats is noted as coming soon; do not overload it.
## 3.4 Files and data analysis
**File uploads** let the model read documents, spreadsheets, images and more. **Data analysis** goes further: it runs real code on your data to compute, chart and transform – the correct home for the exact arithmetic that the raw model gets wrong (D2).
Upload a PDF or doc and ask questions about it. The model reads the content into context. Good for summarising, extracting and answering from a document – but verify the extraction against the source.
Upload a CSV or spreadsheet and ask for totals, filters, pivots or charts. It executes real computation, so a sum or a chart is trustworthy in a way a model-guessed number is not. This is the fix for arithmetic and counting weaknesses.
:::tip[Assessment signal]
"Compute / total / chart this spreadsheet reliably" → data analysis (real computation), not a plain-chat estimate. "Summarise this PDF" → file upload, then verify.
:::
## 3.5 Study mode, image generation and voice
| Feature | What it does | Reach for it when |
| --- | --- | --- |
| **Study mode** | Guides you through a topic with questions and steps instead of just handing answers | You want to *learn*, not just get an answer |
| **Image generation** | Produces images from a description (also available with Thinking) | You need a visual: a slide graphic, a concept image |
| **Voice / voice with video** | Speak to ChatGPT and hear replies | Hands-free brainstorming, practising a conversation |
Study mode is worth calling out because it maps onto the Academy's own teaching stance: it acts as a tutor rather than an answer vending machine, which is exactly the posture the assessment rewards.
## 3.6 GPTs, scheduled tasks, Sites and workspace agents
These are the "beyond a single chat" surfaces.
| Surface | What it is | Use it when |
| --- | --- | --- |
| **GPTs** | Custom, shareable versions of ChatGPT with fixed instructions/tools | You repeat a specialised task and want a reusable, shareable helper |
| **Scheduled tasks** | ChatGPT runs a prompt on a schedule | A recurring digest, reminder or check |
| **Sites** | Publish content ChatGPT helps you build | You need a simple published page |
| **Workspace agents** | Delegated, multi-step work inside a workspace with oversight | A structured task you want carried out with review points |
| **Company knowledge** | ChatGPT answers grounded in your organisation's connected internal sources | Questions whose answers live in internal systems |
**Company knowledge** is the internal-data counterpart to Search: instead of the open web, it grounds answers in your organisation's connected sources – the right choice when the correct answer is inside the company, not on the internet.
```text
where does the answer live?
│
├─ open web, current ───────► Search / deep research
├─ inside our org's systems ─► company knowledge
├─ in this one document ─────► file upload
└─ in the model's training ──► plain chat (then verify)
```
## Decision framework
Use the **surface-selection table**: read down the "I need to…" column, take the surface on the right.
| I need to… | Use | Not this |
| --- | --- | --- |
| Confirm a current fact quickly | Search | Plain chat (stale, may hallucinate) |
| Produce a sourced multi-source report | Deep research | Search (too shallow) |
| Reuse the same instructions and files across many chats | Project | Re-pasting into each chat |
| Draft and revise one document inline | Canvas | Copy-pasting between chat and doc |
| Remember a small durable preference | Memory | A Project (overkill) |
| Compute or chart real numbers reliably | Data analysis | Asking the model to estimate |
| Answer from an internal company source | Company knowledge | Web search |
| Do the hardest reasoning task | GPT-6 Pro / Astra | Luna (under-powered) |
| Do high-volume simple extraction cheaply | Luna | GPT-6 Pro (over-powered) |
| Run a prompt on a recurring schedule | Scheduled tasks | Remembering to do it manually |
The value: it converts a fuzzy "how do I do this in ChatGPT?" into a one-line lookup that stops you defaulting to a plain chat with the top model for everything.
## Common mistakes
| Mistake | Why it happens | What to do instead |
| --- | --- | --- |
| Using the most powerful model for everything | It feels safest | Match the model to the task; Terra/Luna handle most work |
| Asking plain chat for current facts | It answers anyway | Use Search or deep research for anything time-sensitive |
| Using Search when a full report is needed | Search is faster to reach for | Use deep research for multi-source synthesis |
| Re-pasting the same context into every chat | Habit from single chats | Put reusable context in a Project |
| Trusting a model-computed total | The number looks right | Use data analysis for real computation |
| Overloading memory with documents | Memory sounds like storage | Keep memory to small preferences; use Projects/files for bulk |
| Searching the web for internal answers | Search is the default reflex | Use company knowledge for internal sources |
| Copy-pasting a doc between chat and editor | It is what you always did | Use canvas to edit inline |
## Scenario challenge
**Scenario.** Amara, a marketing lead, has four tasks before end of day: (1) confirm a competitor's just-announced pricing, (2) produce a five-page sourced brief on the category's regulatory outlook, (3) clean and total a 500-row campaign spend spreadsheet, and (4) answer a question about her own company's internal brand guidelines that live in a connected internal drive. She is about to do all four in a single plain chat with GPT-6 Pro because "it's the smartest".
**Expert reasoning trace.**
1. **Task 1 needs current facts.** A just-announced price is past any cutoff; a plain chat would either refuse or hallucinate. **Search** returns a quick, cited answer.
2. **Task 2 is a multi-source report.** Search is too shallow; **deep research** browses many sources and produces the organised, cited synthesis she needs.
3. **Task 3 is exact computation on 500 rows.** The model would guess a plausible-but-wrong total; **data analysis** runs real code on the file, so the totals and charts are trustworthy.
4. **Task 4's answer lives inside the company.** Web search would miss it; **company knowledge** grounds the answer in the connected internal source.
5. **Model choice.** GPT-6 Pro is over-powered and slow for the extraction/summarisation parts; Terra (or Luna for the simple extraction) is more cost-effective. Reserve the top model for genuinely hard reasoning, which none of these four strictly require.
**Exam-correct outcome:** Search for the price, deep research for the brief, data analysis for the spreadsheet, company knowledge for the internal question – and a right-sized model per task rather than GPT-6 Pro for everything in one chat.
## Assessment traps
| Trap | Why it is tempting | The discriminator |
| --- | --- | --- |
| "Use GPT-6 Pro for everything" | Top model feels safest | `MOST cost-effective` wants the cheapest capable model; Terra/Luna suffice for most work |
| "Plain chat can answer current questions" | It always replies | Current facts need Search / deep research; plain chat is capped at the cutoff |
| "Search is enough for a full report" | Both use the web | Deep research does multi-source synthesis; Search returns a quick answer |
| "Ask the model to total the sheet" | It produces a number | Data analysis runs real computation; the model estimates |
| "Store the whole handbook in memory" | Memory sounds like storage | Memory is for small preferences; bulk goes in Projects/files |
| "Web search for internal policy" | Search is the reflex | Company knowledge grounds answers in internal sources |
| "Re-paste context into each chat" | It works, sort of | A Project persists shared instructions and files |
## Practice questions
A. GPT-6 Pro (Astra).
B. GPT-5.6 Sol Pro.
C. GPT-5.6 Luna.
D. GPT-5 Thinking Mini.
**Answer: C.** Luna is built for clear, repeatable, high-volume tasks like classification and extraction at the lowest cost. A and B are over-powered and costly. D adds reasoning the task does not need.
A. Plain chat with the default model.
B. Search, which returns a current answer with links.
C. Memory.
D. Canvas.
**Answer: B.** A just-announced price is past the model's cutoff, so Search retrieves and cites the current figure. A would be stale or hallucinated. C and D do not fetch current information.
A. Search.
B. Deep research.
C. Image generation.
D. Scheduled tasks.
**Answer: B.** Deep research browses multiple sources and produces an organised, cited synthesis. Search (A) gives a quick answer, not a multi-source report. C and D are unrelated to research.
A. Re-pasted into each new chat.
B. In a Project with custom instructions and reference files.
C. In a single very long chat kept open forever.
D. In image generation.
**Answer: B.** A Project persists shared instructions and files across many chats, which is exactly the recurring-context case. A is wasteful and error-prone. C hits context limits and clutter. D is irrelevant.
A. Ask the model in plain chat to add the column.
B. Data analysis, which runs real computation on the file.
C. Search.
D. Voice.
**Answer: B.** Data analysis executes real code, so totals and charts are trustworthy, unlike a model-estimated number. A is exactly the arithmetic weakness to avoid. C and D do not compute.
A. Search the open web.
B. Company knowledge, which grounds answers in connected internal sources.
C. Image generation.
D. GPT-6 Pro with no tools.
**Answer: B.** Company knowledge is designed to answer from the organisation's connected internal sources. A searches the wrong place. C is irrelevant. D relies on training data that does not contain internal documents.
A. Study mode.
B. Scheduled tasks.
C. Sites.
D. Data analysis.
**Answer: A.** Study mode guides the learner with questions and steps rather than simply giving the answer. B, C and D serve unrelated purposes.
A. Memory.
B. Canvas.
C. Search.
D. A GPT.
**Answer: B.** Canvas is the side-by-side editing surface for iterating on one document. A stores small preferences. C fetches facts. D is a reusable custom assistant, not an inline editor.
A. Recurring weekly digest → scheduled tasks.
B. Reliable total of a CSV → data analysis.
C. Current news headline → memory.
D. Bulk internal handbook stored → memory.
E. Multi-source report → image generation.
**Answer: A and B.** A scheduled task runs a recurring prompt on a schedule, and data analysis computes reliably on a file. C should use Search. D should use a Project/files. E should use deep research.
A. GPT-6 Pro is over-powered and costlier for simple extraction or summarisation.
B. Terra is a pragmatic all-rounder for most everyday work, and Luna suits high-volume simple tasks.
C. Only GPT-6 Pro can access Search.
D. Cheaper models cannot produce correct answers.
E. Model choice has no effect on cost or latency.
**Answer: A and B.** Matching the model to the task saves cost and latency, and Terra/Luna cover most work well. C and E are false. D is false – cheaper models are correct on suitable tasks.
A. A one-off chat.
B. A custom GPT.
C. Canvas.
D. Search.
**Answer: B.** A custom GPT packages fixed instructions and tools into a reusable, shareable assistant. A does not persist. C is for document editing. D fetches facts.
A. Trust it because the file was supplied.
B. Use file upload to read the document, but verify the stated figure against the source text.
C. Switch to image generation.
D. Store the PDF in memory instead.
**Answer: B.** File upload puts the document in context, but extraction can still misread, so verify the specific figure against the source. A over-trusts extraction. C and D do not address summarising the document.
## Key takeaways
- Pick the least powerful model that clears the task: Luna for high-volume simple work, Terra as the all-rounder, Sol/Sol Pro for complex work, GPT-6 Pro (Astra) for the hardest reasoning.
- **Search** for a quick current fact with links; **deep research** for a multi-source sourced report.
- Hold reusable context in a **Project**, iterate one document in **canvas**, keep small preferences in **memory**.
- Route exact computation to **data analysis**; verify summaries from **file uploads** against the source.
- Use **company knowledge** for internal answers, **Search** for the open web.
- GPTs, scheduled tasks, Sites and workspace agents extend ChatGPT beyond a single chat.
- The right question is always "where does the answer live and what is the least effort to reach it reliably?"