What’s new in GPT‑5 — and how it changes ChatGPT
Here’s what actually matters for teams: the new router, clearer status updates, bigger context, better memory in Projects, and built‑in connectors.
TL;DR — the five real changes
1) Auto‑switching (Router)
ChatGPT now picks between a fast model and a deeper “thinking” model for you. Force ‘Thinking’ only for high‑stakes or messy tasks.
2) Status updates (Preambles)
On tool‑heavy tasks it shares a plan and progress updates, so you can spot mistakes before they ship.
3) Projects + Memory
Stable facts persist across chats inside a project. You can view, edit, or delete what it remembers.
4) Connectors
Search and summarize files from Drive, SharePoint, GitHub, Gmail/Calendar, and more — with approvals before any send.
5) Bigger context
Work across larger inputs and longer interactions (up to 400K tokens depending on the variant) with stronger tool use.
Key differences vs GPT‑4o & OpenAI o3
| Area | GPT‑4o / o3 | GPT‑5 (what changed) |
|---|---|---|
| Model selection | Pick models manually | Auto‑router chooses fast vs thinking; you can override when needed |
| Status updates | Sparse/implicit | Built‑in preambles that share plan + checkpoints on tool‑heavy runs |
| Memory | Per‑chat or basic saved notes | Project‑scoped recall across chats; user‑reviewable memory controls |
| Connectors | Limited workflows | Deeper set (Drive, SharePoint, Dropbox, Box, Outlook, Gmail, Google Calendar, GitHub, Linear, HubSpot, Teams) with approvals |
| Context window | Typically 128–200K | Up to 400K tokens (variant‑dependent) with better long‑context retrieval |
Where you’ll feel it in ChatGPT
Single default model
ChatGPT defaults to GPT‑5 with auto‑switching. You’ll override to Thinking for multi‑step reasoning, synthesis, tricky math, or high‑stakes reviews.
Clearer long runs
Expect short plans and checkpoint prompts (“Proceed to email draft?”). Say yes/no to stay in control.
Projects remember
Within a project, prior chats + saved memories are referenced. Review or delete anything you don’t want carried forward.
Files from your tools
Pull files directly from Drive/SharePoint/etc., summarize, then pause for approval before any outbound send.
5‑minute setup (non‑tech)
- Add to Project instructions: “Answer in 1–3 sentences unless I ask. If facts are time‑sensitive or uncertain, browse and cite or abstain.”
- Pick one real task and use the default (auto) model. Only force Thinking for edge cases.
- Enable Connectors in a sandbox; require approvals before any send.
- Keep a single thread from outline → draft → QA. Store stable facts as memories inside the project.
- For news/finance/health, require browse + two sources and a short verification step.
Governance quickstart (lightweight)
Policies to turn on
- Memory: store only stable, non‑sensitive facts; monthly review.
- Browsing: mandatory for time‑sensitive claims; cite dates.
- Connectors: least‑privilege scopes; approvals before outbound.
- Incidents: abstain → verify → escalate; log and retrain prompts.
Team KPIs
- Re‑prompts per task ↓
- Time‑to‑first‑usable draft ↓
- Unsupported‑claim rate ↓
- Copy‑pastes avoided/week ↑
FAQ (the stuff people ask)
Do I still need to pick models?
Not usually. Start with the default (auto). Override to Thinking when accuracy and reasoning matter more than speed.
Will it still hallucinate?
Less often, yes. Not never. For anything consequential, browse, cite, and verify.
Is Memory private?
You control saved memories: view, edit, or delete. Project recall is scoped to that project.
Are connectors safe?
Use least‑privilege scopes, approvals before outbound actions, and keep audit logs in your wiki.