✓ Verified on September 3, 2026⏱ 4 min read✓ done
Claude's three modes
🎯 Example — Preparing for a Board meeting
Chat: "Analyze this memo and identify the top 5 risks."
Cowork: "Take the Q1 data, build an Excel dashboard and a 10-slide deck for the Board."
Code: "Deploy a web dashboard connected to our internal API."
Cowork: "Take the Q1 data, build an Excel dashboard and a 10-slide deck for the Board."
Code: "Deploy a web dashboard connected to our internal API."
Which one to choose? — decision table
The question isn't "which is best" but "which one for which situation." Simple rule: choose based on the deliverable and the level of autonomy you expect.
| Your need | Mode | Why |
|---|---|---|
| Think, analyze, draft, get an opinion | 💬 Chat | A quick conversation, no file to produce. Your default reflex. |
| Produce a deliverable from your files (Excel, PPT, doc, folder) | ⚡ Cowork | Claude accesses your files and executes the task end to end. |
| Chain several apps / automate a recurring task | ⚡ Cowork | Multi-tool orchestration (email, calendar, SaaS) + scheduled tasks. |
| Build or deploy a custom tool (app, script, integration) | ⌨️ Code | Terminal access, version-controlled code, deployment. For technical teams only. |
| You're away from your desk but a task is running | 📱 Dispatch | Drive Cowork from your phone while the Mac does the work. |
📌 The right reflex
Start in Chat. The moment you catch yourself copy-pasting a result into a file, switch to Cowork: that's the sign a deliverable needs to be produced. Reserve Code solely for development projects handed to your technical teams.
Which model? — the guide's single rule
A "model" is the brain you select at the top of the conversation. Remember one rule, valid everywhere in this guide:
Sonnet 5Everyday work: emails, notes, routine analysis. The default model, the cheapest.
Opus 5Heavy files: reasoning, long assignments, agents. It verifies its own work. Default on Max.
Fable 5.1The summit: thick files, strategic analysis, multi-hour missions. Twice the price of Opus 5: use it deliberately.
| Model | Working memory | What for |
|---|---|---|
| Fable 5.1 ✦ | 1M tokens ≈ 555,000 words | The most powerful (Mythos class): thick files, long assignments, strategic analysis, autonomous agents |
| Opus 5 | 1M tokens | Close to Fable at half the price. Default on Max, the strongest model on Pro. Verifies its own work |
| Opus 4.8 | 1M tokens | Previous generation, still available — takes over from Fable on some sensitive topics |
| Sonnet 5 | 1M tokens | Everyday work: emails, notes, routine analysis. The default model — close to Opus 4.8 for 2 to 3× less |
| Haiku 4.5 | 200K tokens | Small high-volume tasks where speed matters most |
📌 Three useful details
Fable 5.1 formats less and writes denser: ask explicitly for headings and lists when you want them. On sensitive topics (cybersecurity, biology), Fable automatically falls back to Opus — you are told, and nothing is billed. And stop adding "check your work" to your prompts with Opus 5 or Fable: these models already verify themselves, the instruction only burns tokens.
Access by plan
API and usage-based Enterprise: Fable available without restriction ($10 / $50 per million tokens). Max and Team Premium: Fable included up to 50% of your weekly limits. Pro and Team Standard: access through paid usage credits. Detailed prices are in the Pricing appendix. [Terms may change — check claude.com/pricing]
AskUserQuestion — Claude asks you
A key new feature: Claude no longer guesses when information is missing. With AskUserQuestion, it asks you targeted questions (clickable buttons) before getting started, rather than producing an approximate result based on assumptions.
🎯 Why it's a game-changer
Instead of missing the mark for lack of context, Claude clarifies: "For this report, who's your audience — the executive committee, investors, or an internal team?" You click, and it produces the right deliverable the first time. Fewer back-and-forths, more precision.
🛠️ Your turn — 5 minutes
Ask the same thing twice: once in Chat on claude.ai ("analyze this text"), once in the Desktop app pointing at a file. Note what differs.
Expected result : You can say in one sentence when to stay in Chat and when to delegate to Cowork.