Claude App: a practitioner’s field guide.

The five surfaces, one model

"The Claude app" isn't a single thing. Anthropic ships five front doors to the same underlying model, and most of the friction people hit comes from assuming the door is the room. The doors are:

• Web at claude.ai: the canonical surface. Everything else lags this by some amount, except where the OS gives them an advantage.
• Mac desktop app, a thin Electron-style wrapper over the web with menu-bar integration, a global hotkey, and the one surface where Claude can touch your local files and apps (folder access, browser use, computer use).
• Windows desktop app, the same wrapper for Windows, with the same local-machine capabilities and a different system-tray idiom.
• iOS app, a native iPhone/iPad client with Share Sheet, App Intents, Siri Shortcuts, widgets, and a few Apple-only capabilities.
• Android app, a native Android client with system share intents, widgets, and Assistant handoff.

Inside all five, you're talking to the same family of models. As of September 2026 that's the Claude 5 family: Sonnet 5 as the everyday model on every plan (Free included), Opus 5.5 on paid plans, Fable 5.1 at the top end, and Haiku 4.5 for speed. The model picker is the same picker everywhere. What differs is which models the picker offers, which is a function of your plan, not the surface. Part 03 has the full lineup and when to reach for each.

The piece most people get wrong is treating "Claude on my phone" and "Claude on my laptop" as different products. They aren't. Conversations sync. Projects sync. Memory syncs. Artifacts you open on desktop you can hand back to mobile mid-thread. Since July, agentic tasks (what used to be desktop-only Cowork) sync too. The only thing that doesn't sync is voice-mode audio, and that's deliberate: only the transcript is kept.

What this guide is going to do is walk through the four app surfaces (skipping web, which is just the default), call out where each one rewards or punishes you, and end with the limits, etiquette, and troubleshooting playbook that applies across all of them.

Plans, briefly

You can't talk about app capabilities without talking about plans, because most of the "this doesn't work for me" reports trace back to plan ceilings rather than app bugs. Here's the September 2026 picture, simplified:

One thing the table doesn't capture: usage limits are 5-hour rolling windows, not daily, with a weekly limit layered on top for paid plans. If you have a heavy hour at 2pm, you'll have less headroom at 5pm but full headroom by 7pm. Usage is shared across claude.ai, the desktop and mobile apps, and Claude Code, and paid plans can now buy usage credits to keep going past a limit instead of waiting. The mobile and desktop apps are still stingy about telling you where you are in the window. That asymmetry catches people out, especially on long voice-mode walks.

Models: which one, when

The model lineup turned over almost completely between May and September. If your mental map still says "Sonnet 4.6 by default, Opus 4.7 when it's hard," it's two generations stale. Here's the current family, and it's a genuinely big step up:

What actually got better

• Sonnet 5 (June 30) is the one most people will feel, because it's what Free and most daily chats run on. Anthropic calls it its most agentic Sonnet yet; the biggest gains over Sonnet 4.6 are in coding and multi-step, tool-using tasks, and the context window jumped to 1M tokens. It's also the first Sonnet with real-time cybersecurity safeguards, so the occasional refusal on security topics is by design.
• Opus 5.5 (September 22) is the headline. Near-Fable quality on most work, and output more than 30% faster than Opus 5. Two things I care about as a daily user: it writes more tightly, putting the most important information up front instead of burying it, and Anthropic reports it is much less likely than recent models to take hard-to-reverse actions. That second point matters once you're letting Claude do work, not just talk.
• Fable 5.1 (September 1) is stronger on long-running agentic work, multi-step research, and document, spreadsheet, and slide work. Early testers describe it as staying readable over long, multi-step tasks. For high-risk cybersecurity or biology requests it automatically re-runs the request on an Opus model and tells you it did (you can turn that off in Settings → Capabilities).

Effort and thinking

The old "turn on extended thinking to get Opus" dance is gone. The model menu next to the send button now controls three things: the model, the effort level (Low, Medium, High, Extra high on Opus 4.7 and newer, and Max), and whether thinking is on. On Opus 5.5, Fable 5.1, and Opus 5, thinking can't be turned off; effort is the dial. The picker remembers your last choice, so the next conversation starts on whatever you used last.

Mac vs Windows

The desktop apps look almost identical. They are not identical. The differences come from what each OS lets a client app do, and they cluster around: launch behavior, file access, voice, and the keyboard.

Install & update model

Both ship as native installers. The Mac app self-updates silently in the background; you'll notice when a new feature shows up overnight. The Windows app prompts you on launch when an update is available, and you have to click through. If you're on Windows and noticing your friend's Claude app does something yours doesn't, check the version number first.

Distribution: Mac is .dmg from claude.com/download; Windows is .exe from the same place. Neither is in the App Store or the Microsoft Store, which means corporate-managed devices may need IT involvement.

Where it lives

The hotkey difference is the one that matters in practice. ⌘+Shift+Space on Mac collides with nothing common; Alt+Space on Windows collides with the legacy "system menu" command on focused windows. If you're a heavy keyboard user on Windows, change the hotkey first thing in Settings → General → Hotkey.

File pickers and Computer Use

Both apps can read files via drag-and-drop, the attachment button, and (on Pro and above) Computer Use, which lets Claude operate your file system and other applications under permission.

Voice mode

Voice mode is one feature that's reached parity across all four apps in early 2026, but the desktop builds expose it differently. Both Mac and Windows put voice mode behind the microphone icon next to the input box; both default to push-to-talk and let you switch to continuous listening from the in-call menu. You pick from a small preset set of voices, and voice mode is available on every plan, Free included. Two current gaps: Fable models aren't available in voice mode, and in agentic (Cowork) tasks you get dictation, not full voice mode.

The differences:

• Mac uses the system microphone the OS picks unless you override in System Settings → Sound. AirPods Pro switching is handled at the OS level and works reliably.
• Windows respects the default communication device, which can lag the default playback device by a beat. If voice cuts in but Claude can't hear you, your default microphone is wrong. Fix it in Settings → System → Sound.
• Mac shows a menu-bar indicator while voice mode is hot; Windows shows a tray badge and a tooltip. The Windows indicator is easier to miss.

Cowork and the desktop apps

Claude Cowork went GA on April 9, 2026 as a separate mode inside the desktop apps. It didn't stay separate for long. On July 7 it came to the web and mobile apps with cloud-run sessions (beta), and on September 16 Anthropic merged Cowork and chat into one Claude: there's no mode to pick any more. You describe what you want and Claude decides whether to just answer or take on the bigger job (research, a spreadsheet, a deck) and hand back finished files. Your chats, Cowork tasks, projects, connectors, and skills carry over, and Cowork tasks now show up in Recents alongside chats. The merge is rolling out to Pro and Max first, across web, desktop, and mobile, with more plans to follow; Enterprise admins get at least 30 days' notice. See the Cowork guide for the full treatment.

What still needs the desktop app: anything that touches your machine. Reading and writing local files and folders, browser use, and computer use (beta, Pro and Max only) all require Claude Desktop to be open and connected. So on desktop you're never more than a click away from giving Claude permission to operate your machine. That's powerful, and a permission posture worth thinking about.

iOS vs Android

Mobile is where the per-platform differences get pronounced. Both apps share a code base for the chat surface itself, but the OS-integration features diverge sharply because Apple and Google give app developers different primitives to work with.

The basic chat experience

From inside a chat, the two apps are nearly indistinguishable. Same model, same Projects sync, same file uploads, same Artifacts rendering, same voice mode with the same voices, and (since July) the same ability to kick off and follow agentic tasks. Both run identically on phone and tablet form factors. If you're just having conversations, pick whichever phone is in your pocket.

The fork shows up the moment you try to get content into Claude or do something with what Claude returned.

Share sheets & getting content in

The iOS Shortcuts story is meaningfully better than anything on Android right now. If you're an automation-heavy user, that asymmetry is real. On the other hand, the Android Quick Settings tile beats anything iOS offers from a "fastest possible voice prompt" angle, with fewer taps from any app, including the lock screen.

Voice mode

Voice mode is the feature that benefits most from being on a phone. Both apps support continuous voice mode with interruption handling: you can talk over Claude, and it'll yield. Both apps offer the same voices, and voice can use connected tools like Gmail, Google Calendar, and Slack.

Differences worth knowing:

• iOS supports voice mode with screen off (background audio capability), which means you can keep walking and talking without the screen on. The phone's silent switch is honored: flip it and Claude responds in text only.
• iOS integrates with Siri Shortcuts for voice triggering: "Hey Siri, ask Claude…" works if you've set up the shortcut. The handoff is jankier than native Siri because Apple gates intents, but it works.
• Android requires the screen to remain on for continuous voice mode unless you've granted "Display over other apps," a permission that, fairly, freaks people out. Without it, voice mode pauses when the screen times out.
• Android's integration with Google Assistant is read-only: Assistant can hand off to Claude via the share intent, but Claude can't be the system-default assistant.

Camera, vision, and photos

Both apps let you tap a camera icon to take a photo and analyze it without leaving the app. Both let you attach existing photos via the standard photo picker. Both support multiple images per turn (up to plan limits, usually five).

Where they diverge: iOS uses the modern PHPicker, which means Claude only sees the photos you explicitly select. Your full library is never enumerated, even with permission. Android's photo picker is similar on Android 13+, but on older versions you grant broader Photos permission and Claude trusts you to pick wisely. The privacy posture is materially different.

Widgets

Net: iOS wins on lock-screen and StandBy surfaces; Android wins on Quick Settings. Pick whichever pattern fits your physical interaction with the phone.

Background limits and "where did my chat go"

Mobile OSes are aggressive about backgrounded apps, and Claude's behavior diverges:

• iOS rarely kills the Claude app outright unless memory pressure is extreme; conversations resume in place. Voice mode is treated as background audio and survives lock.
• Android's aggressive battery managers (especially Samsung One UI's "Sleeping apps" and Xiaomi's MIUI background killers) will silently kill the Claude process. Add Claude to your battery-optimization allowlist if you use voice mode regularly. Otherwise you'll get cut off mid-sentence.

Capability matrix

Most features are everywhere, but not all of them. Bookmark this table; it answers most "why doesn't X work on my…" questions.

The big change since May: the old "partial" marks on the web and mobile Cowork rows are gone. Since July you can start, follow, and schedule agentic tasks from any surface, and they run in the cloud so they keep going when your laptop is closed. What stays desktop-only is the stuff that needs your actual computer: local files, browser use, and computer use. The merged experience still has a few gaps, too: it doesn't support branching a conversation or "Add from GitHub" yet, there's no web search toggle (Claude searches when it helps), and Research moved to /deep-research or the + button.

Limits, and how they actually manifest

The headline numbers are easy to recite ("up to 1M context, 5-hour windows plus a weekly cap"). What's useful is knowing how those limits show up in practice across the apps.

Context window

On paid plans, the current models (Sonnet 5, Opus 5.5, Fable 5.1) get a 1M-token context window in chat; older models like Opus 4.8 get 500K, and the rest (including Haiku 4.5) 200K. With code execution turned on, Claude now automatically summarizes earlier messages as a conversation approaches the limit, so long chats keep going instead of hitting a wall. That's a real improvement, but a summary is still a summary: detail from early in the thread gets softer. Symptoms: it forgets a stipulation you made earlier, or starts repeating itself, or gives a worse answer to a question it nailed two turns ago.

The apps don't surface where you are in the context budget. There's no "65% used" bar. Watch for the symptoms, not a number. When you see them, start a new conversation and either re-paste the load-bearing parts or use a Project to share context.

Attachment ceilings

• Per-message: typically 5 files, total <100MB.
• Per-PDF: ~32MB and 100 pages on most plans; Enterprise raises both.
• Image limits: ~20MB per image, max 5 per turn. Mobile compresses uploads, so if you're sharing a screenshot of small text, prefer desktop.
• Project knowledge files (Pro+): aggregate budget separate from per-message uploads; Pro is roughly 100K tokens, Team raises it, Enterprise lets you set it.

Rate limits and the 5-hour window

Anthropic frames limits as a 5-hour rolling window plus a weekly limit on paid plans. What you burn per message moves with conversation length, the model (Opus and especially Fable cost more than Sonnet), the effort level, and the features you use: a long agentic task that builds a spreadsheet costs far more than a quick question. Free is the most variable; Pro is steadier; Max is rarely an issue in practice. On paid plans you can buy usage credits to keep going past a limit.

What you'll experience when you hit the cap:

• The web app shows a yellow banner with a countdown.
• The desktop apps show the same banner but truncated; on Windows it sometimes appears for one frame and disappears.
• The mobile apps show a modal dialog with the wait time.
• Voice mode degrades silently: it doesn't end the call, it transparently routes you to a smaller model, and the responses get noticeably less coherent. This is one of the most frustrating failure modes. The only fix is to end the voice call, send one text message to confirm the model fallback, and decide whether to wait or downgrade your expectations.

What's not on Free

• Opus 5.5 and the Fable models (Free gets Sonnet 5 and Haiku 4.5).
• Agentic (Cowork) tasks, scheduled tasks, and computer use.
• Claude Code.
• Research, Claude in Chrome, and the Microsoft 365 integration.
• Claude Design, Slides, and Docs (beta).
• More than five Projects.

Free got a lot better this year: Memory, connectors, Skills, code execution, voice, and a handful of Projects, all on Sonnet 5. It's a genuinely capable chat product. It's just not where the agentic work happens.

Best practices, by surface

The apps reward different patterns. Here's what holds up across months of daily use.

A note on privacy posture

This isn't a guide to Anthropic's privacy policy (that's their job to publish and yours to read), but a practitioner has to know what each app surface is sending where.

• On Free, Pro, and Max, whether Anthropic trains on your conversations is a setting you choose (you were asked at signup or via an in-app prompt). With it on, retention for new chats extends to five years; with it off, the standard 30 days applies. Check where you landed in Settings → Privacy.
• Team and Enterprise plans are commercial terms: no model training on your content by default, and admins control the rest centrally.
• Voice mode does not store audio, only the transcript. Conversations transcribed from voice look the same as typed.
• Memory's contents are visible to you under Topics in Settings → Memory; you can edit and delete entries. They're scoped to your account, not your team.
• Computer Use sessions log the actions Claude took, not the screen content. The action log is in Settings → Activity.

The most underrated control is Settings → Data Controls → Delete account data. It's a hammer, but it works as advertised. If you've ever shared something into a chat that you regret, you can purge it. (On Enterprise, the admin holds the same hammer with finer granularity.)

Troubleshooting playbook

Roughly in the order you'll meet them.

Closing thought

The single most useful frame for thinking about Claude across surfaces is: the model is the same; the integration is what differs. Mac gives you a menu bar; Windows gives you a tray; iOS gives you the Share Sheet; Android gives you Quick Settings; the web gives you the canonical surface where every feature lands first. None of them is "the real Claude." They're all real, with different ergonomic bargains.

Pick the surface that matches your physical relationship with the device, and let conversations flow across all of them. — TWD

If you're new to Claude, start on the web for a week to learn the affordances. Then add the surface(s) where you spend the most ambient time. Voice mode on a phone walk; agentic tasks on the laptop you actually work on, where Claude can reach your files; the mobile share sheet for whatever you're consuming on the go. The trap is treating Claude like an app you "switch to." Its strength shows up when it's the layer underneath whatever else you're doing.

And if any of this is out of date by the time you read it: Anthropic's release notes are where changes get posted. This summer moved fast (a whole new model generation and a product merge in under four months), so check them before trusting any specific claim here. Both desktop apps lag a few days; mobile lags a few weeks.