// The guide

Claude.ai: the complete guide (2026).

2026-07-15 · 6 min read · by Nabil BA-MOH

You hit the usage limit again. Mid-thought, mid-project, with a wall of "you've reached your limit" where your answer should be. And nobody ever explained the rule you just broke — because it isn't written where you'd look.

This guide covers everything claude.ai offers in 2026 — plans, models, Projects, memory, connectors, Research mode. But its real job is the last section: how the usage limits actually work, and the session hygiene that makes the same subscription go two or three times further.

The plans, decoded

• Free — real access, not a demo: web search, memory, file uploads, Projects in read-only quantities. Enough to know if Claude fits you.
• Pro — $17/month billed annually ($20 monthly). Unlimited Projects, Research mode, access to Opus 4.8, Claude Code and Cowork. The default answer for one person doing real work.
• Max — from $100/month, for people who live in Claude all day. Same features as Pro, multiples of the usage.
• Team — per-seat plans (from around $20/seat billed annually, with a premium seat option), admin controls, self-serve since February 2026.
• Enterprise — seats plus API-rate usage, model entitlements, HIPAA self-serve as of July 2026.

One thing most people miss: your limits are shared across every surface — chat, Claude Code, Cowork, mobile. One pool, several doors.

The models: who you're actually talking to

• Sonnet 5 (June 2026) — the default on Free and Pro. Fast, capable, the daily driver.
• Opus 4.8 (May 2026) — the heavyweight, available from Pro up. Use it for hard reasoning, long documents, high-stakes writing.
• Haiku 4.5 — the fast, light one.
• Fable 5 (June 2026) — Anthropic's new Mythos-class tier above Opus. As of this writing it runs on usage credits after an introductory period for paid plans.

Rule of thumb: stay on the default; escalate to Opus when the task is worth more than your patience.

Projects: stop re-uploading the same files

A Project is a workspace with persistent knowledge — upload your documents once, set instructions once, and every conversation inside the Project starts pre-loaded. Unlimited from Pro, and shared between chat and Cowork, so your agent sessions see the same context your chats do.

If you paste the same background into every new conversation, you're doing manually what Projects do automatically — and paying for it in tokens (see the limits section below).

Memory: Claude remembers you now

Since March 2026, memory is on for everyone, including Free. Claude carries what it learns about you and your work across conversations — preferences, ongoing projects, context you'd otherwise repeat. There's also a Monthly recap (July 2026) summarizing what you worked on.

It's genuinely useful, and it has a shape worth knowing: it lives inside claude.ai, on Anthropic's side, scoped to the surface you're chatting in. What it learned about you doesn't travel to other tools.

Connectors: Claude reads your other tools

Connectors (built on MCP, an open standard) plug Claude into Google Drive, calendars, and — since July 2026 — Microsoft 365 with write access. Remote MCP support (January 2026) means web connectors without installing anything. Ask "summarize this week's meeting notes from Drive" instead of downloading and uploading files like it's 2024.

Research mode: the report writer

From Pro up, Research mode runs multi-step web research — many searches, source evaluation, a synthesized cited report. Minutes of waiting instead of an afternoon of tabs. Use it for "compare X, Y, Z with sources"; use plain web search (free, instant) for quick lookups.

The usage limits, finally explained

Here's the section that pays for the scroll. Claude's limits have two layers:

1. A 5-hour rolling window — your short-term budget, which resets as time passes (capacities doubled in May 2026).
2. A weekly cap — the long-term ceiling on top.

Both are shared across chat, Claude Code and Cowork. You can watch the bars in Settings → Usage.

Now the part nobody tells you: what burns quota is not messages, it's context. Every turn, Claude re-reads the entire conversation so far. In a fresh chat, your question costs your question. In a 150-message chat full of pasted documents, that same one-line question drags the whole history behind it — you pay for everything, every turn. Long dirty sessions are quota furnaces.

Session hygiene, in four habits:

1. One topic, one chat. When the subject changes, start a new conversation. Old chats stay searchable; they just stop billing you.
2. Put stable context in a Project, not in the chat. Project knowledge doesn't ride along as conversation history.
3. Harvest and restart. When a long chat produced its answer, ask for a summary of decisions, carry that into a fresh chat, and move on.
4. Match the model to the task. Opus on the default for everything is like taking the truck to buy bread.

People who do this stop hitting limits. Same plan, different discipline.

The ceiling: memory that stays in the app

FAQ