Claude: Chat vs Cowork vs Code

Team reference — which surface to use, and what it costs.
v1.0
May 2026
Print & laminate
The golden rule: Default to Chat. Move to Cowork when you'd describe the task as "go do this for me while I'm in a meeting." Move to Code only when modifying a real codebase. Always plan in Chat before executing in Cowork — saves up to 80% of the tokens.
💬
Chat
claude.ai · web & mobile

What it is

A conversation. You type, Claude replies. Web search, code execution, artefacts (HTML, MD, charts, docs). No filesystem access.

Use it for

  • Anything that fits a back-and-forth
  • Thinking, drafting, brainstorming, research
  • One-off artefacts (docs, slides, charts, mockups)
  • Quick questions and copy work
  • Planning before you delegate to Cowork

Don't use for

Tasks that need to read or write across many files on your computer, or that should run unattended for 10+ minutes.

Relative cost

CHEAPEST · 1×
Cowork
Claude Desktop app · tab

What it is

A delegated coworker. Reads scoped folders on your machine, runs multi-step tasks, scheduled jobs, uses plugins & skills, hands back a finished deliverable.

Use it for

  • Tasks across many files on your computer
  • Long-running work (10+ minutes autonomously)
  • Recurring scheduled tasks (daily digests, weekly reports)
  • Chaining several skills + connectors
  • "Build this exact thing" after planning in Chat

Don't use for

Quick questions, one-off drafts, or anything you'd babysit turn-by-turn. If you're not delegating, you've picked the wrong surface.

Relative cost

HEAVY · 10–30×
{ }
Code
Terminal · CLI tool

What it is

A terminal-based engineering assistant. Reads codebases, edits files in place, runs tests, executes shell commands, works against git. Install: npm i -g @anthropic-ai/claude-code

Use it for

  • Modifying real codebases & repos
  • Work needing git, tests, shell commands
  • Large-scale file manipulation (batch renames, format conversions)
  • Engineering, not knowledge work

Don't use for

Knowledge work, docs, briefs, or anything Cowork could handle. Most expensive surface — easiest to burn limits.

Relative cost

HEAVIEST · 15–100×

If your task is… → open

  • "Draft a one-pager on our new offer" Chat
  • "Help me think through hiring a COO" Chat
  • "Write a LinkedIn post about last week's launch" Chat
  • "Read every PDF in /Clients/AgriCo and produce a debrief pack" Cowork
  • "Every Monday at 7am, summarise last week's Fathom calls" Cowork
  • "Build the financial model we just designed" Cowork
  • "Refactor the auth module and run the tests" Code
  • "Update version numbers in every SKILL.md in this repo" Code

Relative cost per task

Surface · Modelvs Chat
Chat · Haiku
Chat · Sonnet
Chat · Opus
Cowork · Sonnet10–30×
Cowork · Opus20–60×
Code · Sonnet15–40×
Code · Opus30–100×

Directional — actual cost varies with context size, tool calls, and task length. Source: Anthropic Enterprise Consumption Guide + community benchmarks.

API rates · per 1M tokens

ModelInOut
Haiku 4.5$1$5
Sonnet 4.6$3$15
Opus 4.7$5$25

Output is 5× input. Opus 4.7's tokenizer can produce up to 35% more tokens than 4.6 for the same input.

Team rules

  1. Default to Chat. Ask "could I have done this in Chat?" after every Cowork task.
  2. Plan in Chat, execute in Cowork.
  3. Default model: Sonnet. Opus only when needed.
  4. Don't reupload files. Use Projects (Chat) or folders (Cowork).