Using Claude Cowork Efficiently in 2026

Most people using Claude Cowork are quietly burning tokens they don't need to. Not because they're doing anything wrong, but because the defaults aren't optimised for efficiency, and nobody explains the mechanics until you're staring at your usage dashboard wondering where it all went.

We put together a guide on this for clients and figured it was worth sharing more widely. Here's the practical version.

What tokens actually are (and why they matter)

Tokens are the currency of every Claude session, roughly three-quarters of a word each. Every prompt you send, every file you attach, every MCP connector you have active, and every system instruction in your project gets counted before you've typed a single word. The cost varies significantly by model:

• Haiku 4.5 — ~500 tokens per message. Best for quick Q&A, drafts, and rewrites.
• Sonnet 4.6 — ~2,000 tokens per message. The default for real work.
• Opus 4.6 — ~8,000 tokens per message. Around five times the Sonnet cost. Reserve it for genuinely hard reasoning.

That's not a trivial difference once you're doing real volume. The good news is that most of the waste comes from a handful of predictable sources.

The five biggest token drains

• Idle MCP connectors. Gmail, Drive, Notion, Slack all load their full tool definitions into context whether you're using them or not. That can add 40,000+ tokens before you've started.
• Extended Thinking left on. Cowork reasons silently before responding. Worth it for complex problems, wasteful for routine ones.
• Long threads. Every new message re-sends the full conversation history. A 30-message thread is dramatically more expensive than 30 fresh chats.
• Bloated project instructions. Anything over ~500 words is paying tax on every single prompt.
• Vague prompts. "Tell me about this document" forces full processing. "Summarise the risks in section 3" does the same job for a fraction of the cost.

Seven habits that compound over time

None of these are complicated, but they add up quickly when they become habitual.

• Right model, right job. Default to Sonnet 4.6 for real work, Haiku 4.5 for quick questions and rewrites, Opus only when you genuinely need the extra reasoning power.
• One task, one chat. The moment you switch topics mid-thread, you're paying to carry irrelevant history.
• Batch your asks. "Fix the typo, shorten the intro, and add a CTA" is one overhead for three jobs. Three separate prompts is three overheads.
• Be specific from message one. Name the section, the file, the output format, the audience. Specificity is free and saves a follow-up every time.
• Run /context regularly. Type /context in any Cowork chat to see exactly what's loaded: MCPs, skills, memory files, instructions, and how many tokens each one is burning.
• Disable idle MCPs before you start. Each disconnected connector you're not using is thousands of tokens returned to your budget.
• Toggle Extended Thinking off for routine work. On only when you actually need step-by-step reasoning.

The advanced moves

Once the basics are in place, there are a few higher-leverage techniques worth knowing about.

• Build skills for anything you do more than twice. Skills are packaged instruction sets Claude loads on demand: workflows, brand guidelines, document templates, repeatable processes. Cheaper than re-explaining context each session, more reliable than memory, and they keep your project instructions lean.
• Use Plan Mode for big jobs. In Claude Code, `/model opusplan` uses Opus to draft a plan, then auto-switches to Sonnet to execute. Strategic thinking on the smart model, execution on the efficient one.
• Treat Projects as a library. Files in a Project are cached across all chats with no re-upload cost. Keep project instructions under 500 words — they load on every prompt.
• Use YAML instead of JSON for structured data. YAML uses meaningfully fewer tokens for the same information. Knowledge graphs cut token use even further on large reference sets.

Before you start your next session

Run through this quick checklist:

• Default to Sonnet 4.6 (switch to Opus only when needed)
• Disable Extended Thinking unless reasoning is required
• Audit active MCPs and disconnect anything you won't use
• Trim project instructions to under 500 words
• One chat per task — don't mix unrelated topics in one thread

None of this is about obsessing over token counts. It's about making sure the budget goes on the work that matters, not on overhead that's invisible until it isn't.

If you're thinking about how to get more out of AI tooling in your business, we're happy to have a chat. Drop us a message at futureformed.io.