Why Claude Code costs grow faster than expected
Claude Code, Anthropic’s coding agent, reads files, runs commands and reasons before it answers. Every step costs tokens, and every new message sends the whole session so far along with it. On usage-based plans this adds up quickly: a budget planned for a year can be gone within a few months.
Most of that spend comes from habits rather than from the work itself. Long sessions, the largest model for small fixes and a growing CLAUDE.md all make every message more expensive. Better habits are the fastest way to bring costs down, and they pay off from day one.
The token optimization crash course by DataNorth AI is a three-hour, hands-on training for engineering teams that use Claude Code. Engineers learn the habits and settings that reduce token usage, while keeping output quality and speed.
Quick check: is your team overspending on Claude Code tokens?
Eight statements about how your team uses Claude Code today. Every yes points to a habit the crash course deals with.
- Engineers keep one Claude Code session open for most of the day.
- The largest model is the default for every task, including small edits.
- Our CLAUDE.md files have grown over time and nobody trims them.
- Engineers ask Claude Code to look through the whole repository instead of pointing it to files.
- Tasks start without a plan or acceptance criteria, so it takes several attempts.
- Several MCP servers are connected that people rarely use.
- Long test logs and build output get pasted straight into the session.
- Nobody can say what a typical task costs in tokens.
Calculate how long your AI budget lasts
The example starts with 25 engineers who each spend €1,000 a month, and an annual budget of €180,000. Change the numbers to match your team and switch the habits on or off.
Months of the year covered by your budget
At this rate your budget runs out after about 7.2 months. With the selected habits it covers the full year, with about €19,534 to spare.
How this estimate works
Monthly spend is the number of engineers times the average spend per engineer. Each habit lowers what remains by the percentage shown. The effects are multiplied rather than added, because they overlap. The percentages are planning assumptions, not measured results. Combined with a usage baseline before the crash course, you see the real effect in your own numbers.
An illustration based on your own numbers, not a guaranteed saving.
Plan a token optimization crash course for your team
Find out what one task really costs your team. Bring your usage numbers, and we will show you where the tokens go and which habits to change first.
Use the calculator above to estimate what your team could save.
Plan a crash courseDataNorth is independent of AI vendors and has worked with 250 organizations in 20 countries since 2022.
Where the tokens in a Claude Code message go
A message to Claude Code carries much more than your question. It adds its own instructions, your CLAUDE.md, a description of every connected tool and everything that happened earlier in the session. All of that is sent again with every message, so a long session gets more expensive with each turn.
Claude Code instructions
The instructions Claude Code itself sends to the model with every message. You cannot remove this part. It rarely changes and is cached, which keeps it cheap as long as nobody changes settings halfway through a session.
In the crash course: how cache tokens work, and why switching settings mid-session can make this block expensive again.
Tools, MCP servers and skills
Every connected MCP server, plugin and skill describes itself to the model. Tools the team never uses still take up space, and they give the agent more options to wander off into.
In the crash course: only the MCP servers you need, plus reusable commands and skills instead of retyping instructions.
CLAUDE.md
Project memory in CLAUDE.md loads with every message. It tends to grow after every incident, often with whole documents copied in, and rarely gets trimmed.
In the crash course: a short project memory with links instead of copies, and rules that keep build and vendor folders out.
Conversation history
Everything said and done earlier in the session, including dead ends and abandoned attempts. In a long session this is usually the largest block, and it grows with every message.
In the crash course: /clear between tasks, and /compact with a clear focus when a task runs long.
Files, logs and command output
Complete files, long test logs and build output read along the way. Much of it was only needed once, but it stays in the session and is sent again with every message.
In the crash course: pointing to specific files instead of the whole repository, and keeping long logs out of the context.
Your question
The part you actually typed is usually the smallest block. A more specific task costs a few extra tokens here and saves many more in the rest of the session.
In the crash course: small, well-scoped tasks with clear acceptance criteria, so Claude Code needs fewer attempts.
Illustrative example of a long session. A usage baseline shows the actual split for your team.
What the crash course covers: eight modules in three hours
The crash course combines short explanations, live demos in Claude Code and a hands-on exercise on a real task from your own codebase. Every module ends with something participants can apply the same day.
How Claude Code spends tokens
Input, output and cache tokens, why long sessions get expensive, and how to read what a session costs and what fills its context.
Costly habitTreating Claude Code like a chat window with a flat fee.
New habitChecking what a session costs, and knowing what makes it grow.
In Claude Code
/usage # tokens and cost for this session
/context # what fills the context window right nowContext hygiene
Clearing between tasks, compacting with a clear focus, and pointing to files instead of the whole repository. This is where most of the waste sits.
Costly habitOne session for the whole day, and "have a look through the codebase".
New habitA fresh session per task, and a direct pointer to the files that matter.
In Claude Code
/clear # new task, clean context
/compact keep the failing test and the diff
@src/orders/checkout.ts # point to the file, not the repoA lean CLAUDE.md
A short project memory, links instead of copies, and rules that keep build and vendor folders out of the context.
Costly habitA CLAUDE.md that grows after every incident, with whole documents pasted in.
New habitA short file with the essentials and links to the rest.
Example CLAUDE.md
Build: npm run build Tests: npm test
Code style: see docs/style.md
Architecture: see docs/architecture.md
Do not read dist/ or vendor/The right model for the job
Larger models for planning and hard problems, smaller ones for routine edits and subagents.
Costly habitThe largest model for every task, including a typo fix.
New habitA larger model to plan, a smaller one to carry out routine work.
In Claude Code
/model # switch model to fit the taskPlan first, then build
Plan mode, small well-scoped tasks and clear acceptance criteria mean fewer iterations, and every iteration you avoid is a full round of tokens saved.
Costly habitStarting with "build feature X" and correcting the result five times.
New habitA plan and acceptance criteria first, then small steps.
In Claude Code
Shift+Tab # switch to plan mode before buildingReuse and tooling
Custom commands, skills and subagents, caching stable context, and only the MCP servers you need.
Costly habitRetyping the same instructions, with every available MCP server connected.
New habitReusable commands and skills, and a lean set of tools.
In Claude Code
/mcp # keep only the servers you need
/agents # hand routine work to subagentsHands-on exercise
A real task from your own codebase, done twice. Participants compare token use and output quality between the two runs. The exercise has a basic and an advanced track, so beginners and agentic power users both get value.
Baseline runThe task done the usual way, with the habits the team has today.
Optimized runThe same task with the habits from modules 1 to 6.
Checklist and Q&A
Each participant leaves with a personal token checklist. The team agrees on shared habits, and open questions are answered.
Participants receive
- A token cheat sheet
- A CLAUDE.md template
- The session recording
Who the crash course is for
The crash course is for engineering teams that use Claude Code every day, and for the engineering managers who own the budget. Token use differs strongly per role, even within one team, so the content covers four typical profiles. Pick the one that sounds most like your team.
Assisted coder
Snippets, explanations and small fixes. Short sessions and low token use.
Biggest gain: better habits. Clean sessions, specific questions and a smaller model for small fixes.
Feature builder
Daily feature work, refactoring and debugging across larger parts of the codebase. Context size drives the cost.
Biggest gain: context hygiene, a lean CLAUDE.md and planning before building.
Quality and test engineer
Test generation, bug reproduction and requirements analysis. Long logs and test suites in the context add up fast.
Biggest gain: keeping long logs out of the context, and compacting with a clear focus.
Agentic power user
Multi-step agents, subagents, integrations and automation. The highest spend per user, and also the highest potential value.
Biggest gain: the right model for each subagent, reusable commands and skills, and a lean set of MCP servers. The advanced track of the exercise is built for this group.
We tailor the examples to your team’s stack, and the hands-on exercise has a basic and an advanced track. Combined with a usage baseline beforehand, the effect of the crash course becomes measurable in your own numbers. The crash course changes how engineers use Claude Code. A user tier model and monitoring can then decide who uses which tool, and at what cost.
Reduce Claude Code Costs
Token Optimization Crash Course
3 Hour Crash Course
Live and remote options available
10 to 15 developers per group
Basic and advanced track
