← Back to Blog

This Team Dashboard Found 95% of AI Token Input Was Re-Read Context

By Asma Stewart, Partner and Co-founder at Clarity & Motion Collective ·

"Token maxing" usually means a scoreboard: who burned the most, which team looks the most AI-forward. Asma Stewart wants the term to mean something else, namely how much intelligence you get for what you spend. To find out, she built a usage dashboard for her consultancy in Claude Code, fed it the team's real session logs from Claude Code and Codex, and looked at where the tokens went. Most of the spend turned out to be the model re-reading context it already had. This recap covers how the dashboard works, what the data showed, the session habits that cut re-reading, where model choice matters (and where it barely moves the bill), and how to turn the numbers into a 2027 AI budget.



What is token maxing, and why redefine it?

Asma pointed out that the phrase reads two ways. The common version is about volume. Earlier in the AI adoption curve, plenty of companies treated token burn as proof that people were using AI at all. She doesn't find that version useful anymore. Burn is easy to measure. What's hard to measure is what the spend gets you, and that's the version GTM leaders need.

She gave three reasons to care right now:

  • Budgeting. Planning season is here, more agentic workflows are coming online, and many marketing leaders have been guessing at what their AI budget should be.
  • Visibility into how the team works. Where people lean on AI, and where there might be more upside if they leaned in further.
  • Efficiency. Whether you are buying intelligence or buying waste.

She was upfront about the caveats. The build is preliminary and "cheap and cheerful." The goal was directional data on usage across a whole team, solid enough to decide how to harden it later. It was never meant to be engineering-grade precision.

What does the token dashboard track?

The dashboard reads straight from the consultancy's own session logs. It's real usage data, not a case study. Claude Code covers all four seats and Codex covers two of them. Chat usage (Claude chat or ChatGPT) is not included. The focus is Claude Code and Cowork, because that's where agentic workflows run and where spend needs the most scrutiny.

You can filter by time scale (days, weeks, months), view the whole team or a single user, and switch between data sources. Any enterprise tool that produces logs can plug in. Because the team is on monthly subscriptions, the dashboard shows what the usage would cost at API rates. That's a useful reference point, even though it isn't what they actually pay.

The dashboard also tags each session with a work category, using categories the team defined:

Category What it covers
Client work Where most of the tokens go, which is where Asma wants them
Content Content creation
Internal projects Work on the consultancy itself
Personal Non-business use
Other A catch-all she plans to break down further

The categorization works well for them because their work lives in folders tied to specific clients, a structure they set up to make working with Claude and Codex easier. If your files aren't organized like that, you'll need another way to trust the categories.

The team also built a Notion version of the dashboard, since Notion is their system of record. It's less flexible on UI, but it puts the data in front of people in a tool they already open every day.

What did the data show about where tokens go?

The headline numbers, from about 3,000 sessions across four seats going back to March:

Finding Number
Share of input that was the model re-reading context it already had 95%
Share of usage that was the model writing output back 11%
Tokens Asma's conversations typically start with before she types a word 11,000
Tokens one colleague's conversations start with before typing 58,000
Share of Asma's cost driven by model choice 1%
Share of cost driven by model choice for some teammates 11%

Asma had assumed output would drive most of the cost. In practice, almost none of the bill was the model doing new work. It was the model catching up on what had already been said.

She doesn't call that a bug or waste. In current systems, re-reading the context window is how memory works, and the labs are working on it. Her point is that you should know this is happening and let it shape how you run sessions.

The starting-token numbers were the other surprise. Every connector you leave switched on gets shipped along with each re-read, so a pile of active connectors inflates every turn before you've typed a word.

Is Claude really more expensive than ChatGPT?

Asma's research pushed back on a common claim: that Claude costs more because ChatGPT compresses old messages and Claude keeps everything. Both compact. Claude Code appears to compact more often and earlier, and Anthropic recommends compaction for long-running work. The real difference is when it kicks in.

That's why token burn looks like a sawtooth, not a straight line. Usage drops after a compaction and then climbs again. Compaction isn't free either. It throws away detail on purpose, including things you told the model earlier, so the cheaper path costs you some accuracy. Left on autopilot, auto-compaction can also lead to drift: over a long session the model gradually wanders away from the original prompt.

She also cited published research where cheap models looked cheaper per token at first but ended up costing more once tasks got complicated and needed more tokens. Her read is that cost between Claude and ChatGPT is closer to parity than the rumor suggests.

How do you cut re-reading costs?

Since re-reading drives most of the spend, Asma uses two simple levers: make what the model reads smaller, and make it stop re-reading sooner.

How do you make what the model reads smaller?

  • Start over instead of correcting a bad start. If the first response is way off, don't build on it.
  • Ask for the length you want.
  • Structure your context. If five documents each hold one relevant table or a few relevant paragraphs, pull out those pieces and structure them, instead of attaching all five files and making the model work out what matters. "It's called context engineering for a reason."
  • Send the lightest files. PDFs are heavy and get re-read over and over. Screenshots are hard for models to read and burn tokens. Use .md files.
  • Lock in good outputs. When something comes back right, say it's done and move to the next thing instead of letting the model regenerate it.
  • Turn off unused connectors. Asma admits she's guilty of this one, and it's the likely cause of those 11,000 and 58,000 token starts.

How do you make it stop re-reading sooner?

  • Batch questions and feedback. One message with four questions beats four separate messages.
  • Start a new session when the scope changes.
  • Carry results forward. When research is done, take the output into a fresh session for the next step.
  • Compact deliberately. Don't wait for auto-compaction to decide what to throw away.
  • Use subagents with their own context windows for agentic work.

Her subagent example is a competitive analysis across eight companies. Run it in one session, and by the eighth company the model is re-reading everything it found about the first seven. Split it instead: one main agent holds the questions and QAs what comes back, and eight small subagents each research a single company and hand a summary back to the main agent.

When should you start a new session?

Asma called this one "therapy," because it's the habit people struggle with most. After you've built up a session that understands the problem, starting over feels like throwing away work. She argues it isn't. Starting fresh lets you choose what carries forward, so Claude or Codex doesn't decide for you through compaction. It's the option where you have the most control.

Her rules of thumb for switching:

  • The job changes. Research is finished and now you need writing.
  • You've corrected the same thing twice. That's a sign of drift and of the model losing context.
  • You're about to start the biggest piece of work in the session. Give it a clean window.
  • You've finished a discrete part of the project. Decompose agentic projects up front so you know where the session breaks fall.

In Claude Code, she watches the context indicator in the bottom right of the window. Once it shows the yellow circle, the session is essentially spent and it's time to move on. Cowork doesn't have a similar indicator yet, but the same rules apply there.

Does matching the model to the work actually save money?

"Match the model to the work" is the standard advice, and Asma's data only partly backed it up. Like most teams, hers leans heavily on the most capable models. Their work is strategic, with long-running tasks, but the data still told her they over-rely on Opus.

The catch is that for her own usage, model choice drove only 1% of cost, and re-reading drove almost everything else. For some teammates, model choice accounted for 11%. So model choice matters for some people, but session hygiene is the bigger lever across the board.

The team's current split:

Model How they use it
Opus 5 Mainly development work in Claude Code. Fine for that, though very verbose
Sonnet 5 Moving all content work here. In her words, it can do things the most advanced models couldn't do six months ago, and it's faster
Fable Used sparingly. One weekly client run requires it, plus testing while it was available at a lower cost

The team has also found Opus to be a token hog for content creation, where Sonnet 5 often does the job as well or better. They are adding token optimization to its monthly business review so it becomes a standing habit.

How was the dashboard built?

All of it was built in Claude Code with plain-language requests, and the team checked that the math held up. The steps:

  1. Write a plan. Tell Claude Code what's different about your company and what you want it to consider.
  2. Retrieve the logs. With four people, each partner ran a script to pull their local session logs.
  3. Write export scripts and merge everything into one dataset.
  4. Try to break it. Stress-test the output for numbers that look off. Asma compared her results against recently published research and against an individual-user dashboard that inspired the project, and the numbers were similar.
  5. Set up a refresh. A plain script, with no AI involved, refreshes the data daily. She may move it to weekly.

The project grew out of a dashboard someone else built for personal use. Asma's question was how to do the same for a team. For the hardened version, she plans to pull data through the analytics API, which is available on Teams plans and above, instead of local logs. She put together a step-by-step cheat sheet of the build so community members can make their own.

How does token data feed AI budgeting and ROI?

Clients are asking for this kind of visibility because they feel like they're operating blind. Asma's team is now using it in 2027 planning conversations, especially with clients who already run sophisticated agentic workflows and can start to see what those will really cost.

The plan is to track usage over time and mark inflection points, such as industry-wide price changes, so the trend line can be normalized. A breakthrough in token efficiency could change the picture overnight. Even so, her view is that teams should start doing this math now.

The dashboard also gives you a starting point for ROI reporting. Once work is categorized and you can see how AI is used on a given project or initiative, you can add the KPIs you're trying to hit, plus benchmarks for how long the same work took before AI. That's how you get from "tokens consumed" to the efficiency story a board wants to hear. Setting that first benchmark during an initial assessment makes the before-and-after comparison much easier later.

What does this look like in client work?

Clarity & Motion starts every engagement with a readiness assessment across people, process, tech and data, then builds an AI fluency program from it. Builds target big automation opportunities, team friction points, or areas a company wants to innovate on. Most engagements start with the CMO or an SVP of Marketing. In three cases where marketing reported to the CRO, sales came in after seeing what marketing had built.

Her latest build is a time-aware launch engine for product and marketing teams:

  • When a PRD comes in, the agent analyzes it and drafts a messaging framework and launch brief for the team to review.
  • A tiering system based on the brief determines the assets needed.
  • It builds a pre-launch plan for early testing with clients and prospects, then drafts the full launch assets: press release, white paper, webinar deck.
  • It is time-aware. If a product launched in January and got traction, the engine will later flag that the team hasn't talked about it in a while and suggest ways to keep the drumbeat going.

The client is using it for its first big launch in September.

Key Takeaways for GTM Teams

  • Measure what you get, not just what you burn. Token volume alone tells you nothing about value.
  • Expect re-reading to dominate. In Asma's data, 95% of input was context the model already had. Make that context smaller and make the model stop re-reading sooner.
  • Audit your starting tokens. Unused connectors can load tens of thousands of tokens before you type a word.
  • Restart sessions on purpose. Switch when the job changes, after two corrections, or before your biggest task, and carry forward only what matters.
  • Use subagents for multi-part research so each piece gets its own clean context window.
  • Match the model to the work, but don't expect it to fix everything. Session hygiene moved the numbers more than model choice did.
  • Build the dashboard before budget season. Real usage data beats guessing when you set next year's AI budget.

Want to Attend the Next Session?

GTM Clawllective is a community where go-to-market operators share how they're actually using AI in the field. The live discussion stays off the record, and every recap includes the speaker's demo video so you can see the build for yourself. Come to the next one and join the community.


About the Speaker

Asma Stewart is a Partner and Co-founder at Clarity & Motion Collective, a consultancy that helps go-to-market teams roll out AI. The firm passed its one-year anniversary in August 2026. Its partners have each led AI transformation inside companies, and the consultancy now helps other GTM teams do what they did, starting each engagement with a readiness assessment and building an AI fluency program from it.

Her focus is data-driven AI adoption: knowing how a team uses AI, what it costs, and where the upside is, so leaders aren't operating blind.

Connect with Asma on LinkedIn.

Watch the session replay: Token Max: Get More From Your AI Spend