What Insights shows
Insights puts cost, speed, errors and agreement on one page, for a period you choose, with each figure set against the period before it of the same length.
Where to find it. Web: the sidebar's More → Insights. Mac: the sidebar's More → Insights. iPhone: Settings → Insights, and the drawer's More.
The filters
- Range: Last 30 Days, Last 7 Days, This Month, or Custom (up to 92 days). Days end at your own midnight.
- Scope: Only Me, Whole Team, or A Member.
- Project: every project with spend in the period, or one. A project includes the folders inside it.
- Mode: In Turns, Side-by-side or Channel discussions, or all of them.
- Export CSV saves one row per day and model: spend, runs and tokens.
Your range and scope are remembered on each device.
The five figures
- Spend: what was billed or priced at list for your calls. A CLI seat's turns on its own subscription are not money we bill, so they are left out, as on the Spending page.
- Runs: each time the room answered, and how many of those moved a seat to a cheaper model (a small question, a budget nearly spent, or low credit).
- Agreed: of the runs that were reviewed by a second seat, the share that ended with everyone agreeing. A Side-by-side or Channel run is never reviewed, so it is not counted here.
- Median Time: half the runs took less than this. First answer is the median time until the first seat finished speaking.
- Errors: of every seat turn tried, the share that failed, including a seat that went quiet until the room moved on. The family with the highest rate is named when it stands out.
The charts and tables
Spend by Model, per Day draws one line per model family; switch it to Runs or Tokens. Each family keeps one colour everywhere: Claude orange, Gemini blue, ChatGPT green, Grok purple. Hover a day for its values. Show as Table turns any chart into rows.
Where the Money Went splits spend by project, by member (team views), and by release.
Models Side by Side has one row per family, named for the model it ran most: runs, spend, spend per run, median time, errors, how often runs it took part in were agreed, and how often it ran on its cheaper model. Click a column to sort.
Quality Over Time shows, week by week, the agreed share of reviewed runs and the share of runs that were a Keep Going.
Needs a Look
Short sentences about what stands out, each with where to act:
- a model whose errors are well above the others', with the run to look at;
- a budget or credit that runs out within two weeks at this pace;
- a day whose spend is several times a usual day;
- what the prompt cache saved, and whether agreement held on cheaper-model runs.
Who sees what
Everyone sees their own. A team's owner, admins and auditors also see the Whole Team and any one member; a member sees only their own. The page only reads, so an auditor sees exactly what an owner sees.
A team view counts members' discussions that are on our servers. A temporary discussion stays in its owner's Only Me view and never appears in a team's. A private discussion kept on a Mac never reaches our servers: its runs appear in Only Me on that Mac alone, and what it cost appears in your own Spending total without a discussion.
Where the figures come from
Spend and tokens come from the same usage records as the Spending page. Runs, agreement, time and errors come from each run's receipt, which is kept once per run from 21 October 2026. For discussions from before then, each one's last run is counted.