How to analyze LLM eval results in Claude Sheets and Claude Data

Short answer: export a run as CSV from your eval tool, import it into Claude Sheets (in beta on paid Claude plans), and ask Claude to add pass-rate formulas, flag cases that flipped from pass to fail, and chart the result. Claude Data can't take a CSV upload: use it only when your eval scores already sit in a warehouse it connects to, such as BigQuery or Snowflake. Facts checked 11 Nov 2026.

By Evals Review editorial team · Published 11 Nov 2026 · Updated 11 Nov 2026 · Next check 25 Nov 2026

How we checked: product facts come only from Anthropic's launch posts, Help Center articles and the @claudeai launch thread; export steps come from each eval tool's own docs, all read on 11 Nov 2026. This is a docs review: we have not run our own eval results through Claude Sheets or Claude Data. How we review.

TL;DR

  • Claude Sheets is the path for an exported eval CSV: import, formulas, conditional formatting, charts, sharing. [Claude Help Center]
  • Claude Data (formerly Claude Dashboards) needs a warehouse or connected app; each chart shows the SQL behind it. [Claude Help Center]
  • Claude Whiteboards helps the team sort failure themes after the numbers are in. [Claude Help Center]
  • Keep your eval tool for running, storing and tracing evals; this is for analysis and sharing.

What Anthropic launched, and who can use it

On 10 Nov 2026 Anthropic started rolling out Claude Whiteboards and Claude Sheets in beta, and made Claude Dashboards generally available under a new name, Claude Data. All three are on paid plans; the rollout is gradual, so you may not see them yet. [Claude blog] [@claudeai launch thread on X (10 Nov 2026)]

Status and plans (Anthropic pages read 11 Nov 2026)
Claude SheetsClaude DataClaude Whiteboards
StatusBeta [Claude Help Center]Generally available; certified queries and live sharing in beta [Claude Help Center]Beta [Claude Help Center]
PlansPro, Max, Ultra, Team, EnterprisePro, Max, Ultra, Team, EnterprisePro, Max, Ultra, Team, Enterprise
Team plansOn by defaultOn by default, with both beta featuresOn by default
Enterprise plansOff until an owner turns it onTurns on by default on 19 Nov 2026 unless an owner already saved a setting; beta features offOff until an owner turns it on
NeedsYour data (CSV, Excel, Google Sheets import or paste)A data connector: warehouse or connected appNothing; works from what's on the board
UsageCounts toward your plan's usage limits, like other work in Claude [Claude Help Center] [Claude Help Center] [Claude Help Center]

On devices, the launch post says all three work on web, desktop and mobile, while the Sheets help article lists only basic read-only access on mobile. Plan to do the analysis on web or desktop. [Claude blog] [Claude Help Center]

Workflow: from eval run to shared chart in Claude Sheets

  1. Export the run as CSV. Export two runs if you are comparing a baseline with a candidate prompt or model. Remove columns holding user data you shouldn't share: the Sheets beta doesn't support HIPAA, CMEK or zero data retention. [Claude Help Center] Verified export paths:
    How each eval tool exports results (vendor docs read 11 Nov 2026)
    ToolExport path
    Promptfoopromptfoo eval --output results.csv. The CSV includes test variables, the prompt, model outputs, pass/fail status, latency and token usage. [Promptfoo docs]
    BraintrustOpen the menu next to the experiment name and export as CSV or JSON; you can choose whether to download all fields. [Braintrust docs]
    LangfuseMost tables support batch export as CSV or JSON. Filters on the table apply to the export, and all columns are always exported. The file appears under Project Settings, Exports, with a Download button. [Langfuse docs]
    DeepEvalNo CSV export documented. Set results_folder in DisplayConfig and each run is saved as a test_run_*.json file with per-test-case scores and metric reasons (or to SQLite). Flatten the JSON to CSV yourself before importing. [DeepEval docs]
  2. Check it fits. A sheet holds up to 100,000 cells. [Claude Help Center] Multiply rows by columns. Long text columns (full prompts, full outputs) cost the same one cell each, but make the sheet slow to read; keep a case id and a short output column, and keep a trace id or URL so you can find the full run in your eval tool.
  3. Import into Claude Sheets. In Claude, open the Artifacts tab, pick the Sheets tile and import the CSV, or ask for a sheet in a chat and give Claude the file. Claude Sheets can also import from Excel or Google Sheets, or take a paste. [Claude Help Center] Put each run on its own tab.
  4. Ask Claude for the analysis, in formulas. Formulas stay in the cells, so you can check how each number was calculated, and each edit Claude makes is outlined as it lands. [Claude Help Center] Useful asks:
    • "Add a summary tab with pass rate per run and per category, as formulas."
    • "Match the two tabs on case id and add a column that marks cases that passed in the baseline and fail in the candidate."
    • "Highlight those regressions in red and list them at the top of the summary tab."
    • "Add mean latency and token use per run next to the pass rate."
    • Select a range and ask, for example, "why is this category's pass rate lower than last run?"
  5. Check every formula Claude wrote. Open the summary cells and confirm the range and the condition. A pass rate computed over the wrong rows looks just as believable as a right one. If an edit is wrong, version history can restore an earlier version. [Claude Help Center]
  6. Chart it. Bar, line and pie charts are available, plus other chart types Claude can generate. [Claude Help Center] A bar chart of pass rate per run and category, and a short table of regressed cases, usually tell the story.
  7. Share or hand off. Sheets start private; click Share to work on one together. Export as .xlsx or CSV, or use "Send to" for partner apps. PDF export isn't available yet. [Claude Help Center]

This is step 5 of eval-driven development (compare runs example by example, not only on the average), done in a spreadsheet so people outside the eval tool can follow it.

What the regression view looks like

Example data, made up for illustration. These are not real results from any model or tool.

Example data: two runs matched on case id
case_idcategorybaselinecandidateflip
c-014refund policypasspass
c-027refund policyfailpassfixed
c-031how-topassfailregressed
c-042how-topasspass

In this example both runs pass 3 of 4 cases, so the average says nothing changed, yet one case broke and one was fixed. The flip column is what you review before keeping a change.

When Claude Data fits instead

Claude Data turns questions about connected data into dashboards, scrolling data stories, short narrated videos and written summaries. It works against a data warehouse (Amazon Redshift, BigQuery, ClickHouse, Databricks or Snowflake are named) or another app connected to Claude, writes a SQL query for each chart, and shows that query and when the data was last refreshed. [Claude Help Center]

  • It fits when your eval scores already land in a warehouse, for example through a scheduled export your team loads there. Langfuse, for one, documents scheduled exports to blob storage; [Langfuse docs] getting from there into a warehouse is your own pipeline.
  • Define pass rate once. With certified queries (beta), your data team approves the query everyone should use, and charts built on it carry a certified badge. Editing a certified query removes its certification. [Claude Help Center]
  • Check the SQL. Open each chart's query and confirm it counts what you meant, for example that it filters to one dataset version. [Claude Help Center]
  • Be careful with live sharing (beta). Viewers see results from your access, even without access to the source, so check what the dashboard shows before you share it. [Claude Help Center]
  • It doesn't fit a one-off CSV, and Anthropic itself positions dashboards for quick, exploratory questions, sending deeper analysis to another tool. [Claude Help Center]

Claude Whiteboards for the failure review

After the numbers, the team still has to agree on why cases fail. On a whiteboard, everyone adds notes, Claude can regroup them by theme, you can drag across one area to ask about just that part, and the board can become a doc or deck, or be sent to Lucid, Miro or Zoho Vani. Boards export as PNG. Claude works only from what is on the board, not from your connected apps. [Claude Help Center] Each theme you agree on is a candidate for a new pass/fail criterion; see how to write AI evals.

Limits compared with a dedicated eval tool

Claude Sheets beta vs a dedicated eval tool (Anthropic limits read 11 Nov 2026)
Claude Sheets (beta)Dedicated eval tool
SizeUp to 100,000 cells per sheetBuilt for full run history and large datasets
Live dataNo: a sheet can't pull live data from connected apps; you re-import each runRuns land automatically from the SDK or CI
Trace drill-downNo link back to the trace unless you keep a URL or id columnOne click from a failing case to its trace
Rerun or rescoreNo: Sheets analyses results, it doesn't call your model or judgesRerun a dataset, add a scorer, rescore old runs
Spreadsheet featuresNo pivot tables, multi-column sort, filter views, find and replace or PDF export yet (beta)Not a spreadsheet; built-in comparison views instead
Data controlsHIPAA, CMEK and zero data retention not supported on the Sheets betaVaries by vendor; several offer self-hosting
Sharing with non-engineersStrong: share a sheet, edit together, export .xlsx or CSVUsually needs a seat per reviewer

Sources for the Sheets column: Anthropic's current limitations list. [Claude Help Center] The right-hand column describes the category in general; check a specific tool on its review page, or compare them in LLM evaluation tools compared. If the people reading results are product managers, start with LLM evaluation tools for product managers: several of those tools already have dashboards, and the question is whether every reviewer needs a seat.

What can go stale: Sheets and Whiteboards are betas, so the cell cap, chart types and missing features (pivot tables, sorting, PDF export) may change; the Enterprise default for Claude Data changes on 19 Nov 2026; and eval-tool export menus move. We will re-check these on 25 Nov 2026.

FAQ

Can Claude Data read a CSV of eval results?

Not as a file upload. Anthropic's help article says to connect a data warehouse (such as Amazon Redshift, BigQuery, ClickHouse, Databricks or Snowflake) or another connected app first. For a CSV export, use Claude Sheets, which can import CSV files.

Which Claude plans include Claude Sheets and Claude Data?

Anthropic lists Pro, Max, Ultra, Team and Enterprise. Claude Sheets and Claude Whiteboards are in beta; Claude Data is generally available, with certified queries and live sharing in beta. Anthropic describes them as paid-plan features, and they count toward your plan's usage limits.

How many eval rows fit in Claude Sheets?

A sheet holds up to 100,000 cells, so multiply rows by columns. 2,000 test cases with 12 columns is 24,000 cells; 10,000 cases with 12 columns is 120,000 and won't fit. Drop long text columns or split by run.

Is Claude Sheets the same as Claude for Sheets?

No. Claude Sheets is a spreadsheet inside Claude, launched in beta on 10 Nov 2026. Anthropic's own help article points users who work in Google Sheets or Excel to Claude for Google Workspace or Claude for Excel instead. The older Claude for Sheets add-on for Google Sheets is a different product.

Should I replace my eval tool with Claude Sheets?

No. Keep running and storing evals in your eval tool, which can rerun datasets, link failures to traces and gate CI. Use Claude Sheets to analyse an exported run and share the result with people who don't use the eval tool.

Where do Claude Whiteboards fit?

In the review meeting after the numbers. Paste the failing cases or your notes, ask Claude to group them by theme and draft action items, then turn the board into a doc or deck. Whiteboards don't use context from your connected apps.

Sources