Chat Picker

2026 AI Data Export: Comparing Format Support and Migration Ease

Claude documents downloadable account data that cannot be imported into another personal account, while ChatGPT documents data-copy access without stating a file format.

Sources checked 2 Oct 2026

Official pages settle only part of 2026 AI data export: Anthropic’s “Export your Claude data” page documents Claude data exports, while OpenAI’s data-controls page documents data-copy access for eligible personal accounts; the vendor pages we read do not define a shared format or direct import route across ChatGPT, Claude, and Gemini. Chat Picker’s method says it has not tested the assistants for export completeness or migration ease, so use this afternoon plan rather than a ranking.

Know these limits before you start

Use these vendor pages as read on October 1, 2026. Anthropic says individual Claude users on Free, Pro, and Max can export conversation and account data through the web app or Claude Desktop. For Team and Enterprise, only the organization’s Primary Owner can access an export. The emailed download link expires after 24 hours. Claude data cannot be imported into another personal Claude account, although a user joining Team or Enterprise can migrate a personal account directly into the organization.

OpenAI’s data controls say signed-in Free, Go, Plus, and Pro users can request a data copy through settings or the Privacy Portal. That page does not state an archive format, schema, link expiration, or import route. Google’s Gemini Apps Privacy Hub describes processing while signed in, but the page we read does not specify an export format, export location, or cross-account import.

Record the applicable model-improvement setting before using your fixture. OpenAI says personal chats may be used for model improvement when its setting is on. Anthropic says chats and coding sessions are used for improvement when the user allows it. The Gemini pages we read do not state an equivalent app control. A model-improvement setting does not establish what an export contains.

Before paying, separate chat access, export, and API access. OpenAI’s Plus help page lists Plus at $20 a month and says API usage is billed separately. Google’s US plans page lists AI Plus at $4.99 a month and AI Pro at $19.99; its app limits refresh every five hours until the weekly limit. OpenAI’s cancellation page says cancellation stops renewals but does not automatically refund prior charges; canceling at least 24 hours before the next billing date avoids the next period.

For support, OpenAI uses a Help Center chat bubble, while Claude offers asynchronous written support without phone or live chat. Gemini has a community page. OpenAI’s Developer Community excludes ChatGPT app discussion, so treat community answers as leads rather than product specifications.

The test plan

  1. Prepare a fixed fixture. Create migration-fixture.csv with the fields record_id, created_at, title, status, assignee, tags, attachment_name, and source_note. Include accented characters, blank fields, an exact timestamp, a duplicate, a long tag list, an attachment name, and a source note. Keep an untouched reference copy. Give each assistant this exact instruction: “Use only the attached migration-fixture.csv. Return records.txt, records.md, records.json, and records.csv. Keep every field in the original order. Put any missing or altered value in loss_notes.md. Do not infer missing data.” Record the expected value and type for each field.

  2. Request a native export. For each service, use only the export or data-copy control exposed in the test account. Do not count a chat response as an account export. Record the date, plan, eligibility, menu path, processing delay, delivery channel, file and archive names, extensions, byte size, and errors. If no control is visible, write “not found in this account.” If the vendor documentation is silent, write “not documented.”

  3. Check raw export fidelity. Compare the fixture with the raw native export, field by field. Record missing, renamed, flattened, or type-changed fields; altered timestamps; duplicate handling; character encoding; conversation labels; attachment names; and folder structure. Keep any assistant-made conversion in a separate “generated” category so it cannot be mistaken for native export output.

  4. Test import paths. Try each readable export in a clean test account for every target service that exposes an import or migration control. Record accepted formats, required conversion, manual steps, permission changes, attachment handling, duplicates, and whether labels or memory entries carry over. No visible import control means you observed no import path, not that none exists.

  5. Test search and size. Add cobalt-lantern-archive to exactly one fixture record and give every record a distinct title. Search for the phrase in the raw files, after extraction, and in the imported account. Record whether operating-system search, in-app search, both, or neither finds the correct record. Also record total bytes, largest file, extracted size, and any offered compression. Do not infer searchability from a file extension.

  6. Separate UI and API workflows. Repeat the same exact instruction through an API only if your account exposes a documented data-export function. Label model-generated files as “generated,” not native export. When applicable, record separate billing, authentication, pagination, rate or size limits, file schema, and whether the result contains the full account or only the supplied fixture. OpenAI says Plus does not include API usage, while Anthropic lists pay-as-you-go API credits through Console; the Gemini pages we read do not settle an equivalent export path.

How to read your results

Set your required records, fields, and privacy settings before running the test. Then use separate labels: pass, conditional, fail, not documented, and not found. A pass meets every requirement. Conditional means the data is usable only after the manual work you recorded. Fail means you observed missing required data or a blocked import.

Never convert “not documented” or “not found” into a product-wide claim. Compare effort from your log rather than assigning a score: conversion steps, manual corrections, extra paid usage, elapsed time, and support cases. If more than one option passes, choose the one that meets your requirements with the least work. Keep the raw exports, generated files, settings, dates, and logs together so the check can be repeated.

Where the plans differ

The Claude vs ChatGPT vs Gemini matrix compares the rows for free plans, low-cost tiers, main paid plans, heavy-use plans, team plans, context windows, usage limits, training on chats, and ads. For an export test, check team ownership and data controls before price. Published figures carry a vendor source and read date; an unconfirmed figure remains blank under Chat Picker’s method.

Reopen the linked vendor page before subscribing. A tier can change cost or usage without specifying a portable export format. Nothing in the cited Claude export page or Google subscriptions page establishes their cancellation or refund terms, so verify those separately rather than applying OpenAI’s terms to them.

Sources

For current prices and limits, see the dated comparison pages.

Compare assistants