AI Agent Switching Costs in 2026: What OpenAI's Import Changed

Published August 30, 2026By ABD Legacy LLC
AI tools / switching costs / ChatGPT

The switching-cost curve just bent

For most of 2026, the economics of switching AI coding agents were simple and grim: moving a team from Claude Code to Cursor — or from any agent to any other — meant rebuilding everything by hand. Instructions, skills, configuration, project memories, MCP server connections, hooks, slash commands, subagents, and weeks of chat history were each tied to one runtime. The mental model agencies used was "sunk setup": the longer a team had been in one tool, the more expensive leaving became.

That model stopped being accurate on August 11, 2026, when OpenAI's desktop-plus-CLI import wave landed. ChatGPT Work and Codex can now pull an entire working agent setup out of Claude Code, Claude Cowork, or Cursor — configuration, skills, memories, and recent work history — in minutes instead of days [1][2]. The cost curve didn't disappear; it bent. This page is the agency-grade version of that change: what exactly transfers, what still costs money, and why the honest framing is asymmetric, not zero.

What OpenAI's import feature actually transfers

OpenAI's official import documentation (learn.chatgpt.com/docs/import) lists exactly what moves into ChatGPT Work and Codex when you import from another agent [1]:

Source itemDestination in ChatGPT / Codex
Instruction filesAGENTS.md
settings.jsonconfig.toml
SkillsSkills
PluginsPlugins
Existing project foldersProjects (same folders)
Project memories (Claude Code)Memories
Chats from the last 30 daysChatGPT chats
MCP server configurationCodex MCP configuration
HooksCodex hooks
Slash commandsSkills
SubagentsCodex agents

The practical reading for agencies: everything a team configured once — the entire "build & configure" layer — is portable now. AGENTS.md gives the new runtime your project conventions. Skills and slash commands bring your workflow vocabulary. Memories carry project context. MCP configuration brings the external tools you depend on. Hooks preserve lifecycle automation. Subagents preserve the specialist roles your team already built.

Two mechanical details matter when you plan the move [1]:

And the import is non-destructive — OpenAI's docs are explicit: "Importing doesn't change or delete your existing agent setup" [1]. You are not burning the source environment to move; you are copying it, with the option to keep it synced.

How the timeline got here: three releases, one capability

The August 11 wave was the headline moment, but the feature was built in stages across 2026 — which is exactly why it now looks so complete. The official changelog shows [2]:

DateReleaseWhat shipped
June 9, 2026Codex app 26.608First "Import to Codex" flows for Claude Code and Claude Cowork, including during onboarding.
July 21, 2026Codex CLI 0.145.0/import expanded to Cursor and Claude Code — settings, MCP servers, plugins, sessions, commands, and project-scoped memories.
August 11, 2026Desktop + CLI waveDesktop app supports Claude Code, Claude Cowork, and Cursor; CLI /import for Claude Code and Cursor; same release brought the Linux desktop preview.

For an agency evaluating tools, the timeline matters because it tells you the capability is mature enough to plan around: the June and July releases worked out the flow, and the August wave standardized it across desktop and CLI. The official channel for the announcement is the docs changelog rather than a marketing post — treat the changelog as the source of truth when you verify what a specific version supports.

What still costs money: post-import validation is real work

Here is where "switching is now free" gets it wrong. Import transfers configuration; it does not validate it. OpenAI's own docs list exactly what to review after importing [1]:

OpenAI's framing is worth quoting because it sets the right expectation: "A successful import means the configuration was mapped; it does not mean every permission, credential, hook, or integration has been validated for the new runtime" [1]. Translation for agency budgets: the build-and-configure slice of a migration got dramatically cheaper, but a validate-and-harden slice appeared in its place. That is billable effort — someone has to go through the imported permissions, re-authenticate the MCP servers, re-test the hooks, fix the placeholder templates, and re-authorize the plugins before the environment is safe to run client work on.

And one category never transfers at all: credentials and API keys are not on the import list. No API keys, tokens, or provider credentials are documented as importable, so every team that switches re-enters them in the destination [1][3]. That is by design — and it is exactly why "import made switching free" is a slogan, not a cost model.

The asymmetry: import is one-way into OpenAI

The most important structural fact about the feature is the direction it runs. OpenAI documents importing into ChatGPT and Codex from Claude Code, Claude Cowork, or Cursor — but no reverse export route is documented [1][4]. You can bring a Claude Code setup into ChatGPT Work in minutes. Taking a ChatGPT Work or Codex setup back to Claude Code or Cursor still means manual reconstruction.

The consequences for switching-cost math are straightforward:

So the right framing is asymmetric switching costs, not zero switching costs. It is tempting to write "no lock-in" after an import feature ships — resist it. The verified shape is one-way import plus automatic sync plus a post-import review burden; full bidirectional portability is not documented [1][4]. When you evaluate a stack for a client, the direction of the flow is now part of the decision, not a footnote.

This asymmetry also connects to the adjacent story of August 2026: OpenAI's notice that it will stop supplying models to Cursor on November 12, 2026 — the Cursor cutoff is a separate supply decision, but it is one more reason an agency might be weighing the switch right now. We cover that timeline in full in OpenAI Is Cutting Off Cursor: What It Means for AI Coding Agents.

What agencies should do with this: evaluation, standardization, and pricing

Three practical changes follow from the bent cost curve.

1. Stop ranking tools on "sunk setup" as a switching barrier

Evaluation frameworks that still treat "sticky ecosystem / already invested" as a major cost are stale. Migration effort is no longer a days-of-rebuild line item — it is an import plus a review pass [1]. If you compare harnesses for your own stack, the portability score should now reflect what actually transfers and what validation remains, not the sunk cost of the current setup. Our Grok Bot vs Claude Code comparison shows how the harness-level picture looks once setup friction is no longer the deciding factor.

2. Weigh the direction of the flow when standardizing a client stack

The one-way door changes recommendation duty. A client with an established Claude Code or Cursor workflow can move to ChatGPT Work cheaply — but the reverse is not documented, so standardizing a client on OpenAI is a decision with an asymmetric exit cost. Document the direction in the recommendation, and treat "flexibility" as a directional property, not an absolute one. For the vendor-risk side of that call, our AI vendor risk assessment and the running AI vendor risk tracker are the working references.

3. Price migration as import + validation, and budget the validation

If you sell migration services, the effort split changed: configuration transfer is near-zero now, but validation is a real, visible workstream. Quote the post-import checklist — permissions, MCP auth, hooks, prompt templates, plugin re-authorization — as its own line rather than burying it in "setup." And if you are budgeting the agents themselves rather than the switch, remember that agent cost blowups are driven by loops and usage, not setup — see our AI agent cost blowups breakdown for the spend side of the same decision.

Migration effort is now a cost variable

The AI coding agent pricing guide has been updated: import/validation effort now sits alongside usage volume, model selection, and billing model in the cost variables that predict spend.

Read the updated AI coding agent pricing guide (migration effort now a cost variable) →

Frequently asked questions

What are AI agent switching costs?

AI agent switching costs are the time and money required to move a team's agent setup from one tool to another — rebuilding instructions, skills, configuration, memories, and work history in the new runtime. Before OpenAI's import feature, that meant manual reconstruction; today, configuration and recent work transfer in minutes, while validation work remains.

What does ChatGPT Work's Import feature transfer from Claude Code and Cursor?

The official import list covers AGENTS.md instruction files, settings.json (mapped to config.toml), skills, plugins, existing project folders, project memories, chats from the last 30 days, MCP server configuration, hooks, slash commands, and subagents. In the Codex CLI, up to 50 chats from the last 30 days transfer. The import is non-destructive — it doesn't change or delete the existing agent setup.

Do API keys and credentials transfer when importing an agent?

No. Provider credentials and API keys are not on the import list anywhere in OpenAI's documentation — no API keys, tokens, or provider credentials are documented as importable. Teams that import must re-enter credentials in the destination runtime, which is one of the reasons post-import validation is still real work.

What still costs money after importing an agent setup?

Post-import validation. OpenAI's own docs list what to review after importing: permissions and tool restrictions in imported skills and agents, MCP servers using custom auth that may require signing in again, hooks that may behave differently, placeholder-based prompt templates that may break, and plugins or marketplaces that may need manual follow-up or re-authorization. A successful import means the configuration was mapped — not that every permission, credential, hook, or integration has been validated for the new runtime.

Is AI agent switching now free and fully reversible?

No — switching costs are asymmetric, not zero. Import is one-way into ChatGPT and Codex: OpenAI documents importing from Claude Code, Claude Cowork, or Cursor, but no reverse export route is documented. Leaving Claude or Cursor for OpenAI is now comparatively cheap; leaving OpenAI for another agent still means manual reconstruction. The cost of switching depends on the direction of the flow.

How does OpenAI's import feature change agency tool evaluation?

Agencies evaluating tools for their own stack or recommending tools to clients should stop treating migration effort as a days-of-rebuild line item — it's now an import plus a review pass. But the evaluation rubric must also weigh directionality: import lowers the cost of leaving Claude or Cursor for OpenAI while doing nothing for the reverse, so lock-in risk is asymmetric rather than eliminated.

Which agents can you import from, and how does the import work?

On the ChatGPT desktop app, import works from Claude Code, Claude Cowork, or Cursor via Settings > Import. In the Codex CLI, the /import command works from Claude Code or Cursor. The desktop flow supports optional automatic updates that keep imported work in sync with the source agent, and import history is reviewable in the same section. The feature rolled out in stages: first import flows in June 2026, a CLI expansion in July 2026, and the full desktop-plus-CLI wave on August 11, 2026.

Related reading

Sources