← Back to Blog

LLM Vendor Lock-In Is an Artifact Problem

LLM Vendor Lock-In Is an Artifact Problem

We've argued before that enterprises need LLM-agnostic architecture: a provider interface in front of the model, and context and memory stored in your own schema. Teams that built that way can move an application from one model API to another with a config change and an eval run. On the engineering side, switching LLM providers has gotten cheap.

That post was about the systems your engineers build. This one is about the tools everyone else uses, which is where most of the AI work in a company happens. And that's where the lock-in went.

Where LLM Vendor Lock-In Lives Now

Open ChatGPT or Claude inside any large company and look at what people keep there. Chats, first. Then projects, with uploaded files and saved instructions. Then memory, the notes the app keeps about each user so it doesn't have to be told the same things twice. And then artifacts: the reports, dashboards, docs, small apps, and tables that people build in a chat and then come back to every week.

None of that runs through your gateway. It lives in the vendor's app, in the vendor's storage, organized by the vendor's idea of what a project or a folder is. Nobody decided this. It's just the default, and it piles up a little more every working day.

So when we hear an executive say "we're model-agnostic" we ask a follow-up. Is the company model-agnostic, or only the code? For most companies, we think it's only the code. The API can switch in a week. The people can't, because their work is somewhere else.

The LLM Market Keeps Changing Leaders

If the model race had settled, none of this would cost much. You could pick the winner and stop worrying. It hasn't settled.

Menlo Ventures' December 2025 survey of enterprise LLM API spend put Anthropic at 40%, up from 24% in 2024 and 12% in 2023. OpenAI fell from 50% in 2023 to 27%. Google went from 7% to 21%. In coding, Menlo had Anthropic at 54% against OpenAI's 21%. A company that standardized on the market leader in 2023 was on the number two vendor two years later, without changing anything.

Prices move just as fast. On September 22, 2026, Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, 20% below Opus 5. Within about an hour OpenAI released GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50, each half the per-token price of the model it replaced. Two flagship releases and two price cuts, on the same afternoon.

We don't know who leads next year, and we don't trust anyone who says they do. What we do know is that the gap between models changes faster than an enterprise software contract. When a better or cheaper model ships, the company that can move its teams in a week captures the difference. The company that needs a quarter pays for the old one in the meantime.

What Cloud Lock-In Taught Us

The cloud era ran this experiment already, and it took about a decade.

Cloud providers charged egress fees to move data out, and the fees made leaving expensive on purpose. Regulators noticed. The UK Competition and Markets Authority's cloud services market investigation reached its final decision on July 31, 2025. It cited egress fees and technical barriers to switching, and recommended considering strategic market status designation for Microsoft and AWS.

The providers moved first. Google Cloud dropped egress fees for customers leaving its platform in January 2024, and AWS followed in March 2024, on the condition that the customer leaves fully. Then the law caught up. Under the EU Data Act, from January 12, 2027 cloud providers generally can't charge switching or data egress fees at all. Until then they can charge only what the switch actually costs them.

AI vendors don't charge you to leave. There's no egress fee on a chat. The barriers are the other kind the CMA named: formats and access.

Take OpenAI. In a ChatGPT Business workspace, members can't export their own data; OpenAI's help center tells them to ask the workspace owner. That's a defensible choice for compliance, but it means an employee can't pick up their own work and carry it to another tool.

Claude is more open on documents and less open on everything that runs. Anthropic's artifacts documentation lists exports for Docs (Word, PDF, Markdown, Google Docs), Slides, Designs, and Motion. But artifacts that store records, use connected apps, or call Claude run on claude.ai, and Anthropic offers no self-hosted version. Per Anthropic's sharing docs, viewers outside your org or not signed in find that "the parts that use connected apps or Claude don't work for them." You can take a PDF of a dashboard with you. You can't take the dashboard.

Why AI Artifacts Are the Hardest Part to Move

Chats are easy to walk away from. Almost nobody reopens a chat from last spring, and if the new tool has a better model, people start fresh and don't miss much.

Artifacts don't work that way. The weekly pipeline dashboard has a sales team that opens it every Monday. The pricing calculator is linked from Slack threads all over the company. The onboarding doc is the only copy of how a process works. Each of these has users, inbound links, sometimes stored data, and an owner who'd rather not rebuild it.

Moving one means recreating it somewhere else, re-sharing it, and chasing every link. A Markdown doc moves in a minute. An app that keeps its records in a vendor's storage and calls that vendor's model to work doesn't move at all; you rebuild it. Multiply that by every team, and the switching cost of a company that has used one AI app for two years is mostly made of artifacts. We walked through the options for Claude users in our look at Claude artifacts alternatives.

Here's the test we'd give any executive team. If a clearly better model shipped tomorrow, could your teams move this week? Not the API. The people, with their docs and dashboards and the small tools they built. Pick the artifacts your teams open most often and ask, for each one: what format is it in, where is it stored, who can export it, and does it still work if the model behind it changes? If the answers are "the vendor's," "the vendor's," "an admin, maybe," and "no," you're locked in, whatever the architecture diagram says.

A Multi-LLM Strategy Executives Can Enforce

A multi-LLM strategy doesn't need a platform program. It needs a short policy that people can follow without asking IT. Here's ours.

Keep company work in open formats any model reads. Markdown for documents, HTML for pages, CSV for exports, SQL and Postgres for tables. Every current model reads and writes these fluently, and none of them depends on one vendor's app. Store them in systems the company controls, with company sharing and access rules, so that leaving a vendor doesn't mean leaving the work.

Let teams pick models, as long as they share files. Your engineers may prefer one model for code and your analysts another for spreadsheets. That's fine. The rule is that the output lands somewhere the other team's model can read it. Model choice becomes a personal preference, and the shared work stays shared.

For artifacts, one example of a vendor-neutral store is Anchor, which describes itself as a file system for AI artifacts. Per Anchor, it keeps pages as HTML, documents as Markdown, and tables in PostgreSQL, and ChatGPT, Claude, Gemini, Cursor, and other MCP clients can all read and write the same folders. Its enterprise deployment options include on-prem, self-hosted, and managed VPC. It's also another vendor, and it isn't open source. Hold it to the same test as everything else: check that what comes out is plain files you can open somewhere else.

Keep contracts exit-able at renewal. Avoid multi-year commitments that outlast the model they were signed for. Make sure the contract says how you get your data out, in what format, and who on your side can do it. If only a workspace owner can export, know who that is before you need them.

Trial a second model on a live project every quarter.A benchmark tells you which model scores better. A live project tells you whether your team can actually move: whether the files open, whether the prompts transfer, whether anyone can find anything. If the trial is painful, you've found your lock-in while it's still cheap to fix. For the engineering side of this, the provider interface and outcome-based evals from our earlier post do most of the work.

Switching Costs Are Pricing Power

Vendors know your switching cost, often better than you do. Seat counts, usage, the number of shared artifacts, how many teams have built on the app: all of it shows up in their dashboards, and it shows up in your renewal quote. A customer who needs a quarter to leave gets priced like one.

The cloud history makes the same point from the other side. Egress fees came down under competition and regulatory pressure, after years of customers complaining they couldn't leave. With AI vendors, you don't have to wait for a regulator. You can make the threat to move credible yourself. If your work lives in open formats and you ran a live project on a second model last quarter, you can say so at the table, with numbers.

That's the part we think most executives undervalue. The option to switch pays off even if you never use it. It shows up as better terms, faster access to new models, and teams that pick tools on merit instead of habit. And on a day like September 22, it means you can try the new model the same afternoon!

Disclosure: Neural Bridge and Anchor are related companies.