LLM Vendor Lock-In: Enterprise Executive Guide

LLM vendor lock-in happens when your company's work ends up stored inside one AI provider's product. The model is easy to replace. You change an API key, or you pick a different model from a menu. What's hard to replace is everything your people built around it: months of conversations, shared projects, saved instructions, and the reports and apps the AI made for them. If all of that lives with one vendor, switching models means leaving your work behind, and the vendor knows it.

We build Anchor, so we think about this every day, and our advice is short. Keep your options open, keep your company's work out of the model vendor's app, and make it cheap to change your mind.

How fast is the LLM race moving in 2026?

Faster than anyone planned for. At the end of 2023, OpenAI took half of all enterprise spending on LLM APIs and Anthropic took 12%, according to Menlo Ventures. By Menlo's December 2025 survey, Anthropic had 40%, OpenAI 27% and Google 21%. Claude Code drove a lot of that, with Anthropic holding 54% of the enterprise coding market.

Then 2026 happened. Anthropic's annualized revenue went from $9 billion at the end of 2025 to $65 billion at the end of July, TechCrunch reported, against OpenAI's $40 billion. (The two companies may not count revenue the same way, but the direction is hard to miss.) OpenAI fought back. It shipped GPT-5.5 in April, and on September 22 it released GPT-6 Sol and GPT-6 Luna at half the price of the models before them. Anthropic released Claude Opus 5.5 the same day with a 20% price cut. Which one leads now depends on the benchmark you read.

Google had a harder year. It announced Gemini 3.5 Pro at I/O in May and promised it for June. In July, Bloomberg reported that the launch had slipped because the model's coding fell short of Google's own goals, and Alphabet lost close to $200 billion in market value in a day.

Open models closed most of the gap. Moonshot released Kimi K3 in July and published all 2.8 trillion parameters of its weights eleven days later. Nathan Lambert wrote that it ranked third on Artificial Analysis's Intelligence Index, behind one model each from Anthropic and OpenAI, and that open models now trail the closed ones by three to five months. Zhipu's GLM-5.2 and DeepSeek's models ship under the MIT license, so you can run them on your own hardware.

The tools people use to drive these models changed too. Hermes Agent, from Nous Research, went from a new repo in July 2025 to more than 240,000 GitHub stars. Mario Zechner's Pi coding agent has passed 100,000, and it works with more than fifteen model providers. Both let you change the model underneath with one command. A developer on either one can go from Claude to GPT to Kimi before lunch!

Will the LLM race settle down?

We don't think so, and we wouldn't plan around it. Every one of these labs has the money and the people to take the lead back, and each of the big three has held it at some point in the last two years. The model your team loves this quarter has a good chance of being second or third by next year. The best option next year might come from a lab most buyers haven't heard of yet.

So plan for the switch. When a better model shows up, how long would it take your company to move to it? A day? A quarter? Never, because the work is stuck?

Where does LLM vendor lock-in come from?

It comes from the apps the providers build on top of their models.

The model APIs look a lot alike. Prompts move between them with some tuning, and tools like Hermes and Pi exist because the swap is cheap. The providers know this, so they compete on the product around the model. ChatGPT has Projects, memory, custom GPTs and Canvas. Claude has projects and artifacts. Gemini ties into Google Workspace. Each of these features is convenient, and each one stores your company's work inside that vendor's app.

Here's how it bites. Your team spends a year building projects, filling up memory, and saving the reports and dashboards the AI made. Then a better model ships somewhere else, and none of that work comes with you.

Look at what OpenAI's help center says about exports. A personal account can download its chat history from settings. On ChatGPT Business and Enterprise, self-service export isn't available, and users are told to ask their workspace owner. The accounts that hold the most company work are the hardest to get work out of.

We'd put it plainly: don't let the company that sells you the model also hold your files. Keep the two apart and you can replace either one.

What did the cloud era teach us about vendor lock-in?

We've seen this before. AWS, Azure and Google Cloud spent the 2010s fighting for enterprise workloads, and each built services that only ran on its own cloud: its own databases, queues and serverless runtimes. Many of them were good services. They also made leaving expensive, and so did egress fees, the charge for moving your own data out.

In July 2025, Britain's competition regulator, the CMA, published the final decision of its cloud investigation. It found that AWS and Microsoft each held 30 to 40% of the UK market, and that egress fees and technical barriers made it hard for customers to switch or to use more than one cloud. It also found that Microsoft charged more for its software on AWS and Google Cloud than on Azure.

It took regulation to move the fees. Google dropped egress charges for departing customers in January 2024, as the EU Data Act came into force. AWS followed in March and Microsoft soon after. Even then, you had to leave completely, and within 60 days. The Data Act bans switching charges outright from January 12, 2027.

The companies that stayed portable, on open databases and containers, could move a workload when a better offer came along, and they negotiated from that position. The ones that went deep into one vendor's own services took the renewal price they were given, and plenty of them still pay it. Flexera's State of the Cloud report puts the share of enterprises on more than one cloud at 89%. Multi-cloud became the normal way to buy infrastructure. We think multi-LLM will become the normal way to buy AI, and the companies that start now will have an easier time of it.

What does a good multi-LLM strategy look like?

Start with one test: if a better model came out tomorrow, could your company move to it this week? If the answer is no, find what's holding you and move it somewhere you control.

Keep conversations, projects, artifacts and company knowledge in your own systems, in formats any model can read. Markdown, HTML, CSV and SQL have been around for decades, and every model reads them. A vendor's project format works in that vendor's app.

Let people pick their model. Your engineers might prefer Claude, your analysts ChatGPT, and your data team an open model running in your own cloud. That's fine, as long as they all read and write the same files.

Treat the model vendor as a supplier you can replace. Sign contracts you can walk away from at renewal, and try a second model on a live project every quarter, so that switching becomes a habit.

How does a multi-LLM strategy give you bargaining power?

A vendor prices you by how hard it would be for you to leave. If your team can move to another model in a week, every renewal is a negotiation.

The price cuts this month show how much room there is. On September 22, OpenAI halved its prices with GPT-6 and Anthropic cut Opus pricing by 20%. A company that can switch takes those cuts the day they land, and it can send the bulk work to a cheap model like GPT-6 Luna, at 10 cents per million input tokens. A company locked into one vendor's app waits for that vendor's next price sheet.

What does LLM lock-in mean for AI artifacts?

Artifacts are where lock-in grows fastest. They're the reports, dashboards, documents, apps and tables that agents make for you, and today most of them get saved inside the chat app that made them. A dashboard made in Claude lives in Claude. A document drafted in ChatGPT's Canvas lives in ChatGPT. When the company changes models, or when one team uses a different one, those files stay behind.

We built Anchor to keep artifacts apart from the model. Anchor is a file system for AI artifacts. HTML pages, Markdown documents, apps and live tables sit in folders, and every file has a link. Any agent that speaks MCP can read and write them, whether that's Claude, ChatGPT or an open model you host yourself.

So one team can build a report with Claude, another can update it with ChatGPT next week, and the report doesn't care. If you drop a provider next year, your artifacts stay where they are. Pages are HTML, documents are Markdown, and tables are PostgreSQL. They're yours, in formats that were around long before any of these models, whichever model you use to work on them.