Don't Marry Your Model

A few weeks ago I needed a document transcribed.

Not summarized. Not cleaned up. Not “reconstructed.” Just the exact words, in a format I could edit. The kind of job that’s been a five-minute task since OCR existed. I’ll keep the document itself anonymous; it’s enough to say the exact wording mattered. The sort of text where “close enough” is the wrong answer.

I asked ChatGPT. After a couple of clarifying rounds I got back something polished, neat, and subtly different from the original. The response came with this note attached:

Important caveat (don’t skip this). This is not a verbatim legal transcript — it’s a clean, high-fidelity reconstruction. If this ever becomes legal dispute, you should paste the exact original language into each section to remove any ambiguity.

I read it twice. The model knew exactly what I had asked for. It knew “verbatim” was the operative word. And it had silently decided to do something else, called it a “reconstruction,” then warned me that the thing it had just done wasn’t safe for the situation I’d just told it I was in.

That was the moment. I pasted the same photos into Claude. Got back the exact text, character for character. Ninety seconds.

I switched.

This isn’t a “Claude is better” essay

A few months ago I would have written the opposite.

I came to Claude after a long stretch on ChatGPT and it felt like an upgrade. The model asked clarifying questions instead of dumping pages of confident-but-wrong text at me. The hallucinations were fewer and the answers were tighter. Claude Code became part of how I work.

But Claude isn’t pristine either. I hit usage limits routinely. Codex got noticeably better in the same window. The latest GPT release is pretty good. When Anthropic stopped letting OpenClaw authenticate via OAuth, I moved my coding setup to Codex inside a week.

Two months from now I’ll probably switch again. I’m fine with that.

I’m even fine with hallucinations. I understand the technology well enough to know that the same machinery that lets these models be creative is what makes them confabulate. That’s the deal.

What I’m not fine with is a model that quietly decides what I really meant.

The “high-fidelity reconstruction” wasn’t a hallucination. It was a product decision. Someone on a model team decided that when a user asks for verbatim text from a document that looks legally sensitive, the safe move is to soften it and add a disclaimer. Reasonable in the abstract. Wrong for me. Tomorrow that same choice will be right for someone else, and wrong for me again.

The takeaway isn’t about ChatGPT specifically. It’s that every model carries the personality and policy decisions of the team that shipped it. Those decisions change. Your tolerance for them changes. The frontier shifts every quarter.

If your work depends on a single model, you’ve made a bet on a moving target.

The research caught up to my annoyance

Menlo Ventures’ 2025 State of Generative AI in the Enterprise, based on a survey of about 500 U.S. enterprise AI decision-makers, captured the same wobble at scale. Anthropic now earns roughly 40% of enterprise LLM spend, up from 24% a year earlier and 12% in 2023. Over the same period OpenAI lost nearly half of its enterprise share, falling to 27% from 50% in 2023. Google quietly tripled its share. Three labs trade the lead every few quarters.

The Menlo analyst on the report flagged the catch: once an enterprise chooses a vendor, most just keep upgrading to the latest model from that same provider, even when switching costs are low. Most teams build to one model and ride it, even as a better option shows up next door. That inertia is fine right up until the moment your provider ships a behavior change you can’t tolerate.

The architectural answer to this is older than the AI hype: keep an abstraction layer between your work and any one model. NVIDIA researchers made the technical version of the case last summer in a paper called Small Language Models are the Future of Agentic AI. Their argument: smaller, specialized models, under about 10 billion parameters, are sufficient for most of what AI agents actually do, and “heterogeneous agentic systems” that invoke multiple different models are the natural choice when general-purpose conversation is also needed. A planner decides what to do; a small specialist actually does it; a different one critiques the output. The model that summarizes my calls doesn’t have to be the model that drafts my prose, which doesn’t have to be the model that reads my PDFs.

It’s already happening at enterprise scale. ServiceNow signed multi-year deals with both OpenAI and Anthropic in January 2026, within weeks of each other, explicitly to give customers a choice of model based on the task at hand. Snowflake committed $200 million to OpenAI while keeping live partnerships with Anthropic, Google, Meta, and Mistral. Gateways like OpenRouter now front more than 500 models from 60+ providers behind a single API. And a 2026 multi-cloud AI analysis found that enterprises with abstraction layers in their first AI deployment switched primary providers with 60 to 80% less migration effort than teams that built directly against a single vendor API.

The teams that win the next few years aren’t the ones with the best opinion about which model is best today. They’re the ones who’ve built a setup that makes the question boring.

What I actually do

You don’t need a vendor strategy document and a Fortune-500 procurement team. You probably need three smaller habits.

Notice your switching cost. Add it up: rewriting prompts, re-uploading project knowledge, retraining yourself on a different UI, losing months of conversation history. Write that number down. That’s what you’re paying every month to stay where you are. For a lot of us, it’s higher than we think, and lower than it was a year ago.

Separate the work from the model. The prompt that runs your weekly content workflow shouldn’t live inside one vendor’s app. It should live in a file you own, that you can paste anywhere. Same for style guides, examples, system prompts. Treat the model as a renter, not the landlord.

Pick a portable runtime for anything you build. Open-source orchestration projects like OpenClaw can now run the same agent setup against multiple providers. Open protocols like MCP let your tools work with any model that speaks them. Two years from now the model you start with probably won’t be the model you end with. Build like that’s true, because it is.

I’ll probably switch again before the year is out. Maybe to a local model on my own server, with foundation models called only when I actually need real reasoning. Maybe back to ChatGPT, once the “let me protect you from your own request” reflex gets dialed back. Maybe to something none of us are talking about yet.

The choice doesn’t worry me. What I work hard not to do is fall in love with the choice.