Choosing an LLM: pick fast, build to switch

Somewhere in your company right now, a slide deck compares Claude, GPT and Gemini, and a committee is preparing to anoint a winner. Here is the answer to the "which LLM for enterprise use" question before the committee finishes: it matters less than you fear. The leading models are close on most business tasks, the ranking reshuffles with every release cycle, and the price of yesterday’s best capability keeps collapsing. The expensive mistake is welding your systems to any single model, not picking the wrong one.

That is not a dodge; it changes what you should spend your attention on. Pick per task, keep switching cheap, and measure with your own data. The rest of this note is the reasoning, and the parts of the decision that actually deserve committee time.

The gap at the top is small and unstable

For the large majority of enterprise tasks (summarizing, drafting, extraction, classification, question answering over documents), the frontier models from Anthropic, OpenAI and Google are interchangeable in practice. Real differences exist, but they are task-specific: one family handles very long documents better, another is stronger on code, another on multilingual work, and the pattern shifts with each release. Public leaderboards will not tell you which one wins on your contracts, in your language, with your edge cases. Only a test on your own tasks does, which is one more reason an evaluation set is the first artifact a serious AI effort produces.

Enterprise procurement habits map badly onto this market. A three-year commitment to a single provider, decided by a bake-off in month one, optimizes for a ranking that will be obsolete before the contract is signed.

Prices fell faster than quality rose

The quieter story is price-performance. Stanford’s 2025 AI Index report measured the cost of running a model at the level of GPT-3.5 falling more than 280-fold between November 2022 and October 2024, and the slide has continued since. The practical consequence: capability that required the frontier a year ago now runs on small, fast, cheap models. Many companies are still paying frontier prices out of habit for tasks a small model handles indistinguishably, which is often the largest single line in an inflated AI bill, a pattern we dissect in what your AI bill looks like at ten times the volume.

Which LLM for which enterprise task

What works in production is routing, not allegiance. A frontier model takes the genuinely hard work: multi-step reasoning, nuanced drafting, ambiguous judgment calls. A small model takes the volume: classification, extraction, routing, reformatting, which in most business systems is the majority of all calls. European or open-weight models (Mistral, Llama and their peers) take the workloads where data residency, contractual constraints or hosting requirements rule out the default providers, a trade-off with enough depth that we treat self-hosted models versus APIs separately.

One system, several models, each chosen for the task; it is usually both cheaper and better than any single-model answer.

Abstraction beats allegiance

The architecture that makes all of this cheap is unglamorous: a thin internal layer through which every model call passes, so that swapping providers is a configuration change rather than a rewrite. Paired with an evaluation set, it turns model choice from a bet into a routine: when a new model ships, you run your evals against it, read the numbers, and switch or do not. Teams with this setup change models in days and negotiate from strength. Teams without it are hostages to their first choice, a position with real financial consequences that investors have learned to probe, as we describe in pricing model provider dependency.

The corollary: be slow to adopt anything that deepens lock-in early, including fine-tuning a provider’s model before prompting and retrieval have been exhausted.

The questions that deserve the committee’s hour

The model name is the least durable line in the architecture document, so spend the hour on the durable questions. Data terms: is your data used for training, and does an enterprise tier turn that off. Hosting: can processing stay in the EU if your contracts require it. Continuity: deprecation policy, and your evidence that you can switch. Unit cost: price per task at ten times today’s volume, not per token today. Ownership: whether the prompts, evals and integration code are yours. Those answers outlive any leaderboard.

We hold no allegiance ourselves: we ship systems on Claude, GPT, Gemini and open-weight models, chosen per task by measurement. Give the committee two weeks of evidence on your own workload instead of another comparison deck; we run that bake-off at the start of a build, and its results, unlike the leaderboard’s, are yours to keep.