Vertical AI Model vs GPT Wrapper

No, not on its own — and the question is pointed at the wrong thing. Whether a vendor trains its own model is a fact about their cost structure, their latency, and where they can deploy, not a predictor of whether the output clears your bar. A “vertical AI model” means the vendor changed the weights with domain data. A “GPT wrapper” means they did not, and instead built prompts, retrieval, tooling, and a product around somebody else’s API. Both descriptions cover shipping products that work and shipping products that do not.

It is not a quality tier, and it is not a moat by itself. Custom weights do not certify accuracy, and an API call does not condemn it. The buying question is narrower: what does this vendor own that a competitor with the same API key cannot rebuild in a quarter — and does the answer survive the next frontier model release?

Four different things vendors mean by “our own model”

The phrase covers a ladder with roughly 1000x between the rungs in cost and defensibility.

1. Prompt and context engineering. System prompts, tool definitions, output schemas, chunking strategy. This is a wrapper, and calling it a model is marketing. It is also where most of the observed quality difference between two products on the same base model comes from.
2. Custom retrieval and embeddings. The vendor trains the search layer, not the generator. Harvey worked with Voyage AI on voyage-law-2-harvey, an embedding model fine-tuned on more than 20 billion tokens of US legal text, reported at a 25% reduction in irrelevant search results against off-the-shelf embeddings.
3. Post-training on proprietary task data. Supervised fine-tuning or reinforcement fine-tuning over a base model. This is what almost every “purpose-built model” in ops software actually is in 2026.
4. Pretraining from scratch. A new base model. Almost nobody in ops software does this, and the one famous attempt is the reason.

The BloombergGPT lesson

Bloomberg pretrained a 50-billion-parameter finance model on its own corpus and published it in March 2023. It beat comparably sized open models — GPT-NeoX, OPT, BLOOM — on finance NLP tasks. Then GPT-4, with no access to any Bloomberg data, beat it: independent evaluations after GPT-4’s release reported 68.79% versus 43% zero-shot on FinQA, 76% versus 43% on ConvFinQA, and an 83% versus 61% F1 on FIN3 named-entity recognition.

The generalizable read is not “domain models lose.” It is that a mid-size model pretrained on domain data wins against its own weight class and loses to the next frontier release, and frontier releases arrive faster than a pretraining run pays back. Any vendor pitching you on the size of their proprietary corpus is describing a position that gets re-contested every few months.

Harvey ran the whole ladder in three years

Harvey is the cleanest case study because it has occupied every position in this argument:

• 2023 — fine-tuned a case-law model with OpenAI after finding public-API fine-tuning and plain RAG insufficient. In blind tests, attorneys preferred the fine-tuned model over GPT-4 97% of the time.
• May 2025 — added Anthropic and Google models, ending exclusive reliance on OpenAI. Harvey’s stated reasons are worth reading as a buyer: task-level routing because no single model leads everywhere, redundancy so a provider outage or capacity constraint reroutes instead of failing, and admin-level model choice per workspace.
• June 18, 2026 — announced its own legal model series: post-trained open-source models, built with Baseten, Fireworks AI, Applied Compute, Trajectory Labs, and Nvidia, reported as performing close to frontier models. Co-founder Gabe Pereyra framed the goal as frontier-level capability across the product at an affordable price and a strong security posture.

Read the last one carefully. The stated driver is price and deployment posture, not a capability gap the general model could not close. That is the honest version of the vertical-model argument, and it is a different argument from the one most vendors make to buyers.

The numbers that decide it

Harvey’s Legal Agent Benchmark results published with Fireworks, on a 100-task slice under a strict all-pass standard where every criterion must be satisfied:

On the softer mean-score metric the four configurations nearly converge — GLM 5.1 at 0.8921, GPT-5.5 at 0.892, Opus 4.7 at 0.911. Three things fall out of that table. Domain post-training bought roughly an 8x cost reduction at near-parity mean score. The hybrid beat the frontier model outright on all-pass at about 39% of its cost. And every configuration completes under 20 of 100 tasks end-to-end, which tells you the category is nowhere near solved and that mean-score marketing hides it.

What a vendor owns that you cannot copy

Rank these above the weights when you evaluate:

• Data rights. Gong’s position rests on tens of billions of captured sales interactions it has contractual rights to, and it claims 3x the accuracy of off-the-shelf systems on its trained trackers. Eightfold rests on 1.6 billion career profiles and 1.6 million inferred skills. Neither is reproducible with an API key.
• Evaluation infrastructure. Harvey built and published BigLaw Bench and the Legal Agent Benchmark with Snorkel AI. A vendor that cannot tell you how it knows a model change helped is shipping vibes.
• Workflow, permissions, and audit surface. The thing your admins live in, which no model release invalidates.

Diagnostic questions for the vendor

1. Which model runs this feature today, and which ran it six months ago? No answer means version drift is uncontrolled and your accuracy baseline is not reproducible.
2. Show a task where your model beats the frontier model you have access to — then let us rerun it on our documents. Vendor benchmarks on vendor data are a starting point, not evidence.
3. Is our data in your training set, and what is the zero-retention option? “Proprietary dataset” sometimes means your inputs.
4. What happens when your provider deprecates a model? Harvey’s multi-model rationale names outages and capacity limits directly; single-provider vendors carry that risk on your behalf.
5. What is the cost per completed task at our volume? The table above is the reason — the same output has an 11x price spread depending on architecture.
6. All-pass or mean score? Insist on the criterion-level number. Mean score is where partial credit goes to hide.

Watch-outs, each with a guard

“Purpose-built model” that is a system prompt. The claim costs nothing to make. Guard: ask in writing which rung of the ladder above applies — prompt, embeddings, post-training, or pretraining — and put the answer in the contract’s product description.

A fine-tune frozen against a moving base. A model post-trained on a 2024 base decays in relative terms with every frontier release. Guard: ask for the retrain cadence and the date of the current checkpoint, and make it a renewal question.

Own-model claims used as lock-in. If the outputs only exist inside their model, migration cost is the product. Guard: require bulk export of outputs plus the source evidence in a format you can re-ingest.

Buying the corpus story instead of the corpus. Vendors quote dataset size without saying whether they hold rights to train on it. Guard: ask for the data-rights basis, not the row count.

Grounding confused with training. Most legal and GTM accuracy problems are retrieval failures, not weight problems. Guard: work through legal AI grounding vs hallucination and RAG before you accept “we trained our own model” as an accuracy answer.

When it actually matters

It matters when you have a deployment constraint the API cannot satisfy: data residency in a jurisdiction the provider does not serve, an air-gapped or sovereign install, a latency floor for a real-time surface, or volume high enough that an 8x unit-cost difference is a budget line. In those cases a vendor that controls its own weights has options a wrapper does not, and you should ask for them explicitly.

It does not matter for a 40-person team choosing a contract-review or conversation-intelligence product. Run the tool on your own last 20 documents or calls, score the output against what your team would have produced, and buy on that. The weights are the vendor’s engineering problem. For the broader pattern of separating an agent from a feature bundle, see what makes an AI agent for ops.