Should I train my own model?

The question arrives in one of two moods. Either a board member read that a bank trained its own model and wants to know why we have not, or an engineering lead has done the math on a frontier API bill and wants to know whether a smaller model they control would do the job. Both deserve a real answer, and the real answer depends on the task, the industry, and the year. This is the answer for September 2026.

Where the market stands this month

There are now three tiers of model an enterprise can use, and the distances between them have changed in the last twelve months.

The closed frontier. The most capable models from the large American labs, available only as hosted APIs: OpenAI's GPT-6 Astra and the GPT-5.6 family, Anthropic's Claude Fable 5.1 and Opus 5, Google's Gemini 3, and xAI's Grok 4.6. They lead on the hardest work: long-horizon agentic tasks, novel reasoning, complex code, and anything where the answer is not in the training data. They are also the most expensive per token by a wide margin. Published price lists this month run from about two to five dollars per million input tokens for the standard frontier tier, with output several times that, to well over a hundred dollars per million output tokens for the highest-effort variants such as GPT-5.5-pro.

Open weights. Models whose weights can be downloaded, run on hardware you control, and modified. Twelve months ago the best of these trailed the frontier by a year. This month the top of the open-weight rankings is Chinese: DeepSeek V4 Pro and V4 Flash, Alibaba's Qwen3.8, Zhipu's GLM-5.3, Moonshot's Kimi K3, and MiniMax, trailing the closed frontier by months on most public benchmarks, with DeepSeek V4 Pro's result on SWE-bench Verified the strongest published open-weight score on that benchmark. The American and European open models sit behind them on the live rankings: Meta's Llama, Mistral's models including Mistral Small 4, NVIDIA's Nemotron family, OpenAI's gpt-oss, and Google's Gemma. Two facts about this tier matter for an enterprise. Origin now decides eligibility for a growing set of buyers, which the next section covers. And the cost of running near-frontier quality has dropped to where DeepSeek V4 Flash or Mistral Small 4 serve on a two-GPU deployment, and hosted open-weight APIs price output at a few dollars or less per million tokens.

Small and post-trained models. Models from a few hundred million to about ten billion parameters, usually taken from an open-weight base such as Llama, Qwen, gpt-oss, or Nemotron and post-trained on a specific task with a specific company's examples. On the task they were trained for they routinely match or beat frontier models. On anything else they are worse than a cheap general model. Their cost per task is a rounding error next to the frontier, and they run on hardware a company already has.

What each tier is for

The useful question is never "which model is best." It is "which model clears the bar for this class of task at the lowest cost per accepted outcome," where accepted means a person or a downstream check signed off. Measured that way, the tiers sort themselves.

Use the frontier when the task is new every time. Investigations, plans, unusual analyses, code that has to work in an unfamiliar codebase, judgment calls with money attached. Here the extra capability is the product, and paying frontier prices for it is correct. Paying frontier prices for routine drafting is not.

Use open weights when you need control of the runtime. Residency requirements that no hosted region satisfies. Workloads that must not leave a boundary. Latency budgets that hosted APIs cannot meet. Or simply volume: an open-weight model that is 90 percent of the frontier at 5 percent of the price wins every task where 90 percent clears the bar.

Use a post-trained small model when the task is narrow, repeated, and specified. Classification, extraction, tagging, routing, structured drafting from a template, tool-call generation against a fixed set of tools. These are the tasks where a company has thousands of labeled examples lying around in its own history, and where a small model taught on those examples outperforms a frontier model that has never seen the company's conventions.

The constraints the task does not tell you

Two considerations override the task-level analysis, and both are sharper this year than last.

Geography and industry. The open-weight tier's best models are Chinese, and for a growing set of enterprises that is disqualifying regardless of quality. The 2026 defense authorization bars models domiciled in China, Russia, North Korea, or Iran from the Pentagon and its contractors. Several states have banned specific Chinese models from government networks, and bills to extend a ban across all federal agencies are moving. On September 8 the NSA, CISA, and FBI issued a joint advisory naming six Chinese labs for what it called aggressive, targeted distillation of American frontier models. None of this is a ban on open weights as a category, and the American open-weight ecosystem exists and is improving. But a regulated firm, a government contractor, or anyone whose customers ask where the model came from has a shortlist that is shorter than the leaderboard, and the shortlist should be written down before anyone evaluates quality. The constraint is set by the buyer's industry, not by the technology: the legal model in the example above was post-trained from a Chinese open-weight base, which is a reasonable choice for a law firm's contract review and an impossible one for a defense contractor.

Your data and your alpha. Every hosted model comes with terms about what happens to what you send it. The contract questions are the same for every provider: is content stored, can it be used for training, what safety signals are retained, and who can inspect an exception. Zero-retention terms are available from the major labs for enterprise agreements, and they are worth negotiating, but a no-training clause does not by itself cover logs, safety classifiers, or subprocessors. The deeper point is the one our sovereign AI playbook makes: the asset is not the prompts. It is the know-how a year of use produces, and where that compounds is a choice you make by contract and by architecture, not by which model you pick.

When post-training pays

Training a frontier model from scratch is not on the table for any enterprise; the capital and the talent sit with a handful of labs. Post-training is a different question, and it has become routine. It means taking an open-weight base and continuing to train it on your own examples so that it behaves the way your work requires, usually with a low-rank adapter that touches a small fraction of the weights.

The evidence that this works is now ordinary, and it spans four orders of magnitude of effort.

At the small end, a 350-million-parameter model fine-tuned for tool calling scored 77.55 percent on ToolBench, against 26 percent for ChatGPT and under 3 percent for Claude on the same benchmark, on a task the general models were never taught. Convirza runs Llama-3-8B with sixty LoRA adapters, one per call-quality metric, at a tenth of its previous OpenAI bill and an 8 percent improvement in F1. Across published case studies, a 3-to-7-billion-parameter model with five hundred or more curated examples beats GPT-4-class models on the specific task, and the cost reduction against a pure frontier approach lands between 60 and 80 percent.

At the large end, Harvey previewed Tenet in August: a Kimi K3 base post-trained with Fireworks on roughly 1,750 agentic legal task environments, two months on about 150 NVIDIA B300 GPUs, with nearly twice the completion rate of the base model on held-out legal tasks and state-of-the-art results on a contracts benchmark. NVIDIA's Nemotron releases are themselves post-trained open models, and NVIDIA publishes the post-training datasets it used.

The cost and time between those two ends are collapsing. A LoRA adapter on a 3-to-8-billion-parameter base trains in hours on one GPU, on a few hundred to a few thousand curated examples, for a compute bill in the hundreds of dollars. The expensive part is no longer the training. It is assembling and cleaning the examples and writing the evaluation, and both of those are things a company that has been running agents with an audit trail already has: the corrections reviewers made to agent output are the training set.

Four conditions have to hold before it is worth doing.

1. The task is narrow and specified. You can write the evaluation before you write the training set, and a person can grade an output as right or wrong in seconds.
2. You have the examples, and they are yours. Hundreds to low thousands of real, curated cases from your own history, cleaned of what should not be in a training set.
3. The volume is real. Post-training pays back through inference. A task that runs a thousand times a day pays for its training in weeks. A task that runs ten times a week never will.
4. The general model demonstrably fails. Prompting and retrieval come first, because they let you change the data and swap the base model freely. Post-train only after you have measured a gap that prompting did not close.

When those hold, the economics are not close, and this is work Fig does for customers: post-training a model on the customer's own examples, evaluated on their cases, run inside their boundary or ours, and owned by them. When the conditions do not hold, a post-trained model is a maintenance obligation that goes stale every time the frontier moves.

The industry-model debate, and full control

Legal AI happens to be running this argument in public this month, with one leading company shipping a post-trained model and its closest rival announcing a full ontology of law on top of frontier models. It is tempting to read that as a referendum on training versus not training. It is not. Both companies are putting domain structure somewhere a general model does not have it, one in weights and one in an explicit model the software reasons over, and a serious enterprise ends up doing both in some proportion.

The proportion is what matters, and our view on it is simple. The ontology, the objects, links, and actions and the rules that live on them, outlives every model and makes the next one better on day one, so it comes first. A post-trained model encodes some of that structure into weights, wins on speed and cost for the tasks it was trained on, and gets rebuilt when the base or the task changes, so it comes second and only where the four conditions above hold. Industry-specific models are real, and the trend toward them is real, but for an enterprise they are one component in a stack, not the strategy. The strategy is the best result across the whole stack: context, controls, routing, and whichever models, rented or trained, clear the bar for each task.

The strongest form of sovereignty, full control of the model on your own hardware, fits the same frame. The case for it is real: no provider can change the price, the terms, the retention policy, or the model under you, nothing leaves the boundary, and continuity does not depend on a company that might be acquired or shut off. For classified work, some regulated workloads, and institutions with a national mandate, it wins. The costs are real and usually understated: capability lags the frontier by however long it takes you to evaluate and adopt each open release; you own the evaluation harness, the safety testing, the serving stack, the on-call rotation, and the retraining; idle GPUs are the most expensive model on the market; and owning the weights does not give you custody of the logs, an action boundary, or a rehearsed exit, which are the controls that actually protect an institution. Our operating playbook treats full model control as one rung on a ladder the workload charter chooses. Most enterprises should own the ontology, the context, and the controls and rent the intelligence under enforceable terms. Some should own the intelligence too, and Fig will build and run it for them when the charter says so.

The cost, honestly

Three costs are usually compared badly.

The frontier's cost is the token bill, which is visible, plus the retries and human corrections, which are not. The open-weight tier's cost is hardware, power, hosting, and the engineers who keep it running, and it only wins at utilization: an idle GPU is the most expensive model there is. The post-trained model's cost is the training set, the evaluation harness, and the retraining when the base model or the task changes.

The only fair comparison is cost per accepted outcome, with the failed attempts and the review time in the denominator, measured on your own cases. Run it for a class of task and the answer is usually mixed: frontier for the hard ten percent, open weights or a post-trained model for the routine ninety, and the routing between them is where the savings actually come from.

How Fig thinks about it

We do not have a favorite model, and we would be worse at this if we did.

Fig's routing runs the closed frontier, open-weight models, and post-trained models side by side, and sends each task class to whichever clears its bar at the lowest cost per accepted outcome on our evaluations. The model hub is where an organization sets which of those are allowed for which teams and data classes, which is how the geography and industry constraints above become policy rather than memos. When a customer's task is narrow, high-volume, and failing on general models, we post-train a model for it: on the customer's own corrected examples, evaluated on their cases, run inside their boundary or ours, and owned by them. It is part of the engagement, not a separate product.

That indifference is the point. The frontier will move next month, an open model will close another gap, and somebody's post-trained model will beat both on some task nobody expected. A platform built around one of those outcomes has to be rebuilt when the next one arrives. A platform built to measure and route does not.

Key takeaways

• Three tiers, all useful: the closed frontier for novel work, open weights for control and volume, post-trained small models for narrow high-volume tasks.
• The top of the open-weight tier is Chinese this year, and for defense, government, and many regulated firms that is a hard constraint before any quality question.
• Post-training pays when the task is narrow, the examples are yours, the volume is real, and the general model measurably fails. It routinely beats the frontier on those tasks at a fraction of the cost.
• Compare cost per accepted outcome on your own cases, not price per token. The savings come from routing, not from picking a winner.
• Training versus not training is a false choice. The ontology comes first, a trained model second where the conditions hold, and the goal is the best result across the whole stack. Full model control is one rung on the sovereignty ladder, chosen by the workload.

Should you train your own model? For one task in ten, probably. For the platform, no. Own the context and the controls, measure everything, and let the models compete for your work.