TL;DR: The next era of AI products, agentic products and enterprise AI will be built on top of specialized models trained on proprietary data. Construct Labs is the research and deployment partner for teams building vertical agent products. We handle post-training, continual learning, and custom model inference. If you are interested in working with us, contact us via mail or book an intro call.
For two years the reflex has been simple: point the biggest frontier model at every problem and pay by the token. That worked when usage was small. As agent products hit product-market fit and internal tools reach more users, the economics of large generic models become a real bottleneck.
Over the past few months, token spend has grown exponentially. Harvey went from 1 trillion tokens in January 2026 to 12 trillion in May 2026. SemiAnalysis already finds that employees spend the equivalent of roughly 30% of their salary on internal AI tokens, and the curve only steepens. We are entering a period where uncontrolled token spend exceeds payroll.
The first teams to react are not buying more frontier capacity. They are getting deliberate about what runs where. Coinbase moved its default to open models like GLM 5.2 and Kimi 2.7, and cut spend by about half while usage grew.
Open models now track the frontier closely and catch up within months of a new release. Most have already crossed the threshold of broad usefulness.
Swapping to an open source model can match frontier quality at a fraction of the price. But some workflows have to stay at the capability frontier to compete, and a generic cheaper model will not get you there. For those, training a smaller model on your proprietary signal and in-house data builds a real moat on both quality and inference cost. Value accrues at the Pareto frontier of quality and cost, and it accrues per vertical. Generic models drift off that frontier as specialized teams enter the training game. Construct Labs is the research and deployment partner for that; post-training, continual learning, and custom model inference for teams building vertical agent products.
Enter the loop of recursive self improvement
Most AI products today are built on a generic API and stay that way after launch, adapted with prompts and retrieval. You get the same model as every competitor calling the same endpoint. The harness, long system prompts and RAG scaffolding, bend that general model toward your domain, and users adapt to its gaps by stuffing information more into the context. Feedback can only lead to changes in the harness, never the model. The system stays static, and the data and user signals it produces go to the model provider rather than into your own model.
The alternative trains on the signal you already generate. Every correction, edit, accept, reject, and re-query is a labelled record of how you want your work done. You produce this signal for free, every day, and today you are not profiting from it. Open base models are strong enough that this signal, applied through post-training, takes a small model past frontier quality on your task.
However a model trained once still goes stale. Weights freeze at deployment while user behaviour, data distributions, and tools keep moving. Continual learning fixes this by turning your production feedback into weight updates on live rollouts, through algorithms related to On-Policy Distillation, while synthetic RL environments rebuilt from your current tools and data keep training pressure on the parts of the task that shifted. At Construct Labs we focus on algorithms that update your model based on usage data in a zero-touch manner.
For the last two two years post-training a model was a bad idea. Lock-in was high. Training was expensive. Inference economics for a single custom model were terrible. And you forfeited the free improvements every new frontier release delivered. All four constraints have broken. Open base models improve on a fast cadence, and cheap post-training means you carry your data and signal to each new base instead of being stranded on an old one. Shared serving fixes the inference economics. The cost of owning collapsed while the cost of frontier tokens keeps increasing.
We proved this on a real workload. A small open model, finetuned on one job, beats Opus 4.8 on accuracy at roughly 25x lower cost, on held-out data it never saw in training. Full method in the our retrieval agent case study. And a single custom model is not expensive to operate. On a shared autoscaling pool, many finetunes serve at frontier-batch economics. See how we support customers with frontier inference services in our post about multi-LoRa serving.
Work with us
Your finetuned model, served on a shared autoscaling pool sized to your traffic, on a small or large base depending on the job. You bring the data and the workflow. We handle training, serving, and keeping you on the best model over time. If that is the system you want, reach out at hello@constructlabs.com.