Morning briefing · 08:00
Small models make the calls: agents divide the labor
The unit of AI product design is shifting from a single general-purpose model to a combination of dedicated models, each with its own role. Products and research released yesterday all point toward handing agents' branching decisions, user intent detection, design planning and training-data annotation to separate, purpose-built models. But the advantages of combining models are open to attackers too, and capital is still flowing to the largest models.
Dedicated models that split off decisions
The clearest example of this division of labor was OpenAI's Decisions API. It handles short, focused judgments inside apps at low cost, and OpenAI described it as fast, inexpensive decisions usable in any application1. TechCrunch reported it as a broad release of an OpenAI decision model that competes with TypeSafe's judgment model Jev. The API was first introduced on October 20 as something the voice model GPT-Live calls directly. In effect, a structure in which a large model leads the conversation and hands decisions at forks to a small model was first put to use inside OpenAI's own product. Pricing, model names and performance figures have not yet been disclosed.
The same day also brought an example of splitting off the step before the decision into a dedicated model. A lab under China Telecom released Xing4.0-29B-Intent, specialized for intent understanding, as open source24. According to QbitAI, it handles both classification tasks, choosing among options, and generative tasks, filling in free-form output. Intent detection is the first step of routing, where an agent decides where to send a request. Decision models and intent models arriving side by side means an agent's decision path is being defined as independent components.
The same division of labor in generation and data
It is not only decisions that are being split up. Image generation and training-data production also show designs that break work into stages and attach dedicated models. Canva released its first in-house image generation model, Canva Craft, to all users3. According to The Verge, the model has a two-stage structure that plans the design layout before drawing the image. Canva claimed the planning stage raises precision, consistency and control while lowering cost. It is a case of a design platform building a model fitted to its own workflow instead of relying on outside models.
On the data side, NVIDIA released Flora on Hugging Face, a family of open models that automatically annotate actions, time segments, success and causes of failure in task videos6. According to reports, the models were actually used to train GR00T, NVIDIA's humanoid foundation model. NVIDIA said it can cut the cost of producing robot training data by up to 125x, but did not disclose the comparison conditions. Microsoft's ThinkingBox, which post-trains agents by turning database grading results into reinforcement learning rewards, can also be seen as an approach that splits the evaluation role into a separate component9.
The downside of combination, and capital heading to big models
Division of labor is not always put to beneficial use. Security startup Straiker said the hacking group CyberCartel combined several commercial models, including Claude and Grok, to attack about 400 financial institutions in Latin America4. According to reports, AI agents split up target reconnaissance, attack planning and execution. Defenders are responding as well. Anthropic released a new version of Claude Security in public beta that continuously analyzes code changes and tracks whether vulnerabilities are actually fixed, AI Times reported5. It shows that as components multiply, checks must become continuous operations rather than one-off events.
It is also hard to conclude that splitting work across small models reduces the weight of large ones. According to TechCrunch, OpenAI is working to raise $30 billion in new funding at a $1.25 trillion valuation ahead of its IPO2. Meanwhile, the market is reassessing the pace of investment. On October 22 the Nasdaq fell 2.60% and the semiconductor index 6.7%, and Electronic Times (etnews) said concerns over the sustainability of AI investment had grown7. Division of labor is a design for lowering costs, but it moves together with the investment case for the large models that sit at the top.
For companies building agents, the remaining question is less which model is biggest and more which decisions to hand to which model. Splitting roles can cut cost and latency, but it must come with a system that continuously checks the boundaries and permissions between components. Going forward, it will be worth checking the pricing and performance disclosures for the Decisions API, as well as who published the benchmark claim that a Chinese decision model outperforms Microsoft and Jev, and under what evaluation conditions.