An eight-billion-parameter model just beat a 402-billion-parameter model at the one job agents actually perform all day: picking the right tool. Meta and Carnegie Mellon researchers report that a Qwen3-8B model, post-trained with reinforcement learning, selected the correct tool 41.5% of the time on ToolBench G3. Meta's own Llama-4-Maverick — roughly fifty times larger — scored 30.6%.
The inversion is real, and it is also narrower than it looks. The training data came from the same benchmark family the model was tested on, and a 41.5% success rate still means the best model misses more often than it hits when choosing among more than ten thousand tools. That is exactly the point: narrow is where the volume is.
Last week we covered Qwen's 6B-active model posting strong SWE-bench numbers and PilotDeck's routing layer. This week the story is different: the gains are coming from post-training alone, and a new class of purpose-built decision models is absorbing the high-frequency work frontier labs were never cost-effective for.
The gains are coming from post-training alone, and a new class of purpose-built decision models is absorbing the high-frequency work frontier labs were never cost-effective for.
Start with the Meta/CMU result as more than a curiosity. The team's Plan-Retrieve-Select pipeline first decomposes an instruction, then pulls a shortlist of candidate tools, then picks. They trained Qwen3 models from 1.7B to 8B with supervised fine-tuning followed by RL using GRPO, drawing on roughly 21,000 ToolBench G3 instructions. The results held across sizes: the 4B variant hit 40.3%, and even the 1.7B variant reached 37.6% — comfortably clearing the 402B model. The trained models also transferred to different retrievers and new tool corpora, which matters more than the headline number. A specialist that only works on its home benchmark is a demo; one that generalizes across domains and retrieval stacks is infrastructure.
Now widen the lens. DeepSeek's V4-Flash 0731 kept its architecture and size completely frozen — 284B total parameters, 13B active — and rose on all nine agent benchmarks through post-training alone. Terminal-Bench 2.1 went from 61.8 to 82.7. DeepSWE gained 47.1 points. CyberGym gained 38. This was not a new base model. It was the same weights, a better training recipe, and a jump that moved the model into first-tier conversations.
That is the strongest available evidence that the binding constraint on agent capability is not scale. It is the post-training recipe — the reward structure, the task distribution, the execution-focused training data. DeepSeek explicitly did not build a bigger base. It pushed its training toward agent execution, tool calls, and long-horizon tasks. The lesson generalizes: if you want an agent that picks tools well, you train for tool selection. You do not wait for a larger generalist.
The market is already pricing this in. OpenAI's Decisions API — its answer to the decision-model wave — runs at 10 cents per million input tokens, one cent per million cached, and charges nothing for output. It is built on GPT-6 Luna, and an OpenAI product lead told TechCrunch it is, in effect, "just plain old Luna that gets the job done." OpenAI claims near-real-time decisions up to ten times faster than Luna alone through the Responses API, cutting decision time from three seconds to as low as 100 milliseconds. The frontier lab is unbundling its own cheapest model into a decision endpoint.
Ant's LingDT-Decision-1.0 comes in 4B and 27B versions and claims more than 90% token savings versus large language models on the same tasks, with roughly 10x faster response. In data-labeling tests it cut token consumption about 97% versus DeepSeek V4-Flash; in customer service, risk-detection accuracy rose from 76% to 89% versus Qwen-3.7-Flash. Across 23 open benchmarks, the 27B version's composite score led Microsoft-Decision-1 and Jev 1.13.0.
China Telecom's Xing4.0-29B-Intent attacks the same niche from the intent side. It handles both intent classification and parameter extraction — "doing multiple choice and fill-in-the-blank," in the lab's framing — and scored 58 overall across 6,000 evaluation items, ahead of the best general model tested at 49. On pure category decisions it scored 70, ahead of GPT-6-Luna at 65 and well ahead of Jev's two answer modes at 56 and 53.
And then there is the price floor. Together AI released tev1-4B-experimental, a Jev-like classifier fine-tuned on Qwen3.5 4B, serving at $0.042 per million input tokens with zero output cost. The training cost was $17. Seventeen dollars to train a decision model that serves at four cents per million tokens.
Sit with that number against the frontier layer's capital requirements. The specialist layer is trainable for pocket change and priced to match. The frontier layer still costs billions to build and will keep costing billions, because the hardest long-horizon engineering tasks still favor GPT-5.6 Sol, Claude Opus 5.5, and Kimi K3. On DeepSWE, V4-Flash 0731's 54.4 still trails GPT-5.6 Sol's 18.6-point gap in the other direction — the frontier models remain the choice when the chain is long and the stakes are high.
So the agent stack is splitting cleanly. Underneath: a proliferating layer of 4B-to-29B specialists, each RL-trained for one decision shape — tool selection, intent routing, parameter extraction, risk scoring — serving at commodity prices and improving through recipe iteration, not scale. On top: a shrinking set of frontier generalists reserved for the tasks that genuinely need deep, multi-step reasoning across unfamiliar domains.
The specialist layer's rise is the demand-side complement to the routing and harness economics we covered on 09/21 — but the argument here is about the models themselves, not the router that calls them.
The strategic question is no longer whether small models win at tool use. The numbers have settled that. The question is which layer captures durable margin. The specialist layer's moat is thin: a competitor can replicate your $17 training run, and Together AI is publishing the data recipe and tutorial so they will. The frontier layer's moat is capital, and capital is abundant. What neither layer has figured out is where the recurring value sits when the training cost of a competent specialist approaches zero and the frontier model above it gets invoked once per thousand calls.
For founders, the implication is that the decision endpoint is becoming a commodity input, and the defensible work is the task distribution you train against — the proprietary corpus of real decisions your business makes a million times a day. For investors, it means the volume economics of agents will be decided in the 4B-to-29B range, not at the frontier, and the models that capture that volume were never designed to be general. They were designed to pick the right tool, route the right ticket, and extract the right parameter — cheaply, quickly, and every single time.