Does model efficiency matter more than peak capability for real work?

Occamy-1.0 argues that for agents doing "user-directed, multi-step work in a persistent digital environment," practical value depends "not only on peak capability but also on how efficiently that capability is delivered." The reason is economic: "A long task can invoke the model dozens or hundreds of times, so small differences in per-call cost and latency accumulate across the episode." Co-work is defined by "sustained coordination across the complete task rather than by any fixed collection of skills," so the unit that gets billed is the whole episode, not the single answer. The paper reports that Occamy-1.0, obtained by further training the post-trained Qwen3.6-35B-A3B checkpoint, is "consistently among the strongest comparably sized models," stays competitive with "substantially larger frontier systems on several tasks," and, "under our stated evaluation and pricing protocol," sits at the low-cost knee of the observed cost-performance Pareto frontier across four representative benchmarks.

The mechanism the paper offers is a claim about where the work is. Many steps in everyday work "emphasize state tracking, coordination, recovery, and follow-through rather than frontier-scale reasoning." If that holds, a smaller model does not have to rebuild general capability; it can "concentrate learning on the coordination, recovery, and follow-through that real work demands." The training system is built to make that concentration possible: execution-grounded data and environments that connect "task construction to runnable state transitions and task-level outcomes," replayable long-horizon trajectories captured across multiple harnesses, and staged post-training meant to "develop and consolidate complementary execution capabilities." The infrastructure preserves "token-exact trajectories, environment-state replay, and segment boundaries created by history rewrites," which is what lets a long episode be replayed and trained on.

This sits close to What makes an AI system feel like a colleague rather than a chatbot?, which locates the colleague-like quality in state and closure; Occamy adds a cost argument for the same shift and a training route that targets those behaviors directly. It also echoes Can delegation teach models to manage context more actively?, where a trained 30B model rivals models ten times larger: in both, a targeted skill substitutes for raw scale. It complicates Does a single benchmark score actually predict agent readiness? by naming cost per episode as one more axis that peak-capability rankings hide, and it offers a possible reading of the low completion rates in Why do AI agents fail at workplace social interaction?, though the excerpt never engages that benchmark.

The excerpt does not establish how much of the result comes from the execution-centered post-training rather than from the Qwen3.6-35B-A3B starting point; it gives no base-model comparison, benchmark names, scores, prices, data scale, or ablations. The premise that many steps need coordination more than frontier reasoning is stated as motivation, not tested in the excerpt. The Pareto claim is explicitly conditional on the authors' own evaluation and pricing protocol and on four benchmarks. What follows at that strength is narrow: cost accumulated over an episode is a legitimate axis for judging co-work agents, and a compact model post-trained on runnable, replayable environments is one reported way to move along it.