The $1,339 vs. $10,565 Lesson: Why Smarter AI Planning Makes Cheap AI Workers Viable

The $1,339 vs. $10,565 Lesson: Why Smarter AI Planning Makes Cheap AI Workers Viable

What if you could cut your AI software costs by 87 percent without sacrificing quality? That is not a hypothetical. Cursor just ran the numbers, and they are hard to ignore.

Cursor, the AI-powered coding platform recently acquired by Elon Musk's SpaceX for $60 billion, pitted two versions of its agent swarm against each other in a rigorous real-world test. The challenge: rebuild SQLite, one of the most widely deployed databases in the world, entirely in Rust, using only the 835-page SQLite manual. No source code. No test suites. No internet access. The new system, which separates "planner" agents from "worker" agents, scored 100 percent on the sqllogictest benchmark in every configuration tested. The older system, which used a single-tier approach, bogged down in merge conflicts of its own making, generating more than 70,000 conflicts compared to fewer than 1,000 for the new architecture.

The cost data is where this story gets especially relevant for anyone making decisions about AI spending. Four configurations were tested: GPT-5.5 running solo, Grok 4.5 solo, Opus 4.8 as planner paired with the cheaper Composer 2.5 as worker, and Fable 5 as planner with Composer 2.5 as worker. The cheapest configuration, Opus 4.8 plus Composer 2.5, cost $1,339 total. The most expensive, GPT-5.5 solo, came in at $10,565. In the GPT-5.5 run, workers alone consumed $9,373 of that bill. In the Opus-plus-Composer hybrid run, the entire worker fleet cost just $411. Composer 2.5, the cheaper worker model, benchmarks at the same level as Opus 4.7 and GPT-5.5 but costs only $0.50 per million input tokens and $2.50 per million output tokens. Cursor's founder confirmed the model is built on top of the Chinese open-source model Kimi K2.5. The key insight from Cursor is this: only a small portion of any large task actually requires the reasoning power of a frontier model, specifically the upfront task breakdown and key design decisions. Once a frontier planner resolves the ambiguity in a task, cheaper models can execute the plan at a fraction of the cost.

The code quality data drives this point home further. In the Fable 5 configuration, the old swarm produced 64,305 lines of engine code. The new planner-worker architecture needed only 9,908 lines to hit 100 percent on the test suite, an 85 percent reduction in code volume at the same or better results. The old single-agent approach also suffered from what Cursor calls "split-brain design," where two planner agents would independently build the same feature in different places and in different ways, creating compounding rework. The new system addressed this by having agents record decisions in shared design documents, with code linked back to those documents through compile-time references. A neutral third agent stepped in to resolve merge conflicts when they arose.

For small and mid-size business owners, the practical lesson here goes well beyond software development. The planner-worker framework Cursor validated is a model for how AI should be deployed across any complex workflow, whether that is marketing, customer service, content production, or operations. Most businesses default to throwing the most powerful and most expensive AI model at every problem. What this research demonstrates is that the real leverage lies in the planning layer. A capable frontier model that clearly defines the goal, breaks it into discrete steps, and resolves ambiguity upfront enables a much cheaper execution layer to do the rest. The business that masters that separation will outperform the one simply paying for the biggest model.

This also signals something important about how AI tool costs are going to evolve. The gap between frontier model capability and the cost of capable worker models is already dramatic. Composer 2.5 matches the performance of models that cost many times more per token. As more purpose-built, task-optimized models emerge, the businesses with smart AI architecture will have a structural cost advantage over those running everything through a single expensive model. That advantage compounds over time, especially for businesses running AI at scale across marketing campaigns, support workflows, or content pipelines.

This week, map one repeating workflow in your business where AI is currently involved or where you are considering it. Identify the decision-making and planning steps versus the execution steps. Then ask whether those two phases are currently handled separately or lumped together. If they are lumped together, you are likely overspending on execution by using a high-cost model for work a cheaper, well-instructed model could handle. Designing that separation deliberately is the single highest-leverage architecture decision you can make in your AI stack right now.

The businesses that grow fastest with AI will not be the ones with access to the most powerful models. They will be the ones that know exactly where intelligence is required and where efficient execution is enough.

Originally inspired by: Cursor's agent swarm suggests cheaper models can handle most coding when frontier models plan the work (https://the-decoder.com/cursors-agent-swarm-suggests-cheaper-models-can-handle-most-coding-when-frontier-models-plan-the-work/) See how Leads to Conversion can help you build a smarter, more cost-efficient AI strategy. Get your free AI audit