Developer communities circulated a striking data point this week: on the DeepSWE software engineering benchmark, a lightweight model at $0.61 per-task cost scored highest, while Claude Opus, GPT-5, and other models costing several times more failed to beat it. We're not focused on the capability scoreboard — we're tracking the next signal for the AI industry: the "expensive thinks, cheap does the work" division of labor is solidifying.
What this is
The source is OpenAI's Codex code editor user community. Developers have worked out a workflow: keep the expensive reasoning model on the main thread for planning and decision-making, while configuring cheap lightweight models as "sub-Agents" (think AI assistants dispatched to handle grunt work) to execute repetitive tasks like code scanning, file checks, and testing.
The logic is straightforward. Expensive models are indeed stronger, but OpenAI itself acknowledges a phenomenon: the longer the conversation, the more likely the model suffers "Context Rot" — forgetting earlier agreements, repeating mistakes, and hallucinating. The fix is to isolate execution-heavy work into separate threads, keeping the main thread clear-headed.
Industry view
The supportive voices are loud. The DeepSWE leaderboard showing a $0.61 per-task model at the top proves that for a large share of execution tasks, cheap models are already good enough. Developer communities are spreading a calculation: running this configuration on a ChatGPT Plus subscription can deliver 20x the output of a Pro subscription.
But the dissent is worth hearing. One objection: once tasks are split to sub-Agents, the main Agent's coordination costs rise. If the split is poorly designed, the main Agent ends up spending more tokens parsing sub-Agent outputs — overall cost may not actually drop. A deeper risk is more fundamental: AI coding tools cycle through a new "optimal setup" every three months. The carefully tuned sub-Agent configuration today could be legacy baggage by next quarter.
Impact on regular people
For enterprise IT: budget models may need to change. The old approach was "one model handles everything." The future is "expensive handles thinking, cheap handles execution" — accounting logic will increasingly resemble hiring people.
For individual careers: even if you're not a programmer, the "expensive thinks, cheap does" division will seep into everyday AI tools. You may need to get used to "commanding an AI team" rather than "asking an AI."
For consumer markets: consumer-grade AI products like ChatGPT will likely split into "thinking edition" and "working edition" SKUs, billed separately by use case.