Cohere Cuts Multi-Agent AI Costs by 70 Percent

Cohere published a controlled study evaluating how to organize teams of AI coding agents efficiently on ProgramBench, a 200-task benchmark requiring models to reconstruct software executables. Testing multi-agent configurations with GPT-5.6 models—Sol, Terra, and Luna—the researchers found that while basic multi-agent setups achieve up to 10 percentage point gains over solo agents, they typically cost 3 to 7 times as much. For example, a single Sol agent scored 52.1% under a minimal prompt, whereas a blocking team scored 61.8%.

To resolve this efficiency penalty, Cohere tested delegating specific workflow phases: exploration and planning, implementation, and verification. Delegating only planning to workers allowed an asynchronous Sol lead to match full-team performance—scoring 63.6% versus 63.2%—while cutting compute expenses in half from $510 to $249. Swapping expensive worker models for cheaper alternatives yielded even larger savings. Pairing a Sol lead with 16 inexpensive Luna planning workers achieved a 61.8% score at a total cost of $155, achieving 70% savings compared to an all-Sol team delegating every stage.

Cross-model evaluations on a 30-task subset revealed that multi-agent gains depend heavily on model capability. GPT-5.6 and Claude Opus 4.8 gained 6.9 and 5.9 percentage points over solo baselines, respectively. In contrast, Claude Opus 4.7 rarely delegated tasks, and Gemini 3.8 Flash lost 4.6 points in asynchronous setups. For practitioners, these findings show that multi-agent scaling does not require unconstrained compute budgets. Instead, developers can optimize real-world agent architectures by identifying task bottlenecks, deploying heterogeneous model hierarchies, and restricting expensive worker delegation to information-gathering stages.