Fugu Ultra: AI Orchestration's Speed and Cost Trade-offs

Summary

Sakana.ai has released Fugu Ultra, an AI model that orchestrates multiple frontier models like GPT and Gemini to achieve performance comparable to models like Fable and Mythos, but without export control risks. In tests, Fugu Ultra was pitted against Claude Opus 4.8 across 38 tasks. The results showed that 36 out of 38 tasks ended in a tie, with Fugu Ultra being 4.5 times slower and five times more expensive than Opus 4.8. For example, Opus completed some tasks in 6 seconds, while Fugu took multiple minutes for the same task. The cost difference was significant, with Opus costing around $10 for all tasks, and Fugu costing $50. While Fugu Ultra's multi-agent approach, where a manager model delegates tasks to specialized AI agents, offers a powerful way to leverage different AI strengths and could be beneficial for complex collaborative development environments, it comes at a considerable cost in speed and expense. For many individual use cases, the added complexity and cost may not justify the performance, with existing subscriptions like Codex and Claude Code potentially offering better value.