Sakana AI opens Fugu agent orchestrator to public, benchmarks ahead of Fable 5
Sakana AI's Fugu, a small language model trained to route tasks across larger LLMs, is now publicly available with transparent per-token pricing starting at $5/$30 per million tokens for the Ultra tier.
Sakana AI opened public access to Fugu this week, two months after launching the agent orchestration system in closed beta. Fugu is a small language model trained to dynamically route tasks across larger LLMs — deciding which model to call, in what role, and with what subtasks — rather than requiring developers to hard-code workflows and agent hierarchies.
The system's Ultra configuration outperforms Fable 5 and Mythos on several benchmarks, according to Sakana's published results. Standard Fugu charges only for the underlying top-tier model invoked; Fugu Ultra runs $5 per million input tokens and $30 per million output tokens, with subscription tiers also available.
What stands out
- 01Learned orchestration replaces scripted workflows. Fugu is trained to route tasks adaptively rather than following fixed agent hierarchies, reducing engineering overhead for multi-model systems.
- 02Transparent pricing. Standard Fugu passes through the cost of whichever frontier model it invokes with no markup. Ultra's $5/$30 per million tokens is explicit and undercuts several competing agent platforms.
- 03Two-month beta cycle. Sakana announced Fugu in April 2026 and held it in closed beta until this week, suggesting the routing model required tuning before general release.
- 04Benchmark claims against Fable 5 and Mythos. Sakana's published results show Fugu Ultra ahead on their test suite, though the tasks and eval methodology aren't detailed in the public announcement.
- 05Subscription option available. Beyond pay-per-token, Sakana offers subscription pricing for teams with predictable monthly volume.
