Sakana AI Drops RL Conductor: New 7B Model Orchestrates Frontier AIs

As of May 15, 2026 Sakana AI dropped RL Conductor. The 7B reinforcement-learning model decides which frontier system handles each sub-task. It already posts state-of-the-art scores on reasoning and coding benchmarks while trimming token spend and API round-trips. Look, previous orchestration layers were brittle scripts. They picked one model and stuck with it. RL Conductor treats every request as a live decision problem. It watches performance, cost, and output quality in real time then switches mid-workflow. The result shows up fast on multimodal jobs. Image generators, video pipelines, and code agents now hand off without manual prompting.

Here's the thing: the model learns a policy over model selection instead of relying on static rules. It gets rewarded for hitting quality targets at lower cost. That training loop is what separates it from earlier routers. Creators working with video and image stacks feel the difference first. A single prompt can now trigger GPT-5.5 for planning, Claude Sonnet 4 for detailed description, then Gemini 2.5 Pro for final rendering. No extra glue code required. Advances like Sakana’s RL Conductor demonstrate the next wave of intelligent orchestration that will power more controllable, efficient, and creative AI video and image generation pipelines. Similar routing logic already appears in expert breakdowns of tools like Seedance 2.0 for specialized adult content workflows.

The early numbers look solid. Token usage dropped noticeably versus rigid frameworks. Latency stayed competitive even when it sampled multiple models. Most important, it kept quality high on creative tasks instead of defaulting to the cheapest option. - Dynamic model selection that adapts per step

• Lower API spend without quality cliffs
• Strong results on both coding and open-ended generation
• Works with closed frontier systems and open-source alternatives Plot twist: the biggest win may be simplicity. Teams stop maintaining their own switching logic.

Sakana is rolling the model out through its new Fugu platform. Enterprise customers and research groups sit at the front of the line. Independent creators will likely see limited access later this quarter. Not gonna lie — the pricing and exact rollout timeline remain fuzzy. Early signals point to usage-based credits rather than flat subscriptions. That structure could keep costs reasonable for smaller teams once the gates open. Wild. A 7B orchestrator might end up mattering more than the next 100B base model for day-to-day creative work.