Sakana Fugu opens beta with OpenAI-compatible API

• Sakana AI opened beta access to SakanaAILabs' launch post, positioning Fugu as a multi-agent orchestration layer that routes work across a pool of open and closed models via an official OpenAI-compatible API page.
• According to SakanaAILabs' launch post, the beta ships in two modes: Fugu Mini for lower-latency orchestration and Fugu Ultra for deeper reasoning with the full model pool.
• hardmaru's internal-use post says Sakana had already been using Fugu for research and coding internally, while hardmaru's follow-up highlights a recursive mode where Fugu can call itself again to revise its own workflow.
• The headline benchmark claim comes from the official launch thread, which says Fugu set new SOTA results on SWE-Pro, GPQA-D, and ALE-Bench, with more detail on the launch page.
• Access is still gated: SakanaAILabs' beta-credits post says Sakana is approving early testers and handing out free API credits through a signup form.

You can read the launch page, inspect the Japanese version, and even see Sakana sketch one of the more unusual bits in its training diagram, where Fugu first learns to use a pool of LLMs and then gets trained for recursive self-calls. hardmaru's thread adds that the system was already part of Sakana's own research and coding stack before this beta opened.

Sakana is selling Fugu as an external runtime, not a foundation model. In SakanaAILabs' launch post, the company says developers can hit it through an OpenAI-compatible API, while the official beta page frames the product as a system that picks model combinations and assigns roles per task.

The exposed product split is simple:

• Fugu Mini: fast orchestration, lower latency, per the launch post
• Fugu Ultra: full model-pool orchestration for deeper reasoning, per the same post
• Existing OpenAI-style integrations should need minimal changes, according to SakanaAILabs' API description

That packaging is the interesting part. The model routing logic, role assignment, and collaboration pattern sit behind one API surface instead of living in each user's agent harness.

The official diagram in SakanaAILabs' Japanese post shows a two-step training story. First, Fugu learns to use a pool of open and closed LLMs. Then Sakana continues training it to make recursive calls, labeled "inception" in the diagram.

hardmaru puts that into plainer product language: when recursion is enabled, Fugu reads its own prior output and spins up corrective workflows on the fly. Combined with his earlier note that the system dynamically picks the best model mix, the beta looks less like a fixed agent pipeline and more like a controller that can re-open the task mid-run.

The launch claim that matters most is still benchmark performance. SakanaAILabs says Fugu hit SOTA on SWE-Pro, GPQA-D, and ALE-Bench, and the launch page is the canonical source for whatever methodology and leaderboard detail Sakana chooses to publish.

What is new on day two is the rollout mechanic. In its follow-up post, Sakana says it is giving free API credits to early beta testers, and the company is still routing access through an application form rather than opening the endpoint broadly.

That keeps the release in a familiar 2026 pattern: broad claims about autonomous multi-agent performance, but actual usage still starts as a gated beta with credits attached.