June 23rd news, yesterdayJapan AI Companies Sakana AI publishes Fugu, encapsulating multiple Agent programming systems as a single model API. Once the task is in place, it will decide which model to call, which steps to complete, whether to validate the results and whether to re-refer to itself。
Fugu is divided into two versions: Fugu for daily coding, code review and interactive scenes, with emphasis on performance and delay balance; and Fugu Ultra for more in-depth expert Agent pool, for more difficult issues。
The official benchmark data show that Fugu Ultra scored 73.7 points on SWE Bench Pro, 69.2 points above Opus 4.8; 50.0 points on HLE, slightly above 49.8 points on Opus 4.8。
The official claim that Fugu Ultra is on the same level as Fable 5 and Mythos Preview. The news of June 23rd