Tokyo startup Sakana AI has launched Fugu, a system that conducts a swappable pool of language models to rival Anthropic's restricted Fable and Mythos models.
Key Points:
- Sakana AI's Fugu runs as one model yet coordinates a pool of other systems behind a single API.
- Fugu Ultra scored 73.7 on the SWE-Bench Pro coding test, topping several frontier rivals.
- The design is pitched as a hedge against the export controls that locked away Fable and Mythos.
Sakana Fugu Orchestrates Models
The Tokyo lab rolled out Fugu and a heavier Fugu Ultra tier on Jun. 22, both reachable through one OpenAI-compatible endpoint, it confirmed. Depending on the request, it answers a task alone or pulls together a team of other systems.
The model then runs the checks and the synthesis on its own.
Fugu is itself a language model.
Trained to call agents from a swappable pool, it can even summon copies of itself when a single job needs more hands than one model can supply. The base tier targets low latency for everyday coding, chat, and tools like Codex, and lets teams drop specific agents to meet privacy rules. Fugu Ultra instead chases top answer quality on long problems like paper reproduction and security analysis, which a group of roughly 500 beta users tested over recent weeks.
Also Read: Is The Anthropic Perp Sell-Off A Warning For Pre-IPO Crypto Bets?
Mollick And Levie Weigh In
Benchmark figures the firm released put Fugu Ultra at 73.7 on the SWE-Bench Pro coding test, ahead of Opus 4.8, Gemini 3.1 Pro and GPT-5.5 on the same run.
The company says those scores stand level with Fable 5 and Mythos Preview, and its own table shows the orchestrator finishing on top in 10 of the 11 published rows.
Not every tester was sold. AI researcher Ethan Mollick wrote that Fugu Ultra ran "incredibly slow," with routine coding tests dragging to 30 minutes and output that trailed Fable in real use. Box chief Aaron Levie sounded warmer, calling the single-API routing of each task to the best-suited model a step forward for how applied AI gets built.
Others flagged the price, since the orchestration can stack up token costs many times higher than calling a single frontier model directly on a comparable task. Sakana frames the pooled design as insurance against any provider going dark, flagging the fresh export curbs on Fable and Mythos as the kind of shock that can sever access overnight.
Sakana AI's Origins
Sakana AI took shape in 2023 under Llion Jones, a co-author of Google's "Attention Is All You Need" paper. David Ha, once a research lead at Stability AI, joined him as co-founder. The lab built its name on evolutionary model merging and the AI Scientist line of automated research, and it has long argued that coordinated pools of models can outperform any single system on the hardest, long-running work.
Read Next: Mane City Mobile Lands On iOS And Android In 100+ Countries