Evolutionary algorithms and Japanese models
Sakana AI uses evolutionary algorithms to combine existing generative AI models. This method searches for optimal combinations by imitating the evolutionary process of living organisms through many iterations of trial and error to find optimal results. It finds results that human experts miss during manual experimentation. The company created EvoLLM-JP, which combines Shisa Gamma 7B v1 with WizardMath 7B v1.1 and Abel 7B 002. For example, when asked the sum of the month and day of Setsubun where the first day of spring is February 3rd, the model correctly identifies Setsubun as February 2nd and calculates the sum as 4. EvoLLM-JP scored higher than Llama 2 70B and GPT-3.5 on the MGSM benchmark. The model also handles non-calculative Japanese tasks, like describing sushi or telling jokes in the Kansai dialect. EvoVLM-JP merges LLaVa-1.6-Mistral-7B with Shisa Gamma 7B v1 to provide Japanese image descriptions.
| Parameter | Value/Model |
|---|---|
| Series B Valuation | $2.65 billion |
| Fugu Max Input Cost | $2 per million tokens |
| Fugu Max Output Cost | $6 per million tokens |
| EvoLLM-JP Benchmark | Beats Llama 2 70B and GPT-3.5 |
Strategic investments and regional cooperation
Sakana AI reached a $2.65 billion valuation following its $135 million Series B round in November 2025. MUFG, Khosla Ventures, Macquarie, New Enterprise Associates, Lux Capital, and In-Q-Tel participated in that round. In January 2026, Google invested to advance Gemini AI model adoption in Japan. Citi provided a strategic investment in February 2026 to support international expansion. The investment from Mitsubishi Electric aims to create new solutions by combining its manufacturing knowledge and extensive datasets with Sakana AI’s advanced foundation model technologies across manufacturing and infrastructure sectors. Sakana AI also receives support from the Japanese government through the Generative AI Accelerator Challenge (GENIAC), which provides a large-scale GPU supercomputer for research. Japan and ASEAN also committed to collaborative AI model development during a meeting in Hanoi in January 2026.
Multi-agent orchestration with Fugu
Fugu is a multi-agent system that orchestrates Claude, GPT-4, and Gemini through one API. It sends prompts to these models in parallel so they generate independent responses. An aggregator model reviews all three responses to synthesize a final, consolidated answer. Fugu Ultra uses a fixed pool of agents to maximize answer quality on hard, high-stakes problems. It achieves the best or joint-best score on five internal benchmarks, including GDP.pdf, Chartography, DeepSWE, Toolathon, and SWEFish. Fugu Max uses a larger pool of models to push the cost-performance Pareto frontier, making this model 40-60% cheaper than Sonnet 5, GPT 5.6 Terra, and Kimi K3. Fugu Max costs $2 per million input tokens and $6 per million output tokens. Fugu Cyber specializes in cybersecurity reasoning tasks like vulnerability research and threat investigation, achieving an 86.9% success rate on CyberGym and 72.1% on CTI-REALM. Users must submit an access request form and provide verified contact information to use Fugu Cyber. However, multi-agent orchestration adds latency because the system waits for responses from multiple models and runs an aggregation step. You should skip Fugu if real-time response speed matters more than marginal accuracy gains. Will the industry favor specialized small models or orchestrated frontier models? Fugu Cyber requires users to submit an access request form and provide verified contact information.
