Sakana AI

Sakana AI is the Tokyo lab founded in 2023 by David Ha (CEO), Llion Jones (CTO), and Ren Ito, with Ha and Jones both Google alumni and Jones a co-author of the original Transformer paper, and it is selling a different thesis than everyone else on this page. Sakana does not ship a frontier model. It ships Fugu, an orchestrator: a request goes to one API and Fugu routes the work across a pool of other models, then stitches the answers back together. The family moved fast through 2026, from an April beta to June general availability with Fugu Ultra v1, a July Fugu Cyber and Claude Code interface, and an August consumer push through Sakana Chat alongside an NVIDIA partnership that brought Nemotron models into the pool. On September 11, 2026 Sakana split the line in two. Fugu Max v1.0 optimizes cost at $2 per million input tokens and $6 output with cached input flat at $0.25 regardless of context length, reporting best overall score on six benchmarks including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and the internal SWEFish, and an output rate 40 to 60 percent under Claude Sonnet 5, GPT-5.6 Terra, and Kimi K3. Fugu Ultra v2.0 optimizes capability at $5 and $30, rising to $10 and $45 once context passes 272K tokens, taking best or joint-best on five of eight benchmarks and top two on seven of eight, with Chartography at 48.3 against Opus 5 at 27.3 and Fable 5 at 29.5, and DeepSWE at 74.3. The claim that gives the thesis teeth is a footnote on Sakana's own chart: Fable 5, Fable 5.1, and GPT-6 Astra are not in Fugu Ultra v2's pool, so the scores come from orchestrating open and specialized models rather than reselling frontier access, and Sakana pitches the swappable pool as insulation against vendor lock-in, API revocation, and geopolitical supply risk. Both models carry 1 million token context and speak OpenAI-compatible Chat Completions and Responses plus an Anthropic-compatible Messages endpoint. Read the billing model before committing, because it is the genuinely novel part: on Fugu Ultra and Fugu Cyber, orchestration tokens are reported separately inside token_details but bill at full input and output rates, so the invoice covers reasoning the user never sees and cost per request cannot be derived from visible tokens. On Fugu Max, web_search and web_fetch run through internal tools at $0.007 per call because the open models in its pool lack native search, and one query can take several calls. Fugu is not yet available in the EU or EEA while Sakana works toward GDPR compliance. All benchmark figures are self-reported and SWEFish is Sakana's own.