Sakana AI listed Fugu Ultra v2 on OpenRouter this month, and the listing describes something unusual: not one model but a trained router: it hands work to a set pool of specialized and open models, and it can call copies of itself. Sakana pitches it at full-stack software work, autonomous research and hard reasoning that takes many steps, and says no proprietary frontier model sits in the pool.
Per Sakana, it can search the web natively, accept PDFs and images, return structured outputs, call functions, and run at three settings of reasoning effort. Context runs to 1 million tokens, with up to 128,000 tokens of completion. None of this comes with a published benchmark score. OpenRouter’s listing carries no quality evaluation for Fugu Ultra v2, so any claim about how well it reasons relative to GPT class or Claude class models is currently untestable from public data.
What is measurable is cost and behavior, and OpenRouter measured both live. On paper it costs $5 for every million tokens sent and $30 for every million generated; cached reads are $0.50 per million, and each 1,000 web searches add $10. What customers actually paid over the window OpenRouter measured looks different: about $1.277 per million tokens in, dragged down by a cache hit rate near 95 percent, but $33.97 per million tokens out, which is higher than the list price.
Throughput and latency are where the orchestration design shows its cost. Median (P50) throughput was 8 tokens per second. Median latency to first token was 11.99 seconds, and median end-to-end latency was 22.82 seconds. The tail is longer: P99 end-to-end latency reached 1,462 seconds, more than 24 minutes, consistent with a system that can recursively call itself before returning an answer. Tool call errors ran at 0.47 percent and structured output errors at 0.12 percent, both single-digit-percentage failure rates on the mechanics OpenRouter tracks.
Availability is the more concrete signal. OpenRouter recorded 100 percent uptime over three days, meaning the endpoint was reachable, but successful inference completed only 88.30 percent of the time over that same three-day window, improving to 93.87 percent over the trailing 24 hours. That gap between “the server answered” and “the server finished the job” is the number an operator building on this model would want to reproduce before trusting it in a pipeline with a deadline.
Three named applications account for most of the traffic OpenRouter tracked: the largest was Hermes Agent at 112 million tokens, followed by Codex at 63.4 million and an app called pi at 53.1 million. Sakana’s other listings sit at different price points: Fugu Max is the cheaper tier, charging $2 for a million tokens in and $6 out while drawing on open-weights systems such as the Nemotron family from NVIDIA; Sakana Namazu is a reasoning model built on Kimi K2.6 and tuned for Japanese; and the original Fugu Ultra remains listed.
For a team evaluating orchestration systems rather than single models, the number to track is not a benchmark score, since none exists here, but the availability rate against the SLA a production workflow needs: an 88 percent three-day success rate is a different reliability class than the 99-plus percent uptime frontier API vendors typically advertise.
Capability details are drawn from Sakana AI’s own listing; throughput, latency and availability figures are OpenRouter’s live measurements published on its Fugu Ultra v2 model page, collected mid-September 2026.
