Sakana: Fugu Ultra v2

Sakana: Fugu Ultra v2 is a text model from Sakana, and it is the higher-performance member of Sakana AI's Fugu family. Fugu is not a single monolithic model. It is a learned multi-agent orchestration system, in which a language model is trained to coordinate the work of multiple agents. Fugu Ultra v2 sits at the top of that family, so if you are choosing within the Fugu line, this is the variant intended for the most demanding work.

The model accepts text, images and files as input, and it produces text as output. Its context window is 1,000,000 tokens, which is large enough to hold very long documents, whole codebases or extended conversation histories in a single request. The maximum output is 128,000 tokens, so long generations are possible alongside long inputs. Its knowledge cutoff is 2026-08-28, and the provider release date is 2026-09-11.

Because it takes images and files as well as text, Fugu Ultra v2 fits tasks that mix formats: summarizing or extracting data from scanned documents and PDFs, answering questions over screenshots or diagrams, and reviewing long reports where the source material arrives as an attachment. The large context window suits repository-level code analysis, multi-document research, and agentic workflows that need to keep a lot of state in view. The supported parameters point in the same direction. Tools and tool choice let the model call functions, structured outputs let you constrain responses to a schema, and web search options let it reach beyond its training data. Reasoning controls (include reasoning, reasoning, and reasoning effort) let you decide how much internal deliberation the model does before answering, which is useful when you want to trade latency for care on hard problems.

You call the model through api14's OpenAI-compatible API using the model id "sakana/fugu-ultra-v2". Because the API follows the OpenAI format, existing OpenAI SDKs and clients work with only a base URL and key change. The request parameters supported for this model are include_reasoning, reasoning, reasoning_effort, structured_outputs, tool_choice, tools, and web_search_options. If you rely on a parameter that is not on that list, it is not supported here.

One practical consideration is that the model outputs text only. If your pipeline needs generated images, audio or other non-text output, you will need a different model for that stage. The reasoning controls also add a decision: higher reasoning effort can improve answers on complex tasks but may increase response time, so for simple, high-volume requests you may prefer lower effort or a lighter model in the Fugu family.