Sakana: Fugu Ultra

Sakana: Fugu Ultra is a text generation model from Sakana AI, and it sits at the top of the company's Fugu family as the higher-performance option. What makes the Fugu line unusual is its architecture: rather than being a single monolithic model, Fugu is a learned multi-agent orchestration system, in which a language model is trained to route work across multiple agents. Fugu Ultra is the larger, more capable member of that family, so if you are evaluating Sakana's approach, this is the variant aimed at demanding workloads rather than lightweight ones.

On the input side, Fugu Ultra accepts text and images, and it produces text as output. That combination means you can send prompts that mix written instructions with visual content, and receive a single text response. The context window is very large at 1,000,000 tokens, which allows you to work with long documents, extended conversations or sizeable collections of material in one request without chunking them yourself. The maximum output length is 128,000 tokens, so responses can be correspondingly long when a task calls for it. Supported request parameters include reasoning controls (reasoning, reasoning_effort and include_reasoning), structured_outputs, tool use via tools and tool_choice, and web_search_options.

Those capabilities point to a fairly specific set of practical uses. The million-token context suits tasks such as analysing long reports, contracts or codebases, or maintaining continuity across a lengthy session. Image input opens up document understanding, screenshot analysis and mixed media workflows where the model needs to read both prose and visuals. Because tool calling and structured outputs are supported, Fugu Ultra also fits agentic pipelines: you can let it decide when to call external functions and constrain its replies to a schema your application can parse. The reasoning parameters let you dial effort up or down depending on how much deliberation a task needs, and web_search_options is available when a request benefits from search.

Calling the model through api14 is straightforward because the gateway exposes an OpenAI-compatible API. You send your requests to the standard chat completions endpoint and set the model id to "sakana/fugu-ultra". The parameters listed above can be passed in the request body in the usual way, so existing OpenAI client libraries and code that already targets that interface should work with minimal changes, typically just swapping the model identifier and pointing the base URL at api14.

There are a couple of things to keep in mind. Output is text only, so if you need generated images or audio, this is not the model for that. The supported parameter list does not include sampling controls such as temperature or top_p, so if fine-grained control over randomness matters for your use case, check whether another model on api14 exposes those knobs. Otherwise, for long-context, multimodal, tool-using text tasks, Fugu Ultra is a reasonable candidate to test.