Fugu Ultra

Fugu Ultra by Sakana AI

Fugu Ultra is a long-context general-purpose language model from Sakana AI, released on June 24, 2026 and listed on OpenRouter. Its headline differentiator is a 1 million-token context window, positioning it for tasks that require reading, reasoning over, and generating responses from very large bodies of text. Rather than being optimized only for short conversational turns, Fugu Ultra is aimed at workflows where the model must retain coherence across extensive documents, multi-file inputs, lengthy transcripts, research corpora, or accumulated project history.

The model’s core capabilities are long-context reasoning, text generation, summarization, analysis, and instruction following. A 1M-token context window can support use cases such as comparing hundreds of pages of policy text, synthesizing large legal or technical archives, tracing arguments across long research papers, reviewing extensive logs, and maintaining continuity in complex planning sessions. For developers, analysts, and researchers, the main benefit is reduced need for aggressive chunking or retrieval orchestration before prompting the model.

Technical specifications

• Provider: Sakana AI
• Release date: 2026-06-24
• Availability: Listed on OpenRouter
• Context window: 1,000,000 tokens
• Primary modalities: Text input and text output, based on the available listing information
• Capabilities: Long-context reasoning and general-purpose text generation
• Maximum output length: Not specified in the provided listing details; users should verify current limits on OpenRouter
• Pricing: Not specified here and may vary by OpenRouter routing or provider configuration
• Open-weight status / license: No open-weight or license information is indicated in the provided details

Fugu Ultra’s main strength is its ability to operate over unusually large prompts. This can simplify application design where context length is the bottleneck: fewer document chunks, fewer retrieval misses, and less manual condensation of source material. It may be especially useful for workflows that require global consistency, such as producing an executive synthesis from a large evidence base or answering questions where the relevant details are scattered across a long corpus.

The most important caveat is that a large context window does not automatically guarantee perfect long-context understanding. Models can still miss details, overweight nearby text, hallucinate unsupported conclusions, or struggle with precise retrieval inside very large inputs. Cost and latency may also become significant when using hundreds of thousands of tokens per request. In practice, users should still evaluate Fugu Ultra on their own workloads and consider hybrid retrieval, citation checking, or staged summarization for high-stakes tasks.

Compared with earlier generations of general-purpose LLMs that typically supported far smaller contexts, Fugu Ultra’s 1M-token capacity is its defining advance. It sits in the same broad category as other long-context models from major providers, but its appeal will depend on measured quality, output limits, pricing, latency, and reliability under real workloads. Relative to Sakana AI’s previous public emphasis on model innovation and automated AI research techniques, Fugu Ultra appears positioned less as a narrow research artifact and more as a practical long-context text model for production-style reasoning and generation.

A short practical note: models with very large context windows can also help with software maintenance tasks such as reading long changelogs, dependency manifests, audit reports, and migration notes in one pass, though such uses should include verification against source data.