Fugu Ultra

Sakana AI's September 2026 release page introduces Fugu Max and Fugu Ultra v2 as two optimizations of the same core orchestration architecture, targeting the cost-capability Pareto frontier from different angles. Fugu Max is positioned as expanding the Pareto Efficiency Frontier by orchestrating the largest pool of ope Technically, the release frames Fugu Ultra v2 as the same architecture as Fugu Max but retuned for the question "What is the absolute highest capability we can achieve on complex, multi-step tasks?" — in contrast to Fugu Max's "best possible output at the lowest possible cost" mandate. For developers consuming Fugu Ult

Sakana AI shipped Fugu-Ultra v1.1 on July 24, 2026, per a Coursiv explainer tracking Sakana's official X announcement, roughly a month after the original Fugu and Fugu Ultra launch on June 22. The article reports that v1.1 introduces newer frontier models into Fugu's agent pool, with Sakana claiming gains of up to +7.9 The piece documents the version naming formally: fugu-ultra-v1.1 is the new identifier, while the original fugu-ultra-20260615 is renamed to fugu-ultra-v1.0. Pricing is listed jointly for both versions at $5 per 1M input, $30 per 1M output, and $0.50 cached input, with double rates above 272K context. The article flags

An independent Substack explainer from XYZ Labs (June 22, 2026) walks through the Fugu and Fugu Ultra launch, highlighting that Sakana's bet is on a better "manager" rather than a bigger brain. The article documents that Fugu decides which underlying models to use, how to split work, when to verify, and when to call an For developers integrating Fugu Ultra via gateways like Merge Gateway, the article underscores the practical distinction: Fugu is not a static routing rule but a trained LLM that orchestrates an agent pool. The piece positions Fugu Ultra as the performance-maximizing tier with a deeper, quality-first approach through a

An independent Medium analysis argues that Sakana's Fugu Ultra is not a new frontier model but rather a multi-agent orchestrator that uses frontier models to achieve its benchmark numbers. The article cites Sakana's technical report showing that Fugu Ultra routes work to GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.8, th The piece frames the product positioning as potentially misleading for builders, since the complexity of a multi-agent system is hidden behind one OpenAI-compatible API endpoint. It emphasizes that Fugu handles model selection, delegation, verification, and synthesis internally, so callers cannot directly manage which

DataNorth reports that Sakana AI released Fugu and Fugu Ultra on June 22, 2026, delivered through a single API that routes each request across a pool of frontier language models rather than running as one fixed model. Fugu is the lower-latency default and Fugu Ultra is the quality-first variant built for harder multi-s On Sakana's own testing, the article lists Fugu Ultra scores of 95.1 on GPQA Diamond, 93.2 on LiveCodeBench v6, and 54.2 on SWE-bench Pro, with Sakana reporting these place Fugu Ultra above Anthropic Opus 4.6, Gemini 3.1 high, and GPT-5.4 high on all three benchmarks. The design draws on two peer-reviewed papers accept

A Coursiv reference page describes Fugu Ultra as Sakana AI's quality-first version of Sakana Fugu, the multi-agent model orchestration system exposed through one OpenAI-compatible API. Standard Fugu is positioned as the lower-latency default, while Fugu Ultra is the variant Sakana positions for harder multi-step tasks The article notes that the pricing card on Sakana's product page covers both fugu-ultra-v1.1 and fugu-ultra-v1.0 (previously fugu-ultra-20260615). It frames Fugu Ultra as part of Sakana Fugu's learned orchestration system that can coordinate a pool of expert AI agents, with Fugu handling model selection, routing, coord

Verdent's developer guide examines how to read Fugu Ultra's coding numbers correctly, given that it is an orchestration system wearing a single model's API. The guide notes that Fugu Ultra launched with a SWE-Bench Pro score ahead of Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, but flags that all benchmark figures are Sakana The guide grounds Fugu's architecture in two 2026 Sakana research papers: TRINITY, which assigns Thinker, Worker, and Verifier roles to delegate work, and a learned-orchestration approach called the Conductor. Fugu is described as itself an LLM trained to call other LLMs in an agent pool, including recursively instance

Sakana AI launched Sakana Fugu on June 22, 2026 as a multi-agent orchestration system delivered through a single OpenAI-compatible API. According to MarkTechPost's coverage, Fugu is itself a language model trained to call other LLMs in an agent pool, managing model selection, delegation, verification, and synthesis int The coverage positions the Fugu launch as a shift from fixed-model APIs toward learned coordination, where Fugu can solve tasks directly or assemble expert-model teams as needed. It frames the complexity of a multi-agent system as abstracted from the caller's code, presenting a single endpoint that hides orchestration

The BenchLM aggregator page marks Sakana Fugu-Ultra as superseded, noting that Sakana AI has newer models in this line, specifically Sakana Fugu-Ultra v1.1. The page lists 11 published benchmark rows that are visible but notes that no category has enough eligible evidence for a comparative rank, with the model marked N The page reports a 1M token context window for the model and notes that no comparable first-party API token rate is published, meaning Sakana has not provided a first-party hosted token rate for direct comparison. Speed and time-to-first-token metrics are also not measured. The supersession marker and absence of rankin