Fugu Ultra

Sakana AI shipped Fugu-Ultra v1.1 on July 24, 2026, about a month after the original Fugu and Fugu Ultra launch on June 22, 2026, per a third-party recap citing Sakana's official X announcement and product page. The v1.1 refresh is described as a drop-in update that swaps in newer frontier models inside Fugu's agent po The piece catalogs version-specific model handles: fugu-ultra-v1.1 (new) and fugu-ultra-v1.0, renamed from fugu-ultra-20260615, alongside siblings Fugu, Fugu Ultra, and Fugu Cyber listed on sakana.ai/fugu. Distribution surfaces named are an OpenAI-compatible API plus listings on OpenRouter, Vercel AI Gateway, and model

TechTimes reported on July 16, 2026 that NVIDIA confirmed Sakana AI is integrating the Nemotron family of open-weight models into Fugu, adding a US-origin specialist for coding, tool calling, and instruction following to the orchestration pool. The Nemotron addition is explicitly additive rather than replacing existing The integration is roadmap-stage: TechTimes notes an "upcoming release with no specific date announced," so it is not yet shipped as of the article date, and the article does not mention LLM Gateway or any specific gateway distributor. TechTimes is a general tech outlet rather than a primary source, the page appears pa

On June 22, 2026, Sakana AI announced two tiers of its new product Sakana Fugu: Fugu, optimized for interactive use and lower latency, and Fugu Ultra, a quality-first variant that draws on a larger expert-agent pool. Fugu does not primarily try to answer every prompt itself; instead it decides which underlying models s Sakana's launch positioning claims that Fugu Ultra "stands alongside" systems such as Fable 5 and Mythos Preview on high-end coding, reasoning, and scientific benchmarks. The article explicitly flags that this is a claim from Sakana's own release rather than an independent benchmark review. The piece frames Fugu as a b

A third-party technical critique argues that Fugu Ultra is not a new frontier foundation model but a multi-agent orchestrator, and that its headline benchmark wins are achieved by routing work to closed frontier models. Sakana's June 22, 2026 product page describes Fugu Ultra as standing shoulder-to-shoulder with Anthr The Fugu technical report (arXiv:2606.21228) and product page show that the orchestrator's agent pool includes GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.8, which the author argues is why Fugu Ultra outperforms any single-model baseline on Sakana's chosen benchmarks. Both Fugu and Fugu Ultra are delivered through a sin

Sakana AI's Fugu orchestrates a pool of frontier LLMs behind one API. We cover the features, benchmark numbers, pricing, and real-world use cases.

Fugu Ultra explained: compare Fugu vs Fugu Ultra, pricing, benchmarks, API access, use cases, limitations, and when teams should evaluate it.

The Decoder's June 23 coverage of Sakana Fugu, updated June 25 with first hands-on impressions, is the strongest independent counterweight to the launch benchmark claims and directly relevant to developers reaching Fugu Ultra through any gateway including LLM Gateway. Named testers (Ethan Mollick, @LLMJunky, Hamel Husa For LLM Gateway's fugu-ultra listing, this matters because the developer-experience signals suggest Fugu Ultra is currently early-access quality rather than production-ready: latency, quota caps, and jagged abilities across task types are concrete operational risks. The piece also flags a sovereignty caveat — Fugu's po

A Verdent developer guide focuses on how to read Fugu Ultra's coding benchmark numbers, particularly a SWE-Bench Pro score positioned ahead of Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. It stresses that Fugu Ultra is architecturally unusual, an orchestration system wearing a single model's API, which changes how the scores The guide grounds Fugu 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, both used to train Fugu itself as an LLM that calls other LLMs in an agent pool (including, recursively, instances of itself)