In the single-model era, one API call usually implied one relatively legible execution path. You knew which model ran. You had a rough mental model of cost. You could reason about latency and usage from the outside. In the orchestration era, none of that holds. A case study 👇 ---- @SakanaAILabs launched Fugu this week. It looks like one model. You call it through a single API. But what you call and what actually runs are two different things. Fugu is a coordinator trained to break a task apart, route pieces of it to a pool of other models, and call instances of itself recursively to handle sub-tasks. One prompt can fan out into a swarm of sub-calls the user never sees. The whole thing still surfaces as one model response. Two developers ran into the consequences of that almost immediately. @LLMJunky said a single prompt consumed 100% of a five-hour quota. @cortesi paid for Sakana's $200 tier, found the API slow, and said he hit his limit in under an hour. Neither could say which models ran, how many times Fugu called itself, or why one prompt cost what it did. The issue isn't specific to Fugu either. When @AnthropicAI briefly opened access to Fable earlier this month, users reported similar quota burns with similarly little visibility into why. When one request fans out into many model calls, and some of those calls spawn more, the unit the buyer sees no longer maps cleanly to the unit of work happening underneath. Forecasting cost gets harder. So does explaining latency, attributing performance, or knowing what actually ran. Sakana's benchmark figures sharpen the question. Shared numbers for Fugu include 54.2 on SWE-Pro, 95.1 on GPQA-D, and 93.2 on LiveCodeBench v6, with testers noting scores above Opus, Gemini 3.1, and GPT 5.4 on each. But Sakana's own base model is around 7B parameters. Frontier-tier results from a coordinator that size strongly suggest that a meaningful share of the work is being done by the larger third-party models it calls. Behind one endpoint, the system may be making its own decisions about how much work to do, which models to invoke, and how many times to invoke them. For hobby use, opaque execution is an annoyance. For production systems touching real money or real workflows, it becomes harder to dismiss. You can't price what you can't predict, and you can't debug what you can't trace. When the run is a sealed box, it also becomes harder to prove what the agent actually did. Fugu is one early launch, and Sakana will likely smooth out the rough edges. But as orchestration becomes the product, execution visibility becomes core infrastructure. The next layer of competition in AI will be won by the teams that can make increasingly complex agentic runs legible enough to trust. Follow @SaharaAI for analysis of agent infrastructure, the AI economy, and where the real building is happening.

🐡 The era of agonizing over which single LLM is "the best" might be coming to an end. Title: Sakana Fugu Technical Report URL: Sakana AI's Fugu is a commander (orchestrator) model that marshals a team of frontier LLMs and automatically assembles the best lineup for each task. Here are the three highlights worth knowing. 🧭 A learned conductor Fugu adds a lightweight selection head on top of a pretrained backbone and picks the best worker model directly from logits over hidden states. Because it skips autoregressive decoding, it keeps latency on par with calling a single frontier model — an elegant trick. 🕸️ Dynamically writing its own workflows (Fugu-Ultra) The high-end variant writes out full agentic workflows in natural language, defining task decomposition, subtask allocation, and targeted communication strategies. It composes multi-agent topologies — debate, sequential chaining, tree structures — of up to 5 steps, tailored per problem. 📊 Beating any single model 73.7 on SWE-Bench Pro (above Claude Opus 4.8's 69.2), 95.5 on GPQA-Diamond, 93.2 on LiveCodeBench — state-of-the-art among publicly accessible models. In agentic coding it lands a 5–6% relative gain, described as a full "generational" improvement. Reaching the frontier through smart orchestration rather than sheer scale — a fresh new scaling axis. #SakanaAI# #LLM#

TechCrunch on Sakana Fugu and the broader conversation around frontier AI access in Asia. Our position remains that AI is best developed together rather than hoarded. Sakana AI will continue contributing to a more resilient ecosystem.

Ren Ito, Co-Founder and Chairman of Sakana AI, has been appointed to the "AI for Good Global Commission", established by the United Nations and the ITU. As a Japanese AI company, we are honored to join global leaders, experts, and policymakers to actively contribute to building a trusted AI ecosystem, driving responsible innovation, and shaping the future of AI policy. Sakana AI 共同創業者の伊藤錬が、国連AI for Goodのグローバル委員会委員に就任しました。 日本のAI開発企業として本委員会へ参画し、信頼できるAIエコシステムの形成や責任あるイノベーションの支援に向けた国際的な議論に積極的に貢献していきます。

🐡 One day a regulation changes and the model you depend on loses its API overnight—that reality lays bare the risk of single-vendor dependence in AI. Export controls on Fable and Mythos showed access can be cut in an instant. So rather than betting everything on one giant model, maybe the resilient blueprint for AI sovereignty is collective intelligence: many models, orchestrated to collaborate. That is the question Sakana AI poses. Its answer is Sakana Fugu: One Model to Command Them All. Fugu isn't a mere router—it's a language model trained to call various LLMs in an agent pool. You send a request to one endpoint, and Fugu decides whether to solve it directly or assemble a team of specialized models, handling selection, delegation, verification, and synthesis internally. It calls itself recursively, and because agents in the pool are swappable, it can dynamically route around restricted models—the crux of "sovereignty." It builds on Trinity and Conductor (ICLR 2026), running on learned orchestration rather than fixed workflows. 🚀 The accuracy-focused Fugu Ultra stands shoulder-to-shoulder with Fable 5 and Mythos Preview on rigorous reasoning, science, and engineering benchmarks, and on some tasks beats Gemini 3.1 Pro, Opus 4.8, and GPT 5.5. Among 500 beta users, a code reviewer found more than twenty issues where competitors flagged about three, and data-science research progressed with little human intervention. From monolithic scaling toward a collaborative ecosystem. URL: #AIAgents# #LLM#

DAILY SITUATION RECAP: Nvidia launches the Open Secure AI Alliance in order to find and fix vulnerabilities using open-source AI, sort of like an open Project Glasswing. Founding partners include a mix of enterprise software companies (Databricks, Salesforce, IBM, SAP, Siemens, Snowflake), cybersecurity companies (Palo Alto Networks, Red Hat), open-source providers (Hugging Face, LangChain, OpenClaw, Nous, the Linux Foundation), AI labs (SpaceXAI, Thinking Machines, Cognition), and other major companies (Nvidia, Microsoft, Cisco, Palantir, Dell). Moonshot AI releases the Kimi K3 weights and technical report after eleven days since launch. Kimi K3 is a 2.8T parameter mixture-of-experts (MoE) model with 104B active parameters and a 1M token context window. Moonshot also open-sourced much of their infrastructure, including their attention kernels, agent environment platform, and MoE communication library. Just because you can download it in theory doesn’t mean you actually can — the model is far too big to be run on any consumer hardware. Nvidia invests $5B in Ilya Sutskever’s SSI. Sutskever, formerly co-founder and Chief Scientist of OpenAI, founded Safe Superintelligence in 2024 with the sole goal of building a safe superintelligence, with no other products along the way. It has since raised $3B at up to a $32B valuation (likely higher now). SSI is famously very secretive about its research, but Sutskever said it’s “focused on overlooked aspects of how the human brain functions”. The new funding, and access to Nvidia Vera Rubin GPUs, will allow SSI to 10x its compute. More companies sign on to Nvidia’s open source letter. The letter, posted by Jensen Huang on Friday, advocates for a robust American open-source ecosystem with minimal government regulation. New signatories include Google, SpaceXAI, OpenAI, AMD, Cisco, Palo Alto Networks, Nebius, Scale, Fireworks AI, Baseten, Cohere, Sakana AI, Periodic Labs, Core Automation, OpenClaw, and GitHub. Every major American frontier lab except for Anthropic has now signed. CXMT stock surges 466% on its first trading day. The company, formerly ChangXin Memory Technologies, is the largest memory manufacturer in China and the fourth-largest in the world (after SK Hynix, Samsung, and Micron), with a 9% global market share. It now has the second-highest market cap of any Chinese company after Tencent. CXMT doesn’t make the most leading-edge HBM for AI chips, but supplies DRAM to consumer tech manufacturers and data centers. Nvidia may guarantee $250-350B of financing for an OpenAI data center. SB Energy, a subsidiary of SoftBank, is developing a massive 10 GW data center on federal land in Ohio at a total cost of over $500B. The financing guarantee would allow SB Energy to borrow money at lower rates, and possibly allow OpenAI to spend more on Nvidia chips. China begins manufacturing DUV machines. Deep ultraviolet (DUV) lithography machines print intricate nanoscale patterns on silicon wafers, a critical step in chipmaking. The new machines, built by an unnamed state-backed company, will be shipped to local chipmakers including SMIC, Hua Hong Semiconductor, and CXMT. China is still behind on the most advanced extreme ultraviolet (EUV) lithography, which is solely produced by Dutch company ASML. ASML stock fell 6% on the news. Dario Amodei explains Anthropic’s position on open models: open-weight models without dangerous capabilities are a public good, and Anthropic has never supported a full ban. However, we should be worried about the CCP using them for repression, as well as cyber/bio/alignment risk. To that end, we should not sell chips to China, crack down on distillation, and require mandatory safety testing for all sufficiently capable open and closed models. China threatens to respond if the US sanctions their AI labs. The Chinese Ministry of Commerce said US accusations of distillation were “smears” and that China will “take all necessary measures” to defend its rights and interests against any action that substantively harms them. DeepSeek has suspended its recent funding round after comments from a private investor call with CEO Liang Wenfeng were leaked. Written by @theojaffee. Read more at our link in bio.