What Sakana AI's Own Launch Post Actually Says
Published: September 12, 2026 | Category: AI | By Mahesh | Source: Sakana AI's official launch post, primary data current as of September 12, 2026
Today's new product comes from Tokyo-based Sakana AI, founded by Llion Jones, a co-author of the original Transformer paper, and David Ha. On September 11, 2026, the company published a launch post introducing Fugu Max and Fugu Ultra v2, describing them as "the same core orchestration architecture optimized for two distinct missions."[1] The interesting part is not the benchmark scores, which every lab claims. It is the underlying premise: neither product is a single trained model in the way Gemini or Claude is. Both are, in Sakana's own words, a way of routing your request to other people's models and stitching the answer back together, sold to you as one API endpoint.
What Fugu Actually Is
Claim: Fugu is not a foundation model at all, but a coordination layer that decides which model handles which part of a task. Source: Sakana's own launch post frames the entire release around what it calls the Pareto frontier, stating "the frontier that actually matters to real-world tasks is two-dimensional: capability on one axis, cost on the other," and that "a system that deploys a multi-trillion-parameter model to execute a simple data lookup is not intelligent, but wasteful."[1] The company's own product page and prior technical report describe Fugu as, in the words of an independent technical summary quoting Sakana directly, "a Multi-Agent System, Delivered as One Model" that works "by dynamically coordinating and orchestrating a diverse pool of powerful models" behind a single OpenAI-compatible endpoint.[2] Analysis: This is a genuinely different bet than what every other major AI lab is making this year. Depth Grid covered how Google, Meta and OpenAI are each racing to ship bigger, more capable single models in Google Just Shipped Its Third Flash Model in Six Weeks; Sakana's argument is that the race itself is the wrong shape, and that intelligently routing between many smaller, cheaper, swappable models will eventually beat any single giant one on cost, and potentially on capability too. Published: September 11, 2026.
Two Products, One Architecture
Claim: Fugu Max and Fugu Ultra v2 share identical underlying technology but are tuned for opposite priorities. Source: Sakana's post states plainly: "Fugu Max asks: What is the best possible output we can deliver at the lowest possible cost? Fugu Ultra v2 asks: What is the absolute highest capability we can achieve on complex, multi-step tasks?"[1] On Fugu Max specifically, the post claims "performance within striking distance of elite models at two to six times lower cost," achieving "best overall score on six benchmarks, including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish," the last of which the company describes as its own internal coding benchmark.[1] Fugu Max draws on an expanded model pool that Sakana says now includes NVIDIA's Nemotron family, following what the company describes as an August 2026 partnership with NVIDIA.[1] On pricing, Fugu Max costs $2 per million input tokens and $6 per million output tokens, which Sakana states is "40-60% lower than Sonnet 5, GPT 5.6 Terra, and Kimi K3."[1] Analysis: Whether a genuinely cheaper option matters to a given team depends entirely on workload. For simple, high-volume tasks, an orchestrator smart enough to route to a cheap capable model rather than a large expensive one is a real cost lever; for anything requiring deep, sustained reasoning in a single context, the routing overhead itself, however small, is an added layer of complexity a direct model call does not have. Published: September 11, 2026.
The Claim That Actually Matters: Independence
Claim: Fugu Ultra v2 hits its top benchmark scores without using any frontier models from Anthropic or OpenAI in its underlying agent pool, a claim Sakana presents as a structural advantage rather than a limitation. Source: Sakana's post states outright: "Crucially, Fugu Ultra v2 achieves these scores without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool," and goes further: "Fugu Ultra v2 does not rely on individual proprietary frontier models to deliver frontier output. By orchestrating a swappable pool of open and specialized models, it outperforms closed ecosystems while protecting users from vendor lock-in, API revocations, geopolitical turbulence, and sudden service cutoffs."[1] Specific scores cited include 48.3 on Chartography, a visual reasoning and data interpretation benchmark, against Opus 5's 27.3 and Fable 5's 29.5, and 74.3 on DeepSWE, a real-world software engineering benchmark, which the post says beats "models that cost three to five times more per token."[1] Analysis: This is the actual news, and it is worth reading past the benchmark table to see it. Every one of these figures is vendor-published, not independently reproduced, and should be read with the same caution any company's own benchmark claims deserve. But the strategic point stands regardless of whether the exact scores hold up under third-party testing: Sakana is explicitly selling resilience against a single company's export controls, pricing changes or an outage taking down your product, by building an architecture that was never dependent on any one lab's model in the first place. That is a meaningfully different value proposition than "our model is smarter," and it is aimed squarely at enterprise buyers who have been burned by, or are simply nervous about, depending on a single AI vendor. Published: September 11, 2026.
Reading the Benchmark Claims Carefully
A useful discipline with any vendor's launch-day benchmark chart is separating what is independently verifiable from what is not. Sakana's own post is more transparent than most on this front, explicitly dating Fugu Ultra v2's training cutoff to August 28, 2026 and stating directly in a chart footnote that "Fable 5, Fable 5.1 and GPT-6-Astra are NOT in Fugu-Ultra v2's model pool," a specific, falsifiable claim rather than a vague marketing line.[1] An independent technical review of the release similarly notes that "all pricing and benchmark comparisons... come from Sakana's release announcement and remain vendor-reported until independently reproduced," which is the correct caveat to carry into any read of this launch.[3] What is separately verifiable, and matters regardless of the exact benchmark numbers, is that both models are live today through a standard API, at published prices, which Sakana's own post confirms by pointing users directly to its product page and console.[1]
How to Actually Try It
Sakana's own post lists the access paths directly: both Fugu Max and Fugu Ultra v2 are "available today via our standard OpenAI-compatible API," and existing Fugu users can upgrade with "a single-line parameter change," with "no migration" required.[1] To get started, Sakana points users to its Fugu product page or its developer console, and the company has also published its full underlying technical report on arXiv for anyone who wants to read the research behind the orchestration architecture rather than take the launch post's word for it.[1]
Common Questions
Sources
- Sakana AI, "Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier," official launch post, September 11, 2026. sakana.ai
- AI/TLDR, "Fugu Ultra v2: Sakana AI's Orchestration Model," citing Sakana AI's technical report and product documentation, September 11, 2026. ai-tldr.dev
- AlphaSignal, "Sakana AI Splits Fugu Into Max and Ultra v2 to Cut Costs 60%," September 11, 2026. alphasignal.ai
Read More on Depth Grid
- Google Just Shipped Its Third Flash Model in Six Weeks. Here's What's Actually New.
- Meta's New AI Agent Wants to Book Your Travel and Bills
- Why the AI Moat Is Shifting to the Workflow
- The Real AI Chip Bottleneck: CoWoS Packaging
- What Big Tech's Own Filings Say About the AI Spending Bet
Article by Mahesh | Depth Grid