O Sakana Fugu Bate o Opus? Veja tudo que essa IA é capaz de fazer

⭐ Sign up for the Waiting List for the Stop Fighting with AI Course: https://app.horadecodar.com.br/lp/par... 🔴 Vibe Coding Training (Antigravity, Claude Code and more): https://app.horadecodar.com.br/lp/for... 🟪 Hosting I recommend: https://www.hostinger.com/matheushora... (use coupon HORADECODAR for an extra 10% discount) 📘 Engineering Prompt Guide: https://app.horadecodar.com.br/ebookp... Join our Discord server and follow me on social media: 🟣 Hora de Codar Discord:   / discord   🔴 Instagram:   / horadecodar   Sakana Fugu is one of the most unique releases of the year: it's not a new model, it's an orchestrator that commands several cutting-edge models simultaneously. Instead of trying to be the mastermind, it acts as a conductor calling Opus, GPT, Gemini, and other frontier models behind a single API, deciding on its own which model to use for each part of a task. Sakana's thesis is simple: instead of relying on just one model, you use the best of each, coordinated by a system that has learned to assemble this team. In this video, I pit Fugu Ultra against Opus 4.8 in three real-world tests to see if this idea holds up in practice. If you follow AI models and like to see honest tests, without repeating the marketing table, this video is for you. Sakana Fugu comes from Sakana AI, a Japanese laboratory in Tokyo founded, among others, by one of the authors of the paper "Attention Is All You Need," which practically founded the current era of AI. It's not a weekend wrapper; it has serious research behind it (the Trinity and Conductor papers, on learned orchestration). The central idea of ​​Fugu is orchestration. It is, in itself, a small language model (about 7 billion parameters) trained for a specific function: calling other larger models, delegating parts of the problem, verifying the results, and combining everything into a single response. For you, this appears as a single API compatible with the OpenAI format: you send the request to Fugu, and it decides internally whether to solve it on its own or assemble a team of expert models. The complexity of the multi-agent never reaches its code. It comes in two versions: the normal Fugu, more flexible and with provider opt-out, and Fugu Ultra, tuned for maximum quality in difficult and multi-step problems, with a deeper and fixed pool. It's the Ultra that appears in the top benchmark numbers, and it's the one I use in the video tests. Regarding the benchmarks, and here comes the honest part: Sakana claims that Fugu Ultra leads in most code and reasoning benchmarks, especially SWE-Bench Pro, where it comes out ahead of Opus. But there are three caveats I want to point out. First, the competitors' numbers are self-reported by the providers themselves, under different test conditions, so it's a kind of apple-to-orange comparison. Second, Fugu's own numbers vary depending on the source. And third, "orchestrating beats a single model" is a claim that needs to be tested in practice, not just in the table. That's why I didn't just stick to the numbers: I ran three real tasks. There's a technical point that almost nobody mentions and that I think is fundamental: the hidden cost. Since Fugu calls several models underneath, its calculation isn't just what you see as input and output. There are orchestration tokens, expenses incurred in the internal coordination between models, which are included in the bill even if they don't appear in the final text. In heavy responses, these tokens can be much larger than the visible output. In other words, orchestration has a price that isn't printed, and whoever uses it needs to know this. Fugu's strategic positioning is also interesting. Sakana sells it as a response to dependence on a single vendor and the risk of export controls, at a time when cutting-edge models can be blocked in certain regions: by orchestrating models from multiple providers, if one is restricted, the system routes around it. It's a real argument about concentration risk. But an honest counter-criticism is in order: Fugu still depends on the same providers it orchestrates, so it rents the intelligence of these models instead of replacing them. It reduces lock-in, it doesn't eliminate it. Regarding price and access: Fugu comes in plans of US$20, US$100, and US$200 per month, with both models included, or in pay-as-you-go per token. At launch, it is not available in the European Union and the EEA due to GDPR issues. LINKS Sakana Fugu (official website): https://sakana.ai/fugu/ TIMESTAMPS 00:00 Sakana Fugu, Japan's "new AI" 00:45 Understanding how Sakana Fugu and Fugu Ultra work 02:45 How to test Sakana Fugu on the Codex 04:48 Creating a Landing Page with Sakana Fugu 08:35 Creating a full-stack system with Sakana Fugu 16:55 Token consumption and other data from Sakana Fugu 18:55 Final thoughts on Sakana Fugu