01What Saakana Fugu Ultra Is and How It Works

Saakana Fugu Ultra review: it is a multi-agent orchestration model from Saakana.ai that fuses answers from multiple closed and open models into one output. Instead of relying on a single AI, it combines several to get better results than any one model alone.

On benchmarks, Fugu Ultra claims Fable 5-level intelligence. It is made by a small Japanese AI lab. It uses a subscription plan, so you pay a flat rate and get tokens, similar to a Claude subscription.

In this episode, Julian Goldie tested it side by side with Fusion and Opus 4.8 using 42 real builds. He wanted to see if the benchmark claims matched real performance.

02Main Features of Fugu Ultra

  • Multi-agent fusion – combines multiple models to produce a single answer.
  • Subscription-based – flat plan with tokens.
  • Two versions – Fugu Ultra (high quality) and Fugu Mini (cheaper and faster).
  • One-shot generation – you get one response after 15–20 minutes; no back-and-forth.
  • Strong at creating games and interactive web apps.
  • Accessible via Saakana.ai or integrated into an agent operating system like the AI Profit Boardroom.

Compared to Fusion (pay-per-token) and Opus 4.8 (Claude's most powerful model), Fugu Ultra aims to be a step up. But beware: it may lock you out for five hours if you hit token limits.

03How to Set Up Saakana Fugu Ultra

  1. Go to Saakana.ai and sign up for a Fugu Ultra subscription.
  2. Choose a plan that includes enough tokens for your testing needs.
  3. Decide whether to use the standalone interface or integrate it into an agent system like the AI Profit Boardroom.
  4. If using an agent OS, plug Fugu into your existing setup (as shown in the episode).
  5. Start with simple prompts to get used to the one-shot workflow.
  6. Keep track of your token usage to avoid lockouts.

Julian noted that early users were limited by token counts, making the tool slow and causing frequent lockouts. Planning your prompts carefully can reduce frustration.

04How to Use Fugu Ultra with Real Builds

In the episode, Julian tested Fugu Ultra on 42 different builds using the same 42 prompts on three models. He showed examples like a solar system simulation, a voxel temple run, a fireworks display, a Nordic dungeon crawler, and an open-world game.

For each build, he compared Fugu Ultra, Fusion, and Opus 4.8 side by side. Fugu Ultra often produced the most polished results. For the solar system test, Julian preferred Fugu Ultra's version as the most sophisticated.

He also compared Fugu Ultra vs Fugu Mini using the same prompts. Fugu Ultra produced much better graphics and smoother gameplay than Fugu Mini.

One key use tip: since Fugu Ultra is one-shot, you cannot iterate on the same response. If you want changes, you may need to use another model like Opus 4.8 to tweak the code afterward.

05Benchmark Results and Comparisons vs Fable 5

Julian shared Goldie Bench leaderboard results and standard benchmarks. On Live Code Bench, Fugu Ultra scored 93.2, beating Fable 5's 89.8. It also topped GPQA Diamond at 95.5, above Sakana's own MIFOS preview.

However, Fable 5 outperformed Fugu Ultra on SWE Bench Pro and Terminal Bench. Overall, the gap is close but Fable 5 seems stronger on some coding tasks.

In real builds, Fugu Ultra consistently outperformed Opus 4.8, which often got stuck on start screens in the examples. Fusion sometimes matched but rarely beat Fugu Ultra. In the leaderboard, Fusion was top, followed by Fugu Ultra.

Julian concluded that Fugu Ultra is comparable to Fable 5 but not quite as good, and you cannot have an interactive conversation with it.

06Tips, Limits, and Mistakes to Avoid

  • Watch token usage – running out can lock you out for five hours.
  • Expect 15–20 minute response times – plan your workflow accordingly.
  • Fugu Ultra is one-shot – don't expect to refine the same output; use one prompt per task.
  • For highly iterative tasks, use a model like Opus 4.8 instead.
  • Fugu Mini is faster and cheaper but produces noticeably lower quality.
  • Some builds from other models (like Fusion's fireworks) simply didn't work – always test multiple models.

Avoid relying on Fugu Ultra for quick tests or back-and-forth conversations. Use it for one-off high-quality builds where long waits are acceptable.

07Who Saakana Fugu Ultra Is For and Why It Matters

Saakana Fugu Ultra is for developers and AI enthusiasts who want to generate complex, high-quality outputs from a single prompt. It is especially useful for creating games, simulations, and interactive apps without needing to code yourself.

It matters because it shows that multi-agent systems can rival leading models like Fable 5. By fusing multiple models, Fugu Ultra can produce better results than a single model, as seen in the builds.

Julian recommends it if you want a step up from Opus 4.8. However, if you need fast, iterative work, stick with another model. Fugu Ultra is best for one-shot, ambitious projects where you can wait and accept the output as final.

For the full setup, Julian offers the AI Profit Boardroom, which includes Fugu, Fusion, Claude, and Hermes, plus tutorials and community support.