Sakana Fugu Ultra Benchmarks: A Practical Deep Dive into the Numbers (2026)

You’ve seen the headlines: Fugu Ultra beats Claude Opus 4.8 and GPT‑5.5 on engineering benchmarks. But all those numbers come from Sakana itself — no independent replication yet.

Fugu Ultra is a learned orchestration system that coordinates a pool of frontier models behind one OpenAI‑compatible API. It has no own monolithic LLM — the product is the orchestrator. Two variants exist:

Announced June 22, 2026, it is a multi‑agent system marketed “as a model.”

Flat illustration of a pipeline of AI agents (colored cubes) passing data through a central coordinator, with

These prices are for the API. SaaS subscriptions ($20/$100/$200 per month) are also available as wrappers.

All scores below are self‑reported by Sakana as of June 2026. No independent lab has replicated them yet.

Fugu Ultra solves 73.7% of real‑world code‑fix tasks, outperforming every single‑model competitor. However, Anthropic’s Fable 5 (unavailable due to export restrictions) scores ~80.0%.

Fugu Ultra scores ~93.2, ahead of Claude Fable 5 (89.8) and GPT‑5.4, Gemini 3.1 Pro.

Accuracy: 95.5% — slightly above Mythos Preview (94.6%) and well above older models.

Fugu Ultra: 82.1, Fugu: 80.2, Claude Opus 4.8: 74.6, GPT‑5.5: 78.2. The orchestrator shines in complex shell environments.

Fugu Ultra: 50.0, Claude Opus 4.8: 49.8. Essentially tied, but Sakana uses this to argue “at least not worse.”

Some reviews report: MMLU 87.3% vs Fable 5’s 85.1%; HumanEval 92.4% vs 89.7%. These differ slightly from other sources — always check the exact benchmark variant.

Every major number here is from Sakana’s own table. Until third parties confirm, treat them as strong marketing claims, not definitive rankings.

Best practice: Always write “according to Sakana’s self‑reported benchmarks” and avoid absolute statements like “Fugu Ultra is the best.”

Confusion easily arises:

Always include the exact name and version in comparisons.

Fugu Ultra dominates coding and tool‑use (SWE‑Bench Pro, LiveCodeBench, TerminalBench). On general knowledge (HLE, MMLU) it is roughly on par with top models, not dramatically ahead. Pick the right benchmark for your use case.

High scores do not come from a single super‑model. They come from clever orchestration of multiple models + tools. That means latency and cost can be higher than a single LLM. For production, evaluate total cost of ownership (TCO).

According to Sakana’s self‑reported benchmarks, yes on SWE‑Bench Pro (73.7% vs 69.2%) and TerminalBench 2.1 (82.1 vs 74.6). But these results need independent replication. On general‑knowledge tests like HLE, the gap is negligible.

Only if you need top‑tier quality for complex tasks and can tolerate higher latency and cost. For daily use, the standard Fugu tier ($5/$30 per million tokens) might be more practical.

Some sources report 78.9% on “SWE‑Bench” (without “Pro”). That’s a different, likely easier dataset. Always check the exact benchmark name and version.

Fugu Ultra’s benchmarks are impressive — but numbers alone don’t guarantee results in your workflow. The best way to evaluate any AI system is to test it on your own tasks. That’s why we built MakeBox AI: to give automation engineers a transparent, practical way to compare models and orchestrate workflows.