Japan Didn't Beat Mythos. It Built the Spare Key.

Tokyo's Sakana AI didn't train a model that beat Claude Mythos — it shipped a 7-billion-parameter conductor that routes around Washington's export controls and claims vendor-reported benchmark parity ten days after Commerce killed Fable 5 and Mythos 5.

What Fugu actually is:

• Not a foundation model. Sakana's Fugu is a learned orchestrator — a small language model that picks, delegates, and synthesizes across a swappable pool of frontier workers (GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro, open-source models).
• Mythos and Fable aren't in the pool. Anthropic's export-restricted models are unreachable by design — Fugu compares against their published scores, not live head-to-head runs.
• Two tiers: Fugu routes one worker per query for low latency; Fugu Ultra's Conductor writes multi-step workflows averaging three agents deep.
• Launched June 22, 2026 — ten days after Commerce ordered Anthropic to disable Fable 5 and Mythos 5 globally on June 12.

The benchmark claims — and the caveats:

Sakana's technical report (arXiv 2606.21228) says Fugu Ultra matches or edges Mythos Preview on select tests:

• GPQA Diamond (graduate science): 95.5 vs Mythos Preview's 94.6
• Terminal Bench 2.1 (agentic coding): 82.1 vs 80.4
• LiveCodeBench: Fugu Ultra 93.2, base Fugu 92.9, Fable 5 89.8

But Mythos still leads where it matters for engineering:

• SWE-Bench Pro: Mythos 80.3% vs Fugu Ultra 73.7%
• Humanity's Last Exam: Mythos 64.7% (with tools) vs Fugu Ultra 50.0% (text-only)
• ExploitBench (cybersecurity): Mythos 78.0% — Fugu not reported

Sakana's own report notes: "All scores other than Fugu's are reported by the model providers." Business Standard's July 1 analysis called the comparison apples-to-oranges — orchestration layer vs. monolithic frontier model.

Why the timing is the real story:

• CEO David Ha, co-author of Google's "Attention Is All You Need" paper and former Stability AI research head, launched Fugu as a hedge against single-vendor risk.
• Ha on X: "Access to top models can disappear overnight. Collective intelligence is the practical hedge against this concentration of power. Fugu simply routes around vendor restrictions by relying on an entirely swappable agent pool."
• A committee aide told CNBC July 8 that Congress is now examining whether America has "a sufficient open-weight AI strategy" — the same week OpenRouter data showed Chinese models holding 46% of U.S. enterprise tokens.
• Sakana reached a $2.65 billion valuation after a $135M Series B in 2025. Fugu subscriptions start at $20/month; Fugu Ultra pay-as-you-go runs $5/$30 per million tokens — roughly matching GPT-5.5 pricing while billing a single blended rate no matter how many models get called.

The reception:

• VentureBeat's side-by-side game-build test: Fugu Ultra finished a Crossy Road clone in 22 minutes for $7.32; Opus 4.8 took 79 minutes for $37.85 — but Opus produced better quality.
• Prime Intellect engineer Elie Bakouch: "This is a closed source orchestrator on top of closed source models... this is not 'AI sovereignty.'"
• Sakana's earlier AI Scientist and CUDA Engineer projects faced evaluation criticism — researchers are treating Fugu's scores with similar caution.
• Fugu is not available in the EU/EEA at launch while Sakana works toward GDPR compliance.

Convina's view: The Medium headline that Japan "beat Mythos" is vendor-reported scoreboard theater — Fugu never ran against Mythos in the same pool because Washington made Mythos unreachable. What Sakana actually proved is more interesting and more honest: when you kill the frontier models, the market builds conductors. A 7B orchestrator that rent-scores Mythos on GPQA is not a national AI breakthrough — it's the spare key enterprises will reach for when the guest list runs out. The question isn't whether Fugu beat Anthropic. It's whether routing around export controls counts as sovereignty or just a more expensive API wrapper with better marketing.