Sakana AI Fugu Max & Fugu Ultra v2

A dated reading of what is claimed, reported, and independently verified in the current evidence.

Current reading

Readiness band and full integer triple

AI-assisted assembly · derived results

Readiness band

Watch

Primary summary from verified readiness

Evidence strength Strong

Decision

Human editorial judgment · 2026-09-22

Proceed with caution

Why: The core capability is real and well-documented (learned model-routing orchestration at $2/$6 per M tokens spanning open-weight, specialized, and NVIDIA Nemotron models, with a fast April-to-September 2026 release cadence and a genuine enterprise distribution deal), but every benchmark and cost-savings number currently traces back to Sakana's own blog; independent tech-press analysis flags the absence of third-party verification and mixed early real-world tester feedback, and the vendor has a 2025 precedent of retracting a headline performance claim after external users caught reward-hacked benchmarks.
Next: Track through the next refresh cycle; do not cite Fugu's cost or benchmark numbers in pv content as independently validated until a non-Sakana-sourced reproduction appears; watch the Sumitomo 100k-client distribution deal as the real adoption signal to confirm or disconfirm the cost-arbitrage pitch.

Constraints

No named blocker is present in the current public projection.

Evidence summary

AI-assisted assembly

Counts and dates only. Raw signals, private excerpts, trust records, and internal corpus material are not published here.

Publication record

Initial public reading