Can a Fable‑Level AI Model That Evades Export Controls Beat Claude Mythos?

When Anthropic’s Claude Fable 5 was globally disabled, Sakana AI announced its own Fugu model, claiming performance on par with Claude Fable and Mythos while explicitly avoiding export‑control risks.

Benchmark Performance

In the most stringent engineering, scientific, and reasoning benchmark suites, Fugu ties with leading models such as Fable 5 and Mythos. On the SWE Bench Pro and Terminal Bench 2.1 tests, the Fugu‑Ultra variant achieves the current state‑of‑the‑art scores, improving over the runner‑up by 5%–6%, a gain comparable to a full major version upgrade of competing providers.

For scientific reasoning, Fugu surpasses both Mythos Preview and Fable 5, demonstrating a notable edge in complex problem solving.

Core Architectural Advantage

Fugu is not a single monolithic model; it is a multi‑agent system that maintains a freely switchable pool of AI agents. When a single vendor is restricted, the system automatically reroutes to alternative models, dramatically increasing system resilience.

The model uses a pretrained language model backbone that, via its hidden state, coordinates a pool of expert agents—including recursive calls to its own instances. A lightweight selection head runs in parallel with the base model head, receiving the hidden state ℎ and emitting logits for each working model, thereby reducing coordination overhead and latency.

Training follows a two‑stage process: first, large‑scale supervised fine‑tuning (SFT) on diverse single‑step tasks covering programming, mathematics, reasoning, language understanding, and various agent scenarios; second, evolutionary‑strategy optimization on end‑to‑end multi‑step tasks collected from real‑world coding assistants (Claude Code, Codex, OpenCode), incorporating repository context, iterative editing, tool use, execution feedback, and final task completion.

Model Scheduling Over Scaling

The authors argue that intelligent scheduling should become a new performance dimension, reducing reliance on raw compute scaling. Fugu’s adaptive routing demonstrates sustained, diverse adaptability across evaluation tasks, learning to invoke specialized models based on their unique skills.

Comparisons against three frontier baselines—Gemini 3.1 Pro (high), Opus 4.8 (max), and GPT 5.5 (xhigh, anonymized as models A, B, C)—show that Fugu’s orchestrated approach yields comparable or superior results without increasing model size.

Implications

The release signals a shift from “large model domination” to “model orchestration supremacy,” where the ability to dynamically schedule and combine multiple expert models becomes the primary competitive edge.

References: Sakana AI’s GitHub repository (https://github.com/SakanaAI/fugu), release page (https://sakana.ai/fugu‑release/), and the two ICLR 2026 papers underlying Fugu’s design.