OTHER NEW June 22, 2026

Sakana Fugu Launches — First Frontier Alternative to Fable 5

WHAT CHANGED

Tokyo-based Sakana AI launched Sakana Fugu and Fugu Ultra on June 22, 2026. Multi-agent orchestration system presenting as a single foundation model via OpenAI-compatible API. Fugu Ultra claims parity with Anthropic's Fable 5 and Mythos Preview on engineering, scientific, and reasoning benchmarks — while neither of those models is in Fugu's agent pool (export-controlled, inaccessible).

WHY IT MATTERS

First production frontier-level alternative explicitly positioned as 'no export control risk.' Released 10 days after Fable 5 suspension — timing intentional. Architecture novelty: tiny coordinator LLM (~7B) trained to delegate to a pool of frontier models including Gemini 3.1 Pro, GPT-5.5, Opus 4.8. Built on Sakana's ICLR 2026 papers (TRINITY + Conductor). Independent reviewers (Ethan Mollick) report 'sharp gap between benchmark claims and real use' — verify on your workloads.

HOW TO USE IT

Sign up at console.sakana.ai. OpenAI-compatible API — point your existing OpenAI client at Fugu's base URL with Sakana API key. Pricing: $20/mo Standard, $100/mo Pro (10x), $200/mo Max (20x). Pay-as-you-go available at API rates. Free second month for subscribers signing up before July 31, 2026. NOT available in EU/EEA pending GDPR review.

OTHER / PYTHON

from openai import OpenAI

client = OpenAI(
    base_url="https://<your-fugu-endpoint>/v1",
    api_key="YOUR_SAKANA_API_KEY",
)
resp = client.chat.completions.create(
    model="fugu-ultra-20260615",  # or "fugu" for cheaper
    messages=[{
        "role": "user",
        "content": "Reproduce the method in this paper and report the gap."
    }],
)
print(resp.choices[0].message.content)

sakanafuguorchestrationfrontieralternativemulti-agentjapan

ORIGINAL SOURCE

https://sakana.ai

VIEW ORIGINAL SOURCE →

Real-world reports from beta (~500 users): A software engineer found Fugu Ultra surfaced 20+ issues in code review vs the ~3 typical for frontier models. A cybersecurity engineer reported it kept scoped assessments within bounds while producing evidence and retest steps. An enterprise platform exec praised persona stability across long sessions.

Caveat: Fugu Ultra’s pool is fixed — no opt-out of individual models. The Fable 5/Mythos comparison is mixed: depends heavily on which benchmark, which Anthropic model version, and which evaluation configuration. The clearest claim is that Fugu Ultra “stands shoulder-to-shoulder” with Fable 5 on Sakana-reported benchmarks. Independent verification still pending.

Three things to track over the next 30 days: (1) whether the benchmark numbers survive independent replication on agentic and long-horizon tasks, (2) latency and cost overhead of chaining API calls across providers, (3) what happens if any of Fugu’s pool models becomes inaccessible.

← BACK TO UPDATES