Sakana Fugu: the AI that is not a model, it is a conductor

While the US government had Claude Fable 5 switched off, a Tokyo lab shipped a system that hits frontier scores by orchestrating other people's models behind one API. Here is what Fugu actually is, what the numbers really say, and when you should (and should not) use it.

01The story in 60 seconds

JUN 12The US government orders Anthropic to suspend Fable 5 and Mythos 5, citing export-control powers, three days after launch. Anthropic disables both models worldwide to comply.
JUN 22Sakana AI, a Tokyo R&D lab, launches Fugu into that exact vacuum: a single API that promises frontier performance with no dependence on any one provider or government.
JUL 1Fable 5 comes back online after the order is lifted, with a new vulnerability-detection classifier. But by then Fugu had made its point: the "one irreplaceable model" era has a challenger.

The pitch that landed: if any single model can be switched off overnight, the safest bet is a system that treats every frontier model as a replaceable part.

02What Fugu actually is

Fugu is not one big model. It is a small, learned coordinator sitting on top of a pool of frontier models. You send one request to one OpenAI-compatible endpoint. Behind the scenes, the coordinator decides which models to call, in what order, and in what roles:

T
ThinkerPlans the approach and breaks the task down.
W
WorkerExecutes the pieces, often several models in parallel.
V
VerifierChecks the output before it comes back to you as a single answer.

The coordinator itself is tiny: the research behind it (two ICLR 2026 papers, TRINITY and Conductor) uses models of roughly 0.6B and 7B parameters, trained with reinforcement learning to discover coordination strategies instead of following a hand-wired agent graph. If a model in the pool becomes unavailable, the coordinator routes around it. That is the resilience story.

Sound familiar? This is the same orchestrator-and-workers pattern we covered in The Fable Sandwich: one smart brain dispatching cheaper executors. Fugu productizes that pattern and sells it as a model.

03The benchmark numbers

Sakana's published scores, against the frontier models everyone can actually buy today:

Benchmark	Fugu	Fugu Ultra	Opus 4.8	Gemini 3.1 Pro	GPT 5.5
SWE Bench Pro	59.0	73.7	69.2	54.2	58.6
TerminalBench 2.1	80.2	82.1	74.6	70.3	78.2
LiveCodeBench	92.9	93.2	87.8	88.5	85.3
GPQA-Diamond	95.5	95.5	92.0	94.3	93.6
CharXiv Reasoning	85.1	86.6	84.2	83.3	84.1

Read the fine print. The Fable 5 comparisons in Sakana's launch material use provider-reported scores, not head-to-head runs (Fable 5 was offline and is excluded from the pool). The pool membership is officially proprietary; leadership has named Gemini, DeepSeek, Kimi, Opus and GPT as members, but the docs anonymize results as "Model A/B/C". Treat vendor benchmarks as a claim, not a fact.

04The speed story (and its cost)

The demo everyone shared: the same "build a battle royale game" prompt given to three systems. The heavyweight single model produced the most polished build in about 20 minutes. Fugu shipped a working game in about 6, ahead of Kimi. That is the real advantage: parallel workers finish bounded tasks faster than one giant model working alone.

Independent testing adds the other half of the picture:

1
Orchestration overhead is realLight queries have run as long as 108 seconds, with roughly 60% of tokens spent on back-channel coordination calls you never see.
2
Cheaper and faster, not always betterIn one developer benchmark, Fugu Ultra burned 89K tokens ($7.32) in 22 minutes where Claude Opus used 940K tokens ($37.85) in 79 minutes. Fugu won on speed and cost, Opus won on output quality.
3
You cannot audit itBy design you do not know which underlying model handled your request. For regulated data, client work, or anything needing provenance, that is a dealbreaker.

05Variants and pricing

Variant	What it is	Pay-as-you-go (per 1M tokens)
Fugu	Latency-balanced default for everyday coding and interactive work. You can opt out of specific providers.	Standard rate of the underlying model
Fugu Ultra	Deeper expert pool for hard multi-step problems. Pool is fixed.	$5 in / $30 out / $0.50 cached
Fugu Cyber	Security-specialized: vulnerability research and threat analysis.	$6 in / $36 out / $0.60 cached

Subscriptions: $20 / $100 / $200 per month (Standard, Pro at 10x, Max at 20x), both main models included. 1M-token context window; rates step up past 272K tokens. When multiple agents work your request you pay a single rate based on the top model involved, no fee stacking. Not yet available in the EU/EEA.

06When to use it (and when not to)

Reach for Fugu when	Stay with a single model when
Speed matters more than polish: prototypes, internal tools, throwaway builds.	Output quality is the product: client deliverables, production code you will maintain.
You want resilience against any one provider having an outage, price hike, or shutdown.	You need to know exactly which model touched your data (compliance, NDAs, regulated work).
The task decomposes into parallel bounded pieces (multi-file builds, research sweeps).	The task is one long coherent thread where context continuity beats parallelism.

Our take: the interesting thing about Fugu is not the benchmarks, it is the architecture. Orchestration as a product is where agentic AI is heading. You can get most of the benefit today inside Claude Code by running the orchestrator pattern yourself, with full visibility into which model does what.

07Try it in 5 minutes

1
Get a keySign up at console.sakana.ai (or use it through OpenRouter / Vercel AI Gateway with the account you already have).
2
Point your existing code at itThe API is OpenAI-compatible, so it is a two-line change in any OpenAI SDK project:

# same OpenAI SDK, different base URL
client = OpenAI(
    base_url="https://api.sakana.ai/v1",
    api_key=os.environ["SAKANA_API_KEY"],
)
resp = client.chat.completions.create(
    model="fugu",  # or "fugu-ultra"
    messages=[{"role": "user", "content": "..."}],
)

Start with plain fugu on a real task from your backlog. Watch the per-request token report (cost shows up per call) before you let it near anything big.

SANITY CHECK BEFORE YOU COMMIT

Run the same prompt through Fugu and through your current daily model. Compare output quality, wall-clock time, and the reported cost. Fugu earns its place on speed-sensitive parallel work; it has not earned blind trust on quality.

EasyHive · easyhive.pro · Sources: sakana.ai/fugu, Anthropic statement on the Fable 5 suspension, VentureBeat, independent developer benchmarks (July 2026)

All trademarks belong to their owners. Easy Hive is not affiliated with or endorsed by them. This site is not part of Facebook, Instagram or Meta Platforms, Inc., and is not endorsed by Meta. Savings and results shown are examples, not guarantees; yours will vary. Educational content only, not legal, tax or financial advice. Full disclaimers.