Sakana Fugu Launches With 93.2 LiveCodeBench Score After Claude Ban

Sakana AI released Fugu and Fugu Ultra this week, turning its multi-agent orchestration research into a commercial API after Anthropic said a US government directive forced it to suspend access to Claude Fable 5 and Claude Mythos 5. Sakana said Fugu Ultra scored 93.2 on LiveCodeBench, ahead of Fable 5's 89.8, and matched or beat several top models on science and software benchmarks. The Tokyo company is selling the service as one endpoint that can route work across a pool of models instead of tying a customer to one provider.

Anthropic announced on June 12 that the US government, citing national security authorities, directed it to suspend access to Fable 5 and Mythos 5 by foreign nationals, including Anthropic employees. The company said it had to disable the models for customers to comply, while access to other Anthropic models would remain unaffected. Sakana co-founder and Chief Executive David Ha framed that disruption as the commercial opening for Fugu.

"Relying on a single company's model for national infrastructure is a massive risk," Ha wrote on X, according to VentureBeat. "Fugu simply routes around vendor restrictions by relying on an entirely swappable agent pool."

Sakana posts 93.2 on LiveCodeBench

Sakana describes Fugu as a language model trained to call other language models, including instances of itself, and to decide when to delegate, verify and combine work. The company said the system builds on two ICLR 2026 papers, TRINITY and Conductor, which trained coordinator models to assign tasks to specialized agents instead of using fixed workflows.

The benchmark claims are the center of the launch. Sakana said Fugu Ultra scored 73.7 on SWE-Bench Pro, ahead of Anthropic's Claude Opus 4.8 at 69.2 and OpenAI's GPT-5.5 at 58.6, while trailing the restricted Fable 5 score of 80.0. On GPQA-Diamond, Fugu and Fugu Ultra each scored 95.5, above the 94.6 Sakana attributed to Mythos Preview.

The scores come with a caveat Sakana itself notes. Fable 5 and Mythos Preview are not in Fugu's agent pool because they are not publicly accessible, and Sakana has not disclosed which models Fugu uses for a given request. The Verge's Richard Lawler noted that the pitch amounts to using other frontier models more carefully, while leaving customers without a view into which model performed which part of the work.

Fugu Ultra starts at $5 per million input tokens

Sakana is offering two versions. Fugu is aimed at coding, chat and lower-latency work; Fugu Ultra is aimed at harder tasks such as AI research, paper reproduction, cybersecurity assessment and patent search. Both are available through an OpenAI-compatible API.

Fugu Ultra starts at $5 per million input tokens and $30 per million output tokens, with cached input priced at 50 cents per million tokens, according to Sakana's documentation cited by VentureBeat. For contexts above 272,000 tokens, the Ultra rate rises to $10 per million input tokens, $45 per million output tokens and $1 per million cached input tokens.

The standard Fugu service uses variable pricing based on the highest-tier underlying model activated for a request, rather than stacking charges for every agent in the workflow. Sakana also says enterprise users can exclude specific providers or models from the routing pool and can opt out of prompt use for future training. The service is not available in the EU or EEA while the company works through data compliance issues.

Critics point to the black box

Sakana said nearly 500 beta users tested Fugu before launch, with reported uses in code review, security assessment, patent search and automated research. OfficeChai cited an AutoResearch run in which Fugu Ultra ran 123 training experiments over 14 hours on one H100 GPU.

Outside testers have reported mixed results. VentureBeat cited Mark Santos of Mark Studios, who compared Fugu Ultra and Claude Opus 4.8 on a Three.js game build. Santos said Fugu Ultra finished in 22 minutes, used about 89,000 tokens and cost roughly $7.32, while Opus took 79 minutes, used about 940,000 tokens and cost nearly $37.85. He still judged Opus better on final application design and function.

Elie Bakouch, a research engineer at Prime Intellect, objected to Sakana's sovereignty framing. "This is a closed source orchestrator on top of closed source models," he wrote on X, according to VentureBeat. "If before you didn't control the models, now you don't even control which ones are used or how much."

Sakana's next test is adoption rather than another benchmark. The company says it will add open models and its own models to the pool over time, and the commercial version is now generally available outside the restricted European regions. Customers weighing Fugu will have to decide whether one more black box is an acceptable price for less dependence on any single model provider.

AI Models Working in Concert Outperform Solo Systems by 30 Percent, Sakana Study Finds

Japanese researchers prove AI models work better as teams than alone, boosting performance 30%. TreeQuest lets companies mix different AI providers instead of relying on one.

The Implicator

OpenClaw Creator Says Anthropic Is Pushing Developers Off Claude Subscriptions

OpenClaw creator Peter Steinberger says Anthropic is pushing developers off Claude consumer subscriptions and toward API keys.

The Implicator

You are Managing Five Expensive AI Interns

Anthropic multi-agent Claude Code lets developers run several assistants at once. The token bills tell a different story about who benefits.

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai