Sakana AI's Fugu Ultra v2 outperforms GPT-6 Astra in some tests

Summary

Sakana AI has released the Fugu Ultra v2 and Fugu Max models. The company claims that Ultra v2 outperforms GPT-6 Astra and Claude Fable 5.1 in selected coding benchmarks (DeepSWE). The Max model, on the other hand, reduces output costs by 40-60% compared to GPT-5.6 Terra, Claude Sonnet 5, and Kimi K3. Both models are served via an OpenAI-compatible API. Ultra v2 is priced at $5/$30 per million tokens, while Max is priced at $2/$6.

Why it matters

Sakana Fugu is not a single large language model (LLM); it is a model routing orchestrator. Therefore, benchmark claims made against monolithic frontier models blur the line between a 'better model' and 'better routing.' If an orchestrator can match or exceed top-tier models on narrow tasks at a lower cost, it simultaneously puts pressure on the pricing and positioning of open-weight products from OpenAI, Anthropic, and NVIDIA.

Highlights

• Fugu Ultra v2 outperforms GPT-6 Astra and Claude Fable 5.1 on DeepSWE, but does not include either model in its shared model pool.
• Fugu Max adds NVIDIA Nemotron (open-weight) to its pool and claims 40-60% lower output costs compared to GPT-5.6 Terra, Claude Sonnet 5, and Kimi K3.
• Pricing: Ultra v2 is $5 per million input tokens and $30 per million output tokens; Max is $2 per million input tokens and $6 per million output tokens. Both are served via an OpenAI-compatible API.

Critical take

The claim in the headline that it 'outperforms GPT-6 Astra' relies on a limited set of benchmarks (DeepSWE is mentioned) and is based on a system that is fundamentally an orchestrator rather than a trained model. This makes the comparison structurally unbalanced. Furthermore, the article lacks independent verification, task distribution breakdowns, or latency and reliability data. This leaves the 'superiority' claim as, at best, a vendor press release statement.