

Sakana Ai · 9× · vu le 19 juil. 2026
Sakana Fugu is a commercial AI product from Tokyo-based lab Sakana AI that is not a classical monolithic language model, but rather a multi-agent orchestration system delivered via a single OpenAI-compatible API. Fugu itself is a trained language model that learns to dynamically delegate queries to a pool of frontier models (including Gemini, Claude, GPT), verify their work, and synthesize responses rather than answering directly. It exists in two variants, the latency-optimized Fugu and the more powerful Fugu Ultra, and was released on June 22, 2026, after Anthropic had restricted public access to Fable 5 and Mythos Preview due to US export controls. According to Sakana AI, Fugu
Fonctionnalités
| Key Benchmark (%) | SWE-Bench Pro: 73.7% (Fugu Ultra); LiveCodeBench: 93.2% |
| Context Window (Tokens) | 1,000,000 tokens (Fugu and Fugu Ultra) |
| License | Proprietary commercial API product; model weights and orchestration service are closed-source (only technical report/setup repo public) |
| Multimodality | Text and image (vision input) per Codex CLI model catalog |
| Platform | API via console.sakana.ai, OpenAI-compatible (Chat Completions & Responses), integrates with tools like Codex |
| Price per 1M Tokens | Fugu Ultra: $5 input / $30 output (standard context); $10/$45 above 272K tokens; cached input $0.50 ($1.00 above 272K) |
| Release Date | June 22, 2026 (Fugu / Fugu Ultra, model ID fugu-ultra-20260615) |