Overview
Sakana Fugu is an orchestration model from Sakana AI. It is not a single large model but a routing layer: it reads each request, judges how hard the task is, then assembles a combination of open and specialized models to answer it. On September 11, 2026 the line split in two directions — Fugu Max chases the lowest cost at comparable quality, while Fugu Ultra v2 chases peak quality on complex multi-step work. Both share one orchestration architecture and differ only in routing policy.
Across the milestones from an April beta to the September release, the claim Sakana keeps making is the same: an orchestration layer can outperform any single model it routes to.
Key Features
- One API, two targets: Fugu Max serves high-volume pipelines with uneven task difficulty, assigning each request to the lightest model expected to solve it. Fugu Ultra v2 serves long-horizon reasoning, autonomous research and full-stack development. The model IDs are `fugu-max` and `fugu-ultra`, with reasoning depth switched through `reasoning.effort` across high, xhigh and max.
- Priced on the cost-performance frontier: Fugu Max is $2 input and $6 output per million tokens. On Sakana's published comparison its output price sits 40% to 60% below Sonnet 5, GPT 5.6 Terra and Kimi K3, and it pushes the cost-performance boundary outward on seven of ten benchmarks.
- Ultra v2 scores on hard tasks: On Chartography, which requires sustained reasoning over complex visual and structured data, Fugu Ultra v2 scored 48.3. On DeepSWE, a software engineering benchmark, it scored 74.3. Across eight benchmarks it took best or joint-best on GDP.pdf, Chartography, SWEFish, DeepSWE and Toolathon, and placed top two on seven.
- A swappable model pool: Fugu Ultra v2 reached those scores without Fable 5, Fable 5.1 or GPT-6-Astra in its agent pool, relying on open and specialized models instead, including the NVIDIA Nemotron family through a collaboration with NVIDIA. Sakana frames the design as supply chain resilience: a swappable pool means one provider cutting service does not stop the product.
- Fits existing coding workflows: A one-line installer wires Fugu into the Codex CLI and Claude Code, launched afterwards as `codex-fugu` or `claude-fugu`, with Ubuntu and macOS supported. The API layer is compatible with OpenAI Chat Completions and Responses as well as Anthropic Messages, so a standard OpenAI SDK works after changing `base_url`.
- Fugu Cyber for security work: A specialized orchestration variant for security analysis, vulnerability research and threat investigation, reaching 86.9% on CyberGym and 72.1% on CTI-REALM. It is pay-as-you-go only and requires an access request reviewed manually by the Sakana team.
Use Cases
- High-volume pipelines where per-request difficulty varies widely, using Fugu Max to hold costs down
- Autonomous research, cross-file refactors and complex report reasoning, switched to Fugu Ultra v2
- Teams already on the Codex CLI or Claude Code, adding it as a second provider for redundancy
- Security teams doing vulnerability analysis and threat intelligence triage, using Fugu Cyber
Pros
- Moves model selection out of application code and into a routing layer
- Compatible with both OpenAI and Anthropic interfaces, keeping migration cheap
- A swappable model pool cuts dependence on any single vendor
- Max-tier pricing sits on the cost-efficient side of comparable frontier capability
Pricing
Billed per token. Fugu Max is $2 input and $6 output per million tokens. Fugu Ultra v2 is positioned higher, with price following whichever strongest model gets scheduled. Fugu Cyber is pay-as-you-go only. Enterprises needing a specific model pool or provider configuration can contact Sakana.
Summary
Fugu's value is not in any single benchmark number but in moving model selection into the routing layer. For high-volume workloads with a wide difficulty spread, that abstraction saves real money; for peak quality without being tied to one closed vendor, Ultra v2 is a usable option. Two things to keep in mind: the Ultra v2 training cutoff is August 28, 2026, and the benchmark figures currently come from Sakana's own launch materials rather than independent reproduction.
Version History
- Fugu Max / Fugu Ultra v2 (2026-09-11): Fugu Max widens the model pool further at $6 output per million tokens; Fugu Ultra v2 scores 48.3 on Chartography and 74.3 on DeepSWE
- Sakana Chat & NVIDIA partnership (2026-08): Orchestration reaches daily consumer use and starts integrating NVIDIA Nemotron open models
- General availability & Fugu Ultra v1 (2026-06): An orchestration layer matches closed frontier models on hard benchmarks
- Fugu Beta (2026-04): Multi-agent orchestration shown to work as a unified foundation model