Sakana AI

💰 Pricing & Plans

🔧 API & Developer Experience

• •Endpoint: OpenAI-compatible base URL at console.sakana.ai — most OpenAI SDKs work by swapping base_url and api_key. A single key covers all four Fugu variants; switching is a single parameter change in the request payload, not a separate integration.
• •Streaming & Responses: Standard Chat Completions streaming (SSE). The Responses API breaks out orchestration tokens separately from user-visible model work — a Sakana-specific addition not present in standard OpenAI responses.
• •Orchestration Pricing: When a single underlying model is selected, you pay only that model's standard rate. When multiple agents fire, Sakana does not stack fees — only the top-tier model involved is billed at its standard rate.
• •Web Tools: web_search and web_fetch are billed at $0.007 per call on top of token charges. Useful for retrieval-augmented agents; budget for the per-call cost separately because it can dominate token cost for search-heavy workloads.
• •Subscription vs Token Plan: Standard subscription at $20/month is positioned for lightweight daily usage and is served at standard priority. Token Plan (pay-as-you-go) is served at higher priority than subscription tokens — recommended for production workloads even when a subscription is active.
• •Regional & Latency: Japan-origin hosted API; no mainland-China edge. Direct connections from China require a stable overseas proxy. No published SLA tier; Sakana does not currently publish latency benchmarks per region.

🐡 The Fugu Thesis: Orchestration Over Single-Model Stacking

Sakana's defining bet is that the cost-performance frontier is two-dimensional: capability on one axis, cost on the other. Rather than building one massive frontier model and scaling it, Sakana orchestrates a swappable pool of open and specialized models (NVIDIA Nemotron among them) and routes each request to the leanest model capable of solving it. The result is Fugu Max expanding the Pareto frontier northwest — performance within striking distance of elite models at 40-60% lower output cost than Sonnet 5, GPT 5.6 Terra, and Kimi K3 — and Fugu Ultra v2 pushing peak performance upward while explicitly not relying on Fable 5, Fable 5.1, or GPT-6-Astra. The architecture is the product; the underlying weights are replaceable. For teams that want frontier-class results without vendor lock-in to a single model's roadmap, that is the differentiator. The trade-off is structural: orchestration latency is bounded by the slowest agent in the pool, and the cost model has a soft ceiling that does not always reflect the work done.

🌐 China Access & Latency

Sakana AI is a Japan-based lab serving its Fugu API from Japanese infrastructure with no mainland China regional endpoint. Production use from mainland China therefore requires a stable overseas proxy or an aggregator that fronts the Sakana API; direct connections typically time out or face elevated latency. Real-world latency from China is dominated by the cross-border round trip to Japan, making Fugu a poor choice for latency-sensitive interactive workloads and a workable choice for batch jobs and agentic workflows that tolerate a slower first byte. There is no China-specific pricing, billing, or data-residency option. For teams targeting Chinese end users, a domestic LLM (Qwen, DeepSeek, Zhipu, Doubao) is the latency-safe default, while Sakana remains attractive for Japan-hosted workloads, English/Japanese multilingual tasks, and teams that want the orchestration architecture rather than a single-model provider.