Fugu
Choose Fugu when you want one OpenAI-compatible API to coordinate a dynamically assembled model pool instead of selecting and orchestrating every model yourself. Standard Fugu balances performance and latency; Fugu Ultra coordinates a deeper pool of expert agents to prioritize answer quality on complex, multi-step tasks. Skip it when you need a fixed backing model or deterministic per-model costs: Fugu selects and coordinates the participating agents, while customers can exclude specific providers or models from the standard Fugu pool. The Fugu Ultra pool is fixed. For pay-as-you-go usage, Sakana bills one active agent at the underlying model's standard rate; multi-agent runs use one non-stacked rate based on the top-tier participating model. Subscriptions start at $20/month and include both routes. Sakana does not provide Fugu to users in EU or EEA member states, and says access elsewhere may vary with network conditions or local regulations.
Decision note
Fugu is an orchestration surface, not a standalone foundation model. Its learned coordinator assigns Thinker, Worker, and Verifier roles across a changing model pool, so evaluate the system on task quality, latency, billing predictability, and vendor control rather than treating its system-level results as one model's benchmark scores.
Best fit
Teams evaluating Agents or Coding workloads that want a managed, OpenAI-compatible multi-agent endpoint without building their own router, role prompts, or verification loop.
Not for
Teams that require a named fixed model, full visibility into per-request routing, or a coding agent that edits files and runs tools locally. Fugu is the model-orchestration API behind a client or agent surface, not that client or agent surface itself.
Overview
Sakana AI's hosted multi-agent orchestration API. A learned coordinator routes each request across Thinker, Worker, and Verifier roles drawn from a model pool that standard Fugu customers can constrain by provider or model; Fugu balances latency and quality, while Fugu Ultra coordinates a deeper expert-agent pool to prioritize quality. For pay-as-you-go usage, Sakana bills one active Fugu agent at the selected underlying model's standard rate; multi-agent runs use a single non-stacked rate based on the top-tier participating model. Sakana does not provide Fugu to users in EU or EEA member states, and says access elsewhere may vary with network conditions or local regulations.
Compare Fugu with other agent orchestration surfaces on routing control, model transparency, latency, and variable versus fixed pricing.
Best LLMs for this agent
Pricing and Provider Comparison
Hosted agents expose plan pricing. Bring-your-own-key agents pass the main cost through the model provider route you configure.
Agent access tiers
| Tier | Price | Includes |
|---|---|---|
| Standard | $20/mo | Both Fugu and Fugu Ultra; baseline allowance for lightweight daily usage, occasional API calls, and small experiments. |
| Pro | $100/mo | Both Fugu and Fugu Ultra; 10x Standard usage for regular Coding and Agents workloads. |
| Max | $200/mo | Both Fugu and Fugu Ultra; intended for heavy, long-running workloads. Sakana's current pricing surfaces publish conflicting Max allowances, so confirm the allowance before subscribing. |
| Fugu via Sakana AI | Variable by model | Variable pay as you go with a 1M-token context: one active agent is billed at the selected underlying model's standard rate; multiple agents are billed one non-stacked rate based on the top-tier participating model. |
| Fugu Ultra via Sakana AI | $5 in / $30 out per 1M | $0.50/M cached input through 272K context; above 272K, $10/M input, $45/M output, and $1/M cached input. |
| Fugu Ultra via Vercel AI Gateway | $5 in / $30 out per 1M | $0.50/M cached input, 1M-token context, and Vercel pass-through pricing with no gateway markup. |
Featured Model Specs
Model Compatibility Matrix
Underlying Models
Capabilities
| Feature | Supported |
|---|---|
| Chat | Yes |
| Agent mode | Yes |
| Multi-file edit | No |
| Autocomplete | No |
| Terminal | No |
| Code review | No |
| Git integration | No |
| Deploy integration | No |
| Local models | No |
| Bring your own key | No |
| MCP support | No |
| Offline mode | No |
| IP indemnity | No |
| SSO | No |
| On-prem | No |
Compare to Similar Agent Surfaces
FAQ
Should I use Fugu or Fugu Ultra?
Use Fugu for interactive work where latency and quality both matter. Use Fugu Ultra for harder multi-step tasks where answer depth matters more than response time; Sakana says Ultra can coordinate more expert agents and publishes fixed token rates for the fugu-ultra-20260615 route.
Which foundation models does Fugu use?
Sakana describes a model pool dynamically selected and coordinated by Fugu, not a fixed public backing-model list. Standard Fugu customers can exclude specific providers or models from that pool; Fugu Ultra uses the full fixed pool. That is why LLMReference does not attach Fugu to one agentModel row or attribute its system-level benchmarks to a standalone foundation model.
How is base Fugu pay-as-you-go billed?
With one active agent, Sakana charges the standard rate for the underlying model used. With multiple active agents, it does not stack every model fee; it charges one rate based on the top-tier model involved. The exact per-request rate therefore varies with orchestration rather than following one fixed Fugu token price.
What do the Fugu subscription plans include?
Standard costs $20/month, Pro $100/month, and Max $200/month. Every tier includes Fugu and Fugu Ultra, and Sakana describes Pro as 10× Standard usage. Sakana's current primary pages publish conflicting Max usage allowances, so LLMReference does not report a Max multiplier; confirm the current allowance before subscribing. Pay-as-you-go requests receive higher serving priority than monthly-plan tokens.
Is Fugu available in the EU or EEA?
No. Sakana says it does not provide Fugu to users in EU or EEA member states while it works toward compliance with GDPR and EU-specific regulations. It also notes that network conditions or local regulations may limit access in other regions.
Does Fugu replace a coding agent or app?
No. Fugu is an OpenAI-compatible API endpoint. You connect an existing client or coding agent to it; Fugu then handles model selection and multi-agent coordination behind the request. It does not provide the local file, terminal, git, or deployment surface of a coding agent by itself.