Fugu AI: what the agent does and how to choose a model

1. What the Fugu agent actually does

Fugu is a learned multi-agent system. A request can be answered directly or routed across a small team, then synthesized into one response. The orchestration is learned rather than defined by a workflow you maintain. That is the core trade: less workflow code for you, less control over the path.

Official product copy also says Fugu can exclude specific providers or models from its pool, and that new public frontier models are expected to take roughly two weeks of training and evaluation before a Fugu update. That means the behavior you measured can drift on the vendor's schedule, not yours.

2. The orchestration mechanism, in six constraints

• Fugu coordinates a pool of models through a single OpenAI-compatible (and Anthropic Messages-compatible) interface.
• Fugu allows provider or model opt-outs from its pool; Ultra uses a fixed full pool with no opt-outs.
• Ultra routes between one and three agents according to current docs.
• Routing details are proprietary and not exposed to the caller.
• Orchestration work produces extra billable tokens on Ultra and Cyber.
• New frontier models take roughly two weeks of training and evaluation before a Fugu update.

If your audit requirement is "name every worker," the answer is a self-managed graph, not a different Fugu temperature. See Fugu versus self-managed orchestration for the decision rules.

3. Orchestration tokens are extra work you pay for

OpenAI-shaped field names make it tempting to treat token_details as a breakdown of totals you already counted. Official Sakana pricing says the opposite for Ultra: orchestration fields are additional real usage billed at the matching input, cached, or output rate.

{
  "input_tokens": 120,
  "output_tokens": 80,
  "total_tokens": 200,
  "input_tokens_details": {
    "cached_tokens": 0,
    "orchestration_input_tokens": 450,
    "orchestration_input_cached_tokens": 100
  },
  "output_tokens_details": {
    "orchestration_output_tokens": 220
  }
}

Ultra's published standard rates on August 17, 2026 are $5 input, $0.50 cached, and $30 output per million tokens. Above 272K request context the rates become $10 / $1.00 / $45. The 272K threshold is per request, not a monthly total. Walk the arithmetic in orchestration token accounting.

4. Fugu versus Ultra versus Cyber: the selection table

Each row is a decision you can actually make before writing a request. Where a row says "not Fugu," the reason is a documented product rule, not a quality ranking.

The model chooser encodes region, routing visibility, and workload into the same three-way recommendation, including the "do not use Fugu here" outcome.

5. Responses API essentials

Sakana AI recommends the Responses API for new integrations, especially for tools, multimodal input, reasoning controls, and function calls. The request shape is familiar, but several fields do not behave the OpenAI way.

import os
from openai import OpenAI

api_key = os.environ["FUGU_API_KEY"]
base_url = os.environ["FUGU_BASE_URL"].rstrip("/")
if not base_url.endswith("/v1"):
    base_url += "/v1"

client = OpenAI(api_key=api_key, base_url=base_url, timeout=120.0, max_retries=2)

response = client.responses.create(
    model="fugu-ultra",
    instructions="Find correctness risks first. State uncertainty explicitly.",
    input="Review this migration for rollback risks.",
    max_output_tokens=2000,
)
print(response.output_text)
print(response.usage)

Complete client setup and streaming guidance are on the API quickstart.

6. Diagnose failures before you retry

The full failure tree is on the troubleshooting page.

Frequently asked questions

Sources and verification boundary

• Sakana AI Models documentation: IDs, agent counts, supported fields, and usage shape
• Sakana AI Pricing documentation: Ultra rates and orchestration field billing
• Sakana AI Fugu product page: agent behavior, opt-outs, and update cadence

This page describes documented behavior as of the August 17, 2026 recheck. Recheck the official Models endpoint before a 2026 production release.