Independent guide, source checked July 21, 2026
Sakana Fugu, explained for people making a real buying or integration decision
Sakana Fugu exposes learned multi-agent orchestration through an OpenAI-compatible API. This guide separates current product facts from vendor benchmark claims, shows what each model actually costs, and gives you a reproducible way to evaluate it on your own work.
What this site can and cannot verify
Product names, endpoint support, model IDs, rates, and published scores are checked against Sakana AI first-party pages. We do not claim private access or unpublished hands-on results. Benchmark numbers below are Sakana AI reported results unless a separate independent source is named.
Fast orientation
What Sakana Fugu is, in one minute
Think of Fugu as a project lead behind one model name. 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.
Important distinction:
Fugu is not the same product category as OpenRouter or a self-managed LangGraph pipeline. OpenRouter gives broad model access, while a self-managed graph gives you explicit routing. Fugu's main trade is less orchestration work in exchange for less control over the learned route.
Model selection
Fugu vs Fugu Ultra vs Fugu Cyber
The right starting point depends on task shape, not which name sounds strongest.
Fugu
Best first test for interactive coding, chat, and routine work where response time matters.
- Balances quality and latency
- Routes to an underlying model based on the task
- Supports provider or model opt-outs
- Pay-as-you-go cost is variable, not one fixed public rate
Fugu Ultra
Use when answer quality matters more than latency, then verify the gain on your own hard tasks.
- Uses a fixed full agent pool
- Routes between one and three agents
- Fixed token rates make budgeting possible
- Orchestration tokens are billable
Fugu Cyber
Designed for authorized defensive security reasoning, vulnerability research, and threat investigation.
- Requires a use-case access request
- Available only on pay-as-you-go billing
- Has a higher fixed rate than Ultra
- Authorization and scope still remain the user's responsibility
Practical rule:
Start with Fugu for normal work. Run a paired comparison with Ultra on the tasks where Fugu misses important details. Consider Cyber only for authorized defensive security work and only after confirming access.
Use the transparent model chooserCurrent public rates
Pricing that includes the hidden-looking part: orchestration usage
Rates below are per one million tokens and were checked against the official console pricing page on July 21, 2026.
| Model | Input | Output | Cached input | Context over 272K |
|---|---|---|---|---|
| Fugu | Variable: underlying model rate or top-tier participating model rate | Depends on the routed model | ||
| Fugu Ultra | $5.00 | $30.00 | $0.50 | $10 / $45 / $1.00 |
| Fugu Cyber | $6.00 | $36.00 | $0.60 | $12 / $54 / $1.20 |
What the simple token headline misses
Ultra and Cyber report orchestration input, cached input, and output fields in addition to user-visible input and final output. Sakana AI says these fields represent real usage and are charged at the same input, cached, or output rate. A calculator that ignores them can materially understate a long multi-agent run.
Subscriptions are not a clean break-even formula
Standard is $20 per month, Pro is $100, and Max is $200. The public page describes Pro as 10 times Standard usage and Max as 20 times, but it does not publish an exact token allowance that can be compared across every route. Treat the plan prices as budget references, not guaranteed token equivalents.
Source: Sakana AI console pricing. The second-month subscription campaign is advertised through July 31, 2026 and is time sensitive.
Integration path
The API is familiar, but not every OpenAI parameter behaves the same
Sakana AI supports the Responses, Chat Completions, and Models APIs. Its documentation recommends the Responses API for better tool, multimodal, reasoning, and function-call behavior.
Python, Responses API
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)
response = client.responses.create(
model="fugu",
input="Review this function and list correctness risks first.",
)
print(response.output_text)Model IDs
Use fugu, fugu-ultra, or fugu-cyber. A dated Ultra alias is also currently documented.
Ignored fields
Several sampling controls are accepted but ignored. Do not assume temperature, top_p, or penalty fields change behavior.
Bounded retries
Set a finite timeout and a small retry count. A multi-agent request can be long, but an unbounded retry loop can duplicate expensive work.
Usage capture
Persist the full usage object for cost review. Final input and output fields alone do not represent all billable orchestration work.
Evidence, not decoration
How to read the Sakana Fugu benchmark table
The published table is useful evidence about the vendor's test setup. It is not proof that Ultra will beat every frontier model on your repository, prompts, tools, latency budget, or failure modes.
What the table supports
- Fugu and Ultra are competitive across several coding and reasoning suites in Sakana AI's evaluation.
- Ultra leads Fugu on many, but not all, listed tasks.
- Different tasks benefit from different orchestration depth.
What it does not establish
- Independent reproduction across all scores
- Lower latency or lower total cost
- Reliability on your private code and toolchain
- Full visibility into which agent caused a good or bad result
Use benchmarks to choose tests, not to skip tests.
Use the benchmark evaluation playbook
A fair evaluation uses the same prompts, tools, timeouts, scoring rubric, and repeat count for Fugu, Ultra, and your existing baseline. Record failures and costs, not only the best-looking response.
Deep guides
Three focused guides instead of one giant wall of repeated answers
Each guide solves one concrete decision and links every time-sensitive product fact back to a first-party source.
API quickstart and failure checklist
Choose a model ID, validate configuration, stream a response, understand ignored parameters, and keep retries bounded.
Pricing and orchestration token accounting
Reconcile the usage object, calculate Ultra and Cyber charges, and avoid making up a fixed Fugu rate.
Benchmark reading and evaluation playbook
Turn vendor scores into a blind, repeatable test on the work that actually matters to your team.
Decision checklist
A small evaluation that produces a defensible answer
- 1
Choose ten representative tasks
Include routine, difficult, ambiguous, and failure-prone work. Do not build a set made only of prompts that flatter orchestration.
- 2
Freeze the run contract
Use the same prompt, available tools, timeout, output constraint, and retry policy for every candidate.
- 3
Score output blind
Hide the model name while a reviewer scores correctness, evidence, completeness, and unnecessary claims.
- 4
Capture operational cost
Record wall time, failures, retries, and the entire usage object, including orchestration token details.
- 5
Adopt the narrow winner
Use Fugu where it wins. Escalate to Ultra where the quality delta is worth the cost. Keep the existing tool where neither earns a switch.
Verification log
What was rechecked on July 21, 2026
- Models: the official product and console pages now list Fugu, Fugu Ultra, and Fugu Cyber.
- Cyber access: official docs say it is pay-as-you-go only and requires an approved access request.
- Ultra routing: the console model page says Ultra routes between one and three agents.
- Usage fields: orchestration input, cached input, and output are separately reported and billable.
- Model naming: current docs use the stable
fugu-ultraID and retain a dated alias; one pricing surface still shows the dated identifier, so integrations should query the Models API.
Common questions
Sakana Fugu FAQ
Is Sakana Fugu free?
There is no permanent free tier shown on the current pricing page. Subscriptions start at $20 per month, and pay-as-you-go billing is available. A time-limited campaign advertises a free second month for subscriptions started by July 31, 2026.
Which model should I start with?
Start with Fugu for routine or interactive work. Test Ultra on a small set of difficult tasks where additional depth could matter. Use Cyber only for authorized defensive security work after confirming access.
Can I calculate Fugu pay-as-you-go cost in advance?
Not as one reliable fixed rate. Fugu billing depends on the underlying model or the highest tier model involved. Ultra and Cyber have published fixed rates, so the site calculator covers those two models only.
Are the benchmark results independently verified?
The table is published by Sakana AI and some baseline values are provider reported. Treat it as vendor evidence. Our evaluation guide shows how to test the same decision on your own workload without pretending we ran unpublished experiments.
Is SakanaFugu.com operated by Sakana AI?
No. SakanaFugu.com is an independent editorial guide and has no affiliation, endorsement, or sponsorship relationship with Sakana AI.
Primary sources
- Sakana AI product page: models, positioning, pricing overview, FAQ, and regional availability
- Sakana AI model documentation: model IDs, endpoints, supported fields, benchmark table, and usage fields
- Sakana AI pricing documentation: fixed rates, Fugu billing behavior, and subscriptions
- Sakana Fugu Technical Report: training approach and reported evaluation methodology
Corrections are welcome at [email protected]. Please include a first-party source when possible.