Editorial definitions
Sakana Fugu glossary used on this site
These are not official legal definitions. They are the meanings this independent guide uses after checking Sakana AI’s product, Models, and Pricing pages on September 28, 2026. If an official page later contradicts a line below, the official page wins.
Why a glossary exists here
Vendor pages reuse familiar OpenAI field names for behavior that is not identical. Readers then import the wrong meaning into a budget or a client. This page records the local meaning so later articles do not have to redefine the same trap.
Billing and usage
- Orchestration tokens
- On Ultra, and on Cyber when usage fields are returned, official docs say fields such as
orchestration_input_tokensare extra real usage. They are not a restatement ofinput_tokens. The cached counts behave differently:cached_tokensis a subset ofinput_tokensand is discounted, not added. See the accounting guide. - 272K threshold
- A per-request context threshold on the official Ultra table. Crossing it changes published rates. It is not a monthly volume switch. A month of short requests can exceed 272K tokens in total while every request stays below the line.
- Top-tier agent rate
- Official Fugu pay-as-you-go rule: one active agent is billed at that model’s rate; several agents are billed at one rate, the highest-tier model involved. There is no honest public “average Fugu price.”
- Pay-as-you-go priority
- Official pricing copy says consumption-based tokens are served ahead of monthly-plan tokens. That is a product difference, not a slogan.
- Subscription plans
- The $20 Standard, $100 Pro, and $200 Max monthly plans belong to the Sakana AI API Platform only. They include Fugu, Fugu Ultra, and Fugu Max, but not Fugu Cyber, which stays pay-as-you-go. Sakana Chat is a separate product that is free with its own usage limits; an API subscription does not change Chat limits.
- Unpublished Cyber rate
- As of September 28, 2026 the official pricing and product pages tell readers to contact sales. This site therefore does not publish a Cyber token table or calculate Cyber cost.
Models and identifiers
fugu- The default managed model. Official docs say it can exclude providers or models from its pool. Cost on pay-as-you-go is variable.
fugu-ultra- Alias that currently defaults to
fugu-ultra-v2.0. It uses a fixed, deeper agent pool with no provider or model opt-outs. Current official docs do not state an exact agent count, so this site no longer repeats a number. Pinnable older IDs remain available. fugu-ultra-v1.1- The documented earlier Ultra version, still pinnable. Use it only when you intend to pin. See pinning model IDs.
fugu-ultra-v1.0/fugu-ultra-20260615- The oldest documented Ultra version, still pinnable. Use it only when you intend to pin. See pinning model IDs.
fugu-max/fugu-max-v1.0- Alias for the Fugu Max model, which defaults to
fugu-max-v1.0. Its pool is the largest of the family and is fixed, with no opt-outs. - Fugu Max
- The largest-pool Fugu model, optimized for cost-performance. Its published rates are flat regardless of context length: $2 input / $6 output / $0.25 cached input per 1M tokens. Its built-in
web_searchandweb_fetchcalls are billed $0.007 each. See pricing explained. fugu-cyber- Security-specialized model, pay-as-you-go only, access-gated. It is not a general quality upgrade. See Cyber access.
- Models API
GET /v1/modelswith the same key and base URL. Cyber appears only for eligible pay-as-you-go keys. Treat the returned ID as the contract, not the README string.
API behavior
- Accepted but ignored
- Fields the server accepts in JSON and then discards, including
temperatureon Responses and several Chat Completions sampling controls. Do not expose them as quality sliders. Details: API surfaces. previous_response_id- Not accepted on Responses. Conversation state must be sent in
input. - Reasoning effort
- Documented values include
high,xhigh, andmax. The last two are aliases. Do not treat them as three distinct paid tiers unless an official page later says so. - Messages API
- Anthropic-compatible
/v1/messages, listed on the Models page. Compatibility is the envelope, not a full Anthropic feature certificate.
Access and region
- EU / EEA unavailability
- Official product copy says Fugu is not yet available there while Sakana AI works toward GDPR and EU-specific rules. This site treats that as a stop, not a timeout. See EU availability.
- Listed partner
- A mark on the official product page such as OpenRouter, Vercel, or opencode. It is a listing, not proof that official rates, Cyber access, or training opt-out travel with that door. See third-party access.
- Self-managed orchestration
- A graph you can inspect. Official FAQ says Fugu will not expose which underlying models answered. If that map is required, the chooser stops recommending a managed Fugu model.
What this glossary refuses to invent
It does not invent a Standard subscription token allowance, a current Cyber public rate, an independent benchmark, or a hidden routing table. Those omissions are the editorial method. Corrections with a first-party source go to [email protected].
Sources: Sakana AI product page, Models, Pricing. Checked in 2026.