Models

Fugu Ultra

TL;DR

Premium

Multi-agent orchestration across frontier models.

Available on

ProviderMax 10×Max 20×Teams ProEnterprise

Not on the Go plan ($1/mo, open-source models only).

Switch with

/model

Pick Fugu Ultra from the selector.

Intelligence Index

Not yet scored

Speed

tokens / sec

Input

$5

per M tokens

Output

$30

per M tokens

Fugu Ultra in Command Code

Fugu Ultra is Sakana AI's premium model — multi-agent orchestration across frontier models. It runs in Command Code with a 1M-token context window, switchable any time with /model.

Fugu Ultra has no published Intelligence Index yet — it renders as "not yet scored" rather than borrowing a number.

Fugu Ultra specs at a glance

What Fugu Ultra accepts, how much it can hold in context, and what you need to run it.

SpecFugu Ultra
Context window1M tokens
Input modalitiestext + images
ReasoningYes — efforts: high, xhigh
Minimum planProvider

Fugu Ultra vs the Command Code lineup

Fugu Ultra alongside its nearest real alternatives in the lineup — every price straight from the billing tables.

ModelIntelligenceCodingSpeedInput $/MOutput $/MBlended $/MContext
Claude Fable 559.976.5~62 tok/s$10.00$50.00$201M
GPT-5.6 Sol58.977.4~77 tok/s$5.00$30.00$11.251.1M
Claude Opus 4.855.774.3~55 tok/s$5.00$25.00$101M
GPT-5.554.874.9~81 tok/s$5.00$30.00$11.25200K
Claude Opus 4.753.573.6~52 tok/s$5.00$25.00$101M
Claude Sonnet 4.647.263~55 tok/s$3.00$15.00$61M
Fugu Ultra (this page)$5.00$30.00$11.251M

What Fugu Ultra is best for

Fugu Ultra is a premium model from Sakana AI, priced at $11.25 blended per million tokens.

The honest way to place it: run your own session with /model and compare against the lineup table above — the numbers on this page update as the registry and billing tables change.

When to switch away from Fugu Ultra

No single model wins every task. These are Fugu Ultra's computed nearest alternatives — one step up, one step down in cost, one for speed, one from the same family — each switchable mid-session with /model.

Switch to Claude Opus 4.8

Claude Opus 4.8 runs about 11% cheaper blended ($10 versus $11.25 per million tokens) while scoring 55.7 on the Intelligence Index. Switch down for high-volume work where Fugu Ultra's edge isn't earning its rate.

Switch to GPT-5.6 Terra

GPT-5.6 Terra streams ~155 tokens/sec at a comparable blended cost ($5.625 per million tokens). Use it when iteration speed matters more than squeezing out the last point of quality.

What you pay for Fugu Ultra

Fugu Ultra is billed per token at the rates below — the same billing tables the Usage page charges against, so this page cannot quote a different price than you pay.

Blended cost (3:1 input:output, the shape of a typical coding session) works out to $11.25 per million tokens.

Per 1M tokensInputOutputCache read
All requests$5.00$30.00$0.50

In Command Code: caching and taste-1

Two things change the experience of using this model inside Command Code versus calling it directly through the upstream API.

First, prompt caching is on by default. In an agent loop the same context is read across many steps; cache reads are billed at $0.50 per million tokens versus $5.00 for fresh input.

Second, taste-1 sits between the model and the agent loop, rewriting and reranking candidate edits to match your codebase conventions.

Plan availability

Fugu Ultra is a premium model — it needs the Provider plan or higher. Not on the Go plan ($1/mo, open-source models only).

Command Code is a subscription with model usage at API rates. Each plan ships with monthly LLM credits; credits roll over and never expire, and auto top-up keeps you running if you go over.

PlanPrice/moLLM creditsModels
Go$1$10Open-source only
Pro$15$30Open-source + premium
Provider$15Pay as you goOpen-source + premium
Max 10×$100$150Open-source + premium
Max 20×$200$300Open-source + premium
Teams Pro$40 / seat$40 / seatOpen-source + premium
EnterpriseCustomCustomCustom pool, SSO, audit logs

Switching models with /model

In an interactive Command Code session, run /model to open the model selector. Pick Fugu Ultra and it applies to this session and to future sessions until you change it again. Premium models require Pro or higher; open-source models are available on every plan, including Go.

cmd               # start an interactive session
/model            # open the selector and pick Fugu Ultra

All Command Code models, ranked by quality and speed

Quality is the Intelligence Index — an aggregate score across reasoning, math, coding, and knowledge evaluations. Speed is measured output tokens per second. Models without a published score are noted. This table is regenerated from the model registry, so it is always current.

ModelTierIntelligence IndexOutput speed
Claude Fable 5Premium59.9~62 tok/s
GPT-5.6 SolPremium58.9~77 tok/s
Claude Opus 4.8Premium55.7~55 tok/s
GPT-5.6 TerraPremium55~155 tok/s
GPT-5.5Premium54.8~81 tok/s
Grok 4.5Open-source53.8~114 tok/s
Claude Opus 4.7Premium53.5~52 tok/s
Claude Sonnet 5Premium53.4~79 tok/s
GPT-5.4Premium51.4~164 tok/s
GPT-5.6 LunaPremium51.2~234 tok/s
GLM-5.2Open-source51.1~208 tok/s
Muse Spark 1.1Premium50.6~130 tok/s
Gemini 3.5 FlashPremium50.2~236 tok/s
Claude Sonnet 4.6Premium47.2~55 tok/s
Qwen 3.7 MaxOpen-source46~196 tok/s
MiniMax M3Open-source44.4~113 tok/s
DeepSeek V4 ProOpen-source44.3~62 tok/s
GPT-5.3 CodexPremium44.3~106 tok/s
Kimi K2.6Open-source44.2~43 tok/s
MiMo V2.5 ProOpen-source42.2~56 tok/s
Kimi K2.7 CodeOpen-source41.9~46 tok/s
Tencent Hy3Open-source41.2~58 tok/s
DeepSeek V4 FlashOpen-source40.3~106 tok/s
GLM-5.1Open-source40.2~81 tok/s
Qwen 3.6 Max PreviewOpen-source40~46 tok/s
GPT-5.4 MiniPremium40~171 tok/s
Qwen 3.6 PlusOpen-source39.6~53 tok/s
GLM-5Open-source39.5~51 tok/s
Qwen 3.7 PlusOpen-source39~52 tok/s
Kimi K2.5Open-source38.1~51 tok/s
MiniMax M2.7Open-source38.1~49 tok/s
Nemotron 3 UltraOpen-source37.8~204 tok/s
MiMo V2.5Open-source37.2~88 tok/s
MiniMax M2.5Open-source33.7~78 tok/s
Step 3.7 FlashOpen-source30.3~407 tok/s
Step 3.5 FlashOpen-source26~207 tok/s
Gemini 3.1 Flash LitePremium25~300 tok/s
Claude Haiku 4.5Premium23.7~103 tok/s
Kimi K3Open-sourceNot yet scored
Kimi K2.7 Code HighSpeedOpen-sourceNot yet scored
GLM-5.2 FastOpen-sourceNot yet scored
InklingOpen-sourceNot yet scored
Fugu Ultra (this page)PremiumNot yet scored

Frequently asked questions

Fugu Ultra or Claude Opus 4.8?

Claude Opus 4.8 is about 11% cheaper blended ($10 vs $11.25 per million tokens), scoring 55.7 on the Intelligence Index. Use Claude Opus 4.8 for volume work and Fugu Ultra where its edge earns the difference.

How much does Fugu Ultra cost in Command Code?

$5.00 per million input tokens and $30.00 per million output tokens, with cache reads at $0.50. In an agent loop, cached context brings effective input to roughly $1.85 per million tokens.

What plan do I need for Fugu Ultra?

Fugu Ultra requires the Provider plan or higher. The Go plan ($1/mo) covers open-source models only.

Does Fugu Ultra support image input and reasoning?

Yes — Fugu Ultra accepts image input alongside text. It supports reasoning with selectable effort (high, xhigh).

Which Command Code model should I use?

Claude Fable 5 currently leads the lineup on the Intelligence Index (59.9). Grok 4.5 (53.8) leads the open-weights tier, available on every plan. For fast lookups, Step 3.7 Flash streams ~407 tok/s. There is no single right answer — switch per session with /model and let the task pick the model.

Can I mix Fugu Ultra with other models in a workflow?

Yes. Switch per session using /model. Common pattern: keep a default model and switch up for hard problems or down for quick lookups as the task calls for it.

What does Fugu Ultra actually do?

It is an orchestration product: instead of one model answering, it routes across a pool of frontier models and coordinates them. The listed price is the base tier (context ≤272K); above that the upstream bills roughly 2×.

Which plans include Fugu Ultra?

Max 10×, Max 20×, Teams, and Enterprise. It is not available on the Go or Pro plans.

Is Command Code free to try?

The Go plan starts at $1/mo with $10 in LLM credits. It covers open-source models only. Pro at $15/mo unlocks premium models with $30 in LLM credits.

Does Command Code train on my code?

No. Command Code does not train on your code or store your code snippets. taste-1 data is stored locally in your project directory.

Where can I track my usage?

The Usage page in Studio shows per-request cost, token counts, and which model ran. Settings > Billing lets you change plans, buy credits, or enable auto top-up.

Does Command Code replace my editor?

No. Command Code is editor-agnostic — it runs as a CLI and works alongside any editor (Cursor, VS Code, Zed, JetBrains, Neovim, etc.).

Related reading

Ship code that matches your taste

Command Code is the AI coding agent that continuously learns your taste. Start for $1.