Fugu Max vs GPT-5.6 Luna
Pricing, context window and real answers (September 2026)
Fugu Max (Sakana AI) and GPT-5.6 Luna (OpenAI), compared as you can actually call them on FastMetal. Both are served from the same OpenAI-compatible endpoint and API key; switching is a change to the model string.
Specs and pricing
| Fugu Max |  GPT-5.6 Luna | |
|---|---|---|
| Provider | Sakana AI | OpenAI |
| Input (per 1M tokens) | ¥357.4 | ¥35.74 |
| Output (per 1M tokens) | ¥1,072.2 | ¥214.44 |
| Typical cost (1,000 in + 500 out tokens × 1,000 calls) | ¥894 | ¥143 |
| Context window | 1,000,000 tokens | 1,050,000 tokens |
| Release date | 9/11/2026 | 7/9/2026 |
| Input modalities | text, image, file | file, image, text |
| Throughput (p50) | — | — |
| Arena · overall | Unranked | #67 · ELO 1,452 |
| Arena · Japanese | Unranked | #42 · ELO 1,439 |
| Arena · coding | Unranked | #71 · ELO 1,498 |
Prices are FastMetal's yen rates before tax. Arena ranks are as of the latest leaderboard scrape.
Which should you pick?
- On price, GPT-5.6 Luna: ¥214.44 per 1M output tokens, 80% less than Fugu Max.
- For long documents, GPT-5.6 Luna: a 1,050,000-token context window against 1,000,000 for Fugu Max.
- If in doubt, try both on the same key. Switching is a change to the model string, and each is billed at its own rate.
Real answers to the same prompts
Exactly what the FastMetal gateway returned, side by side. Not benchmark scores: actual output.
Count the number of 'r's in 'strawberry'
Count the number of 'r's in 'strawberry'. Explain your reasoning step by step.
Fugu Max
The word **“strawberry”** has **3** letters **“r”**.
Step by step:
1. Write the word: **s t r a w b e r r y**
2. Check each letter:
- s — no
- t — no
- **r** — yes, 1st
- a — no
- w — no
- b — no
- e — no
- **r** — yes, 2nd
- **r** — yes, 3rd
- y — no
So, the total number of **‘r’s** in **“strawberry”** is **3**.
3.7s261 output tokens
GPT-5.6 Luna
The word **“strawberry”** is spelled:
**s – t – r – a – w – b – e – r – r – y**
The letter **“r”** appears at positions **3, 8, and 9**.
**Answer: 3**
3.0s120 output tokens
Debug This Error
I'm getting the following error in my Node.js application: TypeError: Cannot read properties of undefined (reading 'map') at UserList (/app/components/UserList.js:12:25) at renderWithHooks (/app/node_modules/rea…
Fugu Max
The error means `users` is `undefined` when `UserList` renders, so `users.map(...)` throws:
```js
function UserList({ users }) {
// if `users` is not passed, this is undefined
return (
<ul>
{users.map(user => <li key={user.id}>{user.name}</li>)}
</ul>
);
}
```
Common causes:
1. Parent renders `<UserList />` without the `users` prop.
2. State is initialized as `undefined`:
```js
const [users, setUsers] = useState();
```
3. Async data hasn’t loaded yet, but the component already renders.
4. The API returned something unexpected, e.g. `undefined`, an object, or an error response.
## Fixes
### 1. Default the prop to an empty array
```jsx
function UserList({ use…
10.6s1293 output tokens
GPT-5.6 Luna
`users` is `undefined` when `UserList` renders, so calling `.map()` on it fails. `.map()` only exists on arrays.
Common causes include:
- The parent does not pass a `users` prop.
- The prop name differs, such as `userList` instead of `users`.
- Users are loaded asynchronously and are initially undefined.
- The API response has a different shape, such as `{ data: [...] }`.
### Provide a default value
```jsx
function UserList({ users = [] }) {
return (
<ul>
{users.map(user => (
<li key={user.id}>{user.name}</li>
))}
</ul>
);
}
```
The parent should pass an array:
```jsx
<UserList users={users} />
```
If the list comes from state, initialize it as an empty…
4.8s507 output tokens
Code Review
Please review the following Python function and suggest improvements for readability, performance, and best practices: def get_data(url, retries=3): import requests import time for i in range(retries):…
Fugu Max
Here’s a review with concrete improvements.
## Main issues
1. **Bare `except:`**
- Catches everything, including `KeyboardInterrupt`, `SystemExit`, and programming errors.
- Catch `requests.RequestException` instead, plus `ValueError` for JSON parsing if needed.
2. **Retries everything, including non-retryable errors**
- A `404` or `401` will be retried even though retrying will not help.
- Retry only transient failures: network errors, `429`, and `5xx`.
3. **Fixed 2-second sleep**
- No exponential backoff. Repeated retries can hammer the server.
- Use exponential backoff, optionally with jitter.
4. **Silent failure**
- Returning `[]` after all retries makes it hard…
24.7s3795 output tokens
GPT-5.6 Luna
### Main issues
- Imports should be at module level, not inside the function.
- A bare `except:` catches everything, including programming errors and `KeyboardInterrupt`.
- The function sleeps even after the final failed attempt.
- It retries all HTTP statuses, including permanent errors such as `404`.
- `status_code == 200` excludes other successful `2xx` responses.
- `if item["active"] == True` should generally be `if item.get("active") is True`.
- A `requests.Session` can reuse connections and improve performance.
- Returning `[]` for every failure makes it impossible to distinguish “no active items” from “request failed.”
- Retries should generally use backoff rather than a fixed delay.…
23.3s1978 output tokens
Frequently asked questions
- Which is cheaper, Fugu Max or GPT-5.6 Luna?
- Per 1M output tokens, Fugu Max is ¥1,072.2 and GPT-5.6 Luna is ¥214.44 on FastMetal (yen, before tax), so GPT-5.6 Luna is cheaper.
- How do the context windows of Fugu Max and GPT-5.6 Luna compare?
- Fugu Max takes 1,000,000 tokens; GPT-5.6 Luna takes 1,050,000.
- Can I use Fugu Max and GPT-5.6 Luna with the same API key?
- Yes. On FastMetal's OpenAI-compatible endpoint, switch by passing "fugu-max" or "gpt-5.6-luna" as the model. Each is billed at its own rate from the same prepaid balance.
Try both on one API key
Create an account and add credit to call Fugu Max and GPT-5.6 Luna from the browser chat and the API. No monthly fee.