Model guide · September 2026

Sakana Fugu Max, fully specified

Sakana's orchestration model ships an agent pool at $2/$6 per million tokens — output priced 40% under Sonnet 5 — while keeping its context window, cache-write and batch rates unpublished.

The short version

Fugu Max is the cheap workhorse of Sakana's Fugu family: $2 input, $6 output, $0.25 cache hit per million tokens, fixed regardless of context length. Its output rate lands 40% below Claude Sonnet 5 and roughly 8x under Claude Fable 5.1. The sibling Fugu Ultra v2 costs $5/$30 and doubles above 272K tokens. Sakana publishes no context window, no cache-write or batch rates, and no confirmed API model ID — treat every gap on this page as genuinely unknown, not zero.

Full specification

Every figure below is Sakana's own published number, read from the same model registry this site prices everything with. The gaps are gaps, not zeros.

SpecificationSakana Fugu Max
Release date2026-09-11
API model nameNot confirmed — no model ID appears on the vendor page
Input / 1M tokens$2.00
Output / 1M tokens$6.00
Cache hit$0.25
Cache writeNot published
Batch APINot published
Context windowNot published
Max outputNot published
WeightsProprietary (closed weights)
Data retentionNot published
Regional availabilityNot available in EU/EEA
Subscription plans$20 / $100 / $200 per month (Standard / Pro / Max)

Verified against Sakana’s published Fugu pricing page on 2026-09-19. All rates are USD per million tokens.

Fugu Max against Fugu Ultra v2

Two paid tiers of the same pool: the fixed-rate workhorse, and the long-context flagship that doubles past 272K tokens.

RateFugu MaxFugu Ultra v2
Input / 1M tokens$2.00$5.00
Output / 1M tokens$6.00$30.00
Cache hit$0.25$0.50
Above 272K tokens (in / out / cache)Fixed rate — no surcharge$10.00 / $45.00 / $1.00
Context windowNot publishedNot published
WeightsProprietary (closed weights)Proprietary (closed weights)

Claude Fable 5.1 lists at 5x Fugu Max's input rate and 8x its output rate; against Claude Sonnet 5, Fugu Max's output rate is 40% lower — the comparison Sakana itself markets on.

Fugu rates verified 2026-09-19; Fable 5.1 and Sonnet 5 rates verified 2026-09-14. All figures are vendor-published and none have been independently benchmarked.

How the pool billing works

Fugu is an orchestration layer over a model pool, and most of its pricing quirks come from that.

  • One charge, no fee stacking

    When several agents in the pool run at once, Sakana bills a single rate based on the top-tier model involved, rather than stacking one charge per agent. That makes multi-agent runs cost-predictable — and it is also why the underlying per-model rates are not published.

  • Web tools are metered per call

    web_search and web_fetch bill $0.007 per call on top of token rates. If your workload is search-heavy, tool calls can move the bill more than the model rates do.

  • Subscriptions bundle the family

    Standard ($20/month), Pro ($100/month, 10x usage) and Max ($200/month, 20x usage) all include Fugu, Fugu Ultra and Fugu Max. Pay-as-you-go token pricing is the alternative.

  • No EU/EEA access

    Sakana's pricing page states the product is not available in the EU/EEA. Teams regulated there need a different vendor regardless of price.

What to watch next

Three gaps that could rewrite this page.

  • The context window is still unpublished

    The 272K figure circulating in comparisons is Fugu Ultra v2's pricing threshold, not a published context window. If Sakana publishes real windows, the long-context picture changes and this page follows.

  • The API model ID is unconfirmed

    Fugu Ultra v2 appears as fugu-ultra on OpenRouter and Vercel AI Gateway, but Sakana’s own page lists no API model names. Do not wire an integration against a guessed ID.

  • EU availability could arrive

    The EU/EEA restriction is a policy line, not a technical one — if it lifts, the availability row here changes the same day we re-verify.

Where these numbers come from

Every figure on this page was read from Sakana’s own Fugu pricing page and stored in this site’s model registry with a verification date. When Sakana publishes context windows, cache-write rates or an official API model ID, the registry changes first and this page follows. Community-measured throughput or quality numbers are not reproduced here.

Pricing it against the alternative

A rate sheet is not a monthly bill. Turn the numbers above into an estimate you can argue with, or see how the rest of the 2026 field compares.

Questions people ask