Editorial verification

Source-check log: what official Sakana Fugu pages said

This is not a changelog from Sakana AI. It is this site’s dated comparison of first-party pages across rechecks, newest first. The point is to show which numbers and IDs a reader should stop trusting.

How the rechecks were done

We opened the live product page, Models documentation, and Pricing documentation on August 17 and again on September 28, 2026, and compared them with the statements this site had published earlier. We did not receive a private briefing and we did not run unpublished model evaluations. Where a page was silent, the log says silent.

September 28, 2026 source recheck and corrections

Added

Fugu Max joins the family

Official pages now list a fourth model, fugu-max (defaults to fugu-max-v1.0). It orchestrates the largest pool, uses a fixed pool with no opt-outs, and publishes flat rates of $2 input / $6 output / $0.25 cached per 1M tokens at any context length. Its built-in web_search and web_fetch calls bill at $0.007 each.

Updated

The fugu-ultra alias moved to v2.0

The alias now defaults to fugu-ultra-v2.0. The older fugu-ultra-v1.1 and fugu-ultra-v1.0 (the former fugu-ultra-20260615) remain documented for pinning. Published Ultra rates are unchanged.

Clarified

Subscriptions are for the API, not Sakana Chat

The $20 / $100 / $200 tiers are Sakana AI API Platform subscriptions and include Fugu, Fugu Ultra, and Fugu Max — not Cyber. Sakana Chat is a separate product that is free with its own usage limits, and a subscription does not change those limits.

Dropped

"One to three agents" for Ultra is no longer documented

Current docs describe Ultra and Max as fixed pools but no longer state an exact agent count. This site removed the specific figure and now describes the pools instead of a number.

Corrected

Residual Cyber estimates were still misleading

The August entry below recorded the withdrawal of public Cyber rates, but worked examples on the pricing guide still quoted exact Cyber totals and a 20% premium. We removed those remaining estimates in this recheck. There is no published rate from which to calculate them.

Corrected

Cached tokens are a subset of the input total, not an extra charge

Our pricing formula and worked examples had added cached_tokens to input_tokens. The official usage shape follows the ordinary OpenAI convention instead: cached_tokens is already inside input_tokens, and orchestration_input_cached_tokens is already inside orchestration_input_tokens. We now subtract each cached count from its own input total before applying the full input rate, and add the two cached counts at the discounted cached rate. The orchestration worked example therefore falls from $37.875 to $34.125, and the pricing-guide rows move to $11.38 / $34.13 / $43.50 for Ultra and $3.06 / $7.25 for Max. The orchestration fields themselves remain extra usage outside the ordinary request counts. Sakana's own Namazu section on the same page states that cached_tokens is a subset of the input token count, and OpenAI documents the same relationship for cached_tokens.

Facts that changed on August 17, 2026

Withdrawn

Public Fugu Cyber token rates

July 21 material on this site repeated a Cyber table of $6 / $36 / $0.60 per million tokens, with a long-context row. On August 17 both the product page and the pricing page said to contact sales instead. Those numbers are removed from the calculator and from pricing pages.

Updated

Ultra and Cyber versioned IDs

The Models page now lists fugu-ultra-v1.1 as the default behind fugu-ultra, keeps fugu-ultra-v1.0 as the former fugu-ultra-20260615, and lists fugu-cyber-v1.0.

Added

Anthropic-compatible Messages API

Official Models documentation now lists /v1/messages beside Responses and Chat Completions. Responses remains the recommended generation surface.

Listed

Partner marks on the product page

The product page currently shows OpenRouter, Vercel, opencode, Creao, and Merge. That is a listing, not a verified copy of official billing or Cyber policy. See the access map.

Noted

Sakana Namazu on the same pricing page

The official pricing page now also documents sakana-namazu-v1.0 with its own rates and tool fees. This site does not fold Namazu into Fugu estimates.

Facts that stayed the same across both rechecks

  • Ultra published rates remain $5 / $30 / $0.50, and $10 / $45 / $1.00 above 272K context.
  • Orchestration token fields remain extra billable usage at the same category rates.
  • Standard Fugu pay-as-you-go remains a single top-tier-agent rate, not a public fixed Fugu rate.
  • Subscriptions remain $20 / $100 / $200 with 1× / 10× / 20× language and no public token allowance; as of September 28 they also include Fugu Max.
  • Pay-as-you-go tokens are still described as higher priority than monthly-plan tokens.
  • EU and EEA unavailability is still stated on the product page.
  • Cyber remains pay-as-you-go and access-gated.
  • Routing remains undisclosed. Agent opt-out remains a Fugu feature, not an Ultra feature.

What this site changed because of the rechecks

We expanded the guide library so each new official gap has a page with a decision, not a rewritten homepage, and we removed unpublished Cyber math from the calculator. The September 28 recheck added Fugu Max and its flat rates, moved the Ultra alias to v2.0, reworked every subscription section around the API-versus-Chat split, and corrected the cached-token formula in the pricing guide and calculator. Corrections that follow another official change should go to [email protected] with the page URL and the first-party source.

The cost calculator now handles Max tool-call fees, rejects invalid counts, and clears an invalid quote. Mobile navigation can be closed with Escape, returning focus to its toggle. The analytics banner is removed from the accessibility tree and tab order when dismissed, returning focus to its settings button; reopening it focuses the first choice. It is not a certified advertising consent-management platform. We also replaced a legacy sharing image that described a website template rather than this guide, completed sharing metadata, and stopped assigning one revision date to every sitemap URL.

Accepted large counts now retain exact decimal costs through currency rounding, rather than losing cents through floating-point arithmetic. Long totals wrap on small screens. The local acceptance server now uses the same Pages routing as the deployed site, including extensionless guide URLs.

The analytics banner now states the same boundary as our privacy policy: choosing analytics does not grant advertising consent. It no longer implies that a certified advertising consent platform is already configured.

If you are reading a cached copy of this site, check the date in its header. A July copy still carries the withdrawn Cyber rate table and an outdated promo banner; the August 17 copy still calls the family three models and stops at Ultra v1.1.