Add Sakana Fugu Cyber Model

[stefvk]

Summary

Adds to the Sakana AI available models fugu-cyber, which is scheduled to become available on 24/7 and reachable via the same OpenAI compatible endpoints as the existing Sakana AI models, fugu and fugu-ultra.

Changes

• models/sakana/fugu-cyber.toml - provider-agnostic model metadata
  (1M context window, text + image input, reasoning, tool calling,
  structured output).
• providers/sakana/models/fugu-cyber.toml - Sakana serving details via
  base_model = "sakana/fugu-cyber": reasoning effort options
  (high, xhigh), 1M output limit, and pricing ($6/M input, $36/M output,
  $0.6/M cache read; tiered to $12/$54/$1.2 above 272K context).

No new provider or model family is introduced. Sakana AI and the fugu
family already exist, so provider.toml and family.ts are unchanged.

Validation

• TOML files parse cleanly and conform to the schema in
  packages/core/src/schema.ts (mirrors the existing fugu-ultra pair).
• reasoning = true is paired with reasoning_options as required.
• Merged provider model includes required limit.context and limit.output.
• bun validate passes
• packages/web bun run build passes

[github-actions]

Action items

• [medium] [possible mistake] models/sakana/fugu-cyber.toml:5-6 - Check: release_date/last_updated consistency with claimed availability. Why: Both dates are set to 2026-07-24, which is in the future relative to today (2026-07-17), yet the PR body states fugu-cyber "is reachable as via the same OpenAI compatible endpoints" (present tense), implying it is already served. A model that is reachable now should not carry a release date a week in the future; this is internally inconsistent. Action: Verify the actual go-live date of fugu-cyber on Sakana's API and set release_date/last_updated to the real availability date, or clarify in the PR body whether this is a pre-announcement.

• [low] [possible mistake] providers/sakana/models/fugu-cyber.toml:6-15 - Check: Pricing source citations. Why: The PR introduces new pricing ($6/$36/$0.6 input/output/cache_read, tiered to $12/$54/$1.2 above 272K) that is materially higher than the existing fugu-ultra pricing ($5/$30/$0.5, tiered $10/$45/$1), but the PR body provides no link to Sakana's pricing page, model documentation, or API reference supporting these values. Action: Add direct first-party citations in the PR body stating which page/reference supports each price (standard and tiered).

• [low] [possible mistake] providers/sakana/models/fugu-cyber.toml:2 - Check: Evidence for xhigh reasoning effort on fugu-cyber. Why: The reasoning-options audit requires explicit provider evidence for xhigh (it is not an OpenAI-compatible baseline value). The PR copies ["high", "xhigh"] from fugu/fugu-ultra without citing Sakana's API reference or model documentation proving fugu-cyber accepts those effort values. Action: Cite Sakana's API docs or model page proving high and xhigh are accepted effort values for fugu-cyber, or adjust values to what is verified for this model.

[stefvk]

Action items

• [medium] [possible mistake] models/sakana/fugu-cyber.toml:5-6 - Check: release_date/last_updated consistency with claimed availability. Why: Both dates are set to 2026-07-24, which is in the future relative to today (2026-07-17), yet the PR body states fugu-cyber "is reachable as via the same OpenAI compatible endpoints" (present tense), implying it is already served. A model that is reachable now should not carry a release date a week in the future; this is internally inconsistent. Action: Verify the actual go-live date of fugu-cyber on Sakana's API and set release_date/last_updated to the real availability date, or clarify in the PR body whether this is a pre-announcement.

The release date is indeed in the future, I clarified this in the comment.

• [low] [possible mistake] providers/sakana/models/fugu-cyber.toml:6-15 - Check: Pricing source citations. Why: The PR introduces new pricing ($6/$36/$0.6 input/output/cache_read, tiered to $12/$54/$1.2 above 272K) that is materially higher than the existing fugu-ultra pricing ($5/$30/$0.5, tiered $10/$45/$1), but the PR body provides no link to Sakana's pricing page, model documentation, or API reference supporting these values. Action: Add direct first-party citations in the PR body stating which page/reference supports each price (standard and tiered).

The pricing for fugu-cyber is indeed more expensive than fugu-ultra. As the release is in the future, the pricing documents will be made available on 24/7 along with all API documentation.

• [low] [possible mistake] providers/sakana/models/fugu-cyber.toml:2 - Check: Evidence for xhigh reasoning effort on fugu-cyber. Why: The reasoning-options audit requires explicit provider evidence for xhigh (it is not an OpenAI-compatible baseline value). The PR copies ["high", "xhigh"] from fugu/fugu-ultra without citing Sakana's API reference or model documentation proving fugu-cyber accepts those effort values. Action: Cite Sakana's API docs or model page proving high and xhigh are accepted effort values for fugu-cyber, or adjust values to what is verified for this model.

High and xhigh are accepted effort values just as they are for fugu-ultra. All API documents will be made available on the release date 24/7.