Sakana Chat adds Fugu Max and cross-session memory

Sakana AI has updated Sakana Chat, its browser-based playground for the Tokyo lab’s models. The release moves the Fugu Max orchestrator into the chat interface and adds memory that can carry user-provided context across conversations.

The release post extends Sakana’s broader approach to model orchestration: route prompts among specialized open models to balance capability and inference cost. For developers, the update provides a browser interface for evaluating Fugu Max before integrating the model through Sakana’s OpenAI-compatible API.

Fugu Max reaches the chat window

Sakana Chat launched alongside the alpha version of Sakana Namazu. An August update added the original Fugu model, code execution, and file attachments. Fugu Max, previously available through the API, now replaces that earlier Fugu option in the model picker.

Fugu Max uses a learned router to select a model from a pool of open models for each prompt. Sakana says this arrangement can match the performance of frontier systems while avoiding the cost and constraints of serving every request through one large model.

Sakana Chat model picker with Namazu and Fugu Max options
Users can select Fugu Max from the Sakana Chat model menu.

Sakana’s cost and benchmark claims

Sakana prices Fugu Max at $2 per million input tokens and $6 per million output tokens. The company says its output price is 40% to 60% lower than Sonnet 5, GPT 5.6 Terra, and Kimi K3.

A Pareto-efficient result means that no evaluated alternative delivered higher performance at the same or lower cost under the reported measurements. Sakana lists Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and its internal SWEFish coding benchmark among the six tests where Fugu Max achieved the highest overall score. These figures come from Sakana’s evaluation and should be assessed against each application’s workloads, latency requirements, and quality criteria.

Using the new backend in Sakana Chat requires no configuration. Users select Fugu Max from the model menu and submit prompts through the existing interface.

Memory carries context forward

Cross-session memory allows Sakana Chat to reuse details supplied in earlier conversations, including a user’s role, current project, and preferred response format. This reduces the need to repeat the same instructions when starting a new chat.

  • Memory works with both Namazu and Fugu Max.
  • Users can review stored memories from the settings screen.
  • The feature can be disabled entirely.
  • Only conversations created after this update contribute memories; earlier chats are excluded.

Sakana presents memory as a feature of the hosted chat product. The announcement does not indicate that API requests inherit memories from Sakana Chat, so API integrations should continue to supply their own conversation state and user context.

Chat features and API access differ

The release gives developers two ways to assess Fugu Max. Sakana Chat provides an interactive environment with hosted tools and memory, while the OpenAI-compatible API supports integration into applications that manage their own state and tooling.