Sakana AI

Fugu is an LLM trained to call other LLMs, including itself recursively, behind one OpenAI-compatible endpoint that handles selection, delegation, verification, and synthesis internally. Two tiers at launch, per-agent opt-outs for data compliance, and a stated hedge against single-vendor export cont

Sakana AI refreshed Fugu-Ultra to incorporate newer frontier models, reporting improvement on every benchmark shown, up to 7.9 points over v1.0, with the strongest results on ProgramBench and Terminal Bench 2.1. Pricing stays the same as v1.0.

Sakana claims parity with cyber-specialized frontier models on real-world security benchmarks and ships the result as a separate Fugu endpoint. The accompanying argument is that scores alone do not survive contact with enterprise defense: false positives, integration with proprietary code, and staff

Sakana AI turned its Namazu model into a public API tuned for Japanese language and business context, with web search and code execution built in. The lab shows it running weekly market research end to end, support and order analysis, and generative visual work.

Sakana AI's orchestration models, Fugu and Fugu-Ultra, are now available as providers in OpenCode, the open-source coding agent the team used to verify the models while building them. Model and API endpoints are listed for direct use.

Sakana AI's top orchestration model is now available through OpenRouter. Rather than answering from one set of weights, Fugu-Ultra dispatches a request across a pool of frontier models, and the lab has published release notes alongside the listing.

Sakana AI's first commercial product takes a research topic as its only input, then spends up to roughly eight hours forming hypotheses, gathering material and verifying its own findings before returning summary slides and a long report. It builds on the lab's AB-MCTS work.

Sakana AI and KPMG AZSA built a benchmark where six agent-run firms negotiate, order and manage stock across 90 simulated days, scored on net income. Model behavior diverges sharply, and Claude Haiku 4.5 shows an idle-drift failure: it analyzes its position but never acts.

An ICML 2026 paper with an official JAX/Flax release. Each agent solves a local convex subproblem and reconciles with neighbors through ADMM, with a network sheaf specifying which overlapping parts must agree and a running memory of unresolved conflicts. Multi-agent Sudoku: 93% solved versus 11% for

Sakana AI enforces Dale's principle by splitting each layer into excitatory and inhibitory streams with non-negative weights, routing a single global error signal to hidden units and extending it to multi-class problems. Integrated into PPO it holds up against backpropagation.

Sakana's orchestration model ships a Claude Code-compatible interface with the v1.1 Fugu-Ultra update. Instead of one model writing, debugging, and executing, a pool of frontier models is coordinated from the terminal, with release notes published alongside.

Sakana AI and NYU release Dream-Cubed, a Minecraft world dataset of tens of billions of balanced cubes, plus transformers trained with both continuous and discrete diffusion objectives that generate terrain and structures players can immediately edit and play. Code and paper are public.

Sakana AI, MIT and NYU rebuilt the Picbreeder open-ended evolution experiment with vision-language agents that branch from a shared archive, publish favourites and rate each other. Agents refine rather than leap, and cannot convert accidents into sustained discovery the way humans do.

Nature Communications, with IT University of Copenhagen and Autodesk. Each cubic brick runs the same small neural network and talks only to physically attached neighbors. The collective infers its global shape, detects missing modules at 95% accuracy, and guides its own regrowth, scaling past 18,000

Sakana Chat now runs on Fugu plus the Namazu Japanese model and gains code execution, so a prompt in Japanese produces a working web app, tool, or game in the browser. Drop in an Excel file and it runs Python for the analysis, charts, and report. Free, no login required.

Sakana AI unifies evolution strategies and consensus-based optimization under a single update rule, then derives hybrid optimizers AdaPol and SchedPol. Applied to evolutionary model merging, they locate multiple optima on a small eval set and avoid the overfitting that sinks unimodal search.

Sakana Fugu exposes multi-agent orchestration behind one endpoint, splitting a request between a swappable model pool and a conductor that decides who does what. Sakana AI retrained that conductor on open-weight Gemma 4 and reports parity on its internal evaluation set.

Sakana Translate moves to the Namazu model generation, built for Japanese cultural and contextual nuance rather than literal mapping. The team scored 160 Japanese-to-English pairs with its in-house TransEvalnia method and reports winning over half of the head-to-head comparisons against every compet

Sakana AI's ICML 2026 slate, with abstracts: test-time scaling for masked diffusion models, fast-weight product key memory that writes at inference and adapts to a new domain after 4M tokens, rubric extraction from reviewer comments, and how far a speech front-end can be shrunk.

Sakana AI's Fugu-Ultra is routable through Vercel AI Gateway, callable with the same key and provider abstraction as the other models there rather than through a direct Sakana integration.

Sakana AI's UnMaskFork has several masked diffusion language models share the unmasking of a single answer, with Monte Carlo Tree Search choosing which model fills which step. Each model inherits the others' partial output and commits where it is most confident.

Writing in Project Syndicate after the US restricted foreign access to two Anthropic models, Sakana AI co-founder Ren Ito argues sovereignty is not a national ChatGPT but retained access to frontier models and the freedom to evaluate, select and orchestrate among many.

The model behind Sakana Chat is now available through OpenRouter, bringing Japanese and business-context reasoning with web search and code execution into any workflow already routing through that gateway. Pricing is listed at $0.95 input and $4 output per million tokens.

Sakana AI keeps a general purpose vision model frozen and fits a Gaussian decision rule to its features, outperforming the strongest released AI-image detector on the Percept-Lens suite of generator, prompt and domain shifts. The signal was already in the representations.

Sakana AI launched Fugu Max and Fugu Ultra v2, orchestration systems that route tasks across pools of open-weight and specialized models (including NVIDIA Nemotron) to match or beat frontier model benchmarks like Opus 5 at a fraction of the cost, without relying on any single closed model.

Sakana AI's PC-ALM generalizes predictive coding using an augmented Lagrangian instead of energy minimization, enabling local, backprop-free learning rules to train 1000-layer networks, a depth prior predictive-coding methods failed to scale to.

A Japanese research-meeting report on Sakana AI's ESCA system, which groups social posts into hierarchical narratives rather than keywords, flags AI generation and context tampering per post, and applies a game-search-style novelty search to surface small narratives before they spread.