Sakana AI
1 App · 10 Tools · 33 Intel
Sakana AI is a Tokyo-based frontier AI research and development company founded in 2023 by David Ha (CEO), Llion Jones (CTO), and Ren Ito (Chairman). Jones is a co-author of the landmark 2017 paper "Attention Is All You Need," which introduced the Transformer architecture, and Ha is a former research director at Google Brain. The company takes its name from the Japanese word for fish, reflecting its core philosophy of collective intelligence — developing nature-inspired AI approaches such as evolutionary algorithms and model merging rather than relying solely on brute-force scaling.
Is this you? Sign in with X to claim this profile.
Apps
Autonomous research agent that reasons for hours to produce structured strategy reports and options.
Tools
Agentic system that generates hypotheses, runs experiments, and writes scientific papers using progressive tree search.
Framework for automated scientific discovery in which LLMs propose hypotheses, run experiments, and write papers.
Hypernetwork approach that generates task-specific LoRA adapters for LLMs from a written task description.
Framework that lets language models adapt to unseen tasks in real time by composing expert modules on the fly.
Hypernetwork that converts a document into a LoRA, baking its factual knowledge into model weights.
Browser chat interface for the TinySwallow-1.5B model that runs entirely on-device with no API calls.
Multi-agent system behind one OpenAI-compatible API that orchestrates multiple LLMs on complex, multi-step tasks.
Python chat interface for the TinySwallow-1.5B model that runs entirely offline.
Japanese-tuned language model with a 256K-token context window, callable via Vercel AI Gateway.
Intel
Sakana AI launches Fugu, an LLM trained to coordinate a pool of other models behind one OpenAI-compatible API. Tiers are Fugu, Ultra, Max and Cyber, with per-token pricing, subscriptions, and self-reported benchmarks. Not available in the EU/EEA.
Sakana AI released Fugu-Ultra v1.1, which adds newer frontier models to its orchestration pool. It reports better scores on every benchmark shown, up to 7.9 points over v1.0, with the largest gains on ProgramBench and Terminal Bench 2.1. Pricing is unchanged.
Sakana AI releases Fugu-Cyber, a gated API endpoint of its multi-agent orchestration model, reporting 86.9% on CyberGym and 72.1% on CTI-REALM (self-reported). The post argues enterprise defense needs verification workflows and human expertise, not raw model access alone.
Sakana AI opened Namazu, a Kimi K2.6 derivative tuned on in-house Japanese business data, as an OpenAI-compatible API with web search and code execution. Priced per token with separate tool fees. Not available in the EU/EEA, UK or Switzerland; inputs may be used for training unless you opt out.
Sakana AI made its Fugu and Fugu-Ultra orchestration models available as providers in OpenCode, the open-source coding agent its team used to verify the models during development. The post lists provider and API entry points but gives no benchmarks.
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.
CoffeeBench puts LLM agents in a 90-day coffee supply chain with farmers, roasters and retailers, scored on net income. Strong models negotiate actively, while Claude Haiku 4.5 keeps writing sound analysis yet only waits, in all three trials. Code and trajectories are released.
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 Fugu orchestration models now plug into Claude Code and Codex through claude-fugu and codex-fugu launchers or ANTHROPIC_* variables. Fugu-Ultra v1.1 adds a max reasoning level. The docs note mostly cosmetic Claude Code compatibility gaps and the manual update path.
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 Picbreeder with VLM agents that branch from a shared archive, publish and rate each other. Agents fixate on attractors, longer memory hurts, and persona diversity helps most. Humans still convert accidents into larger conceptual leaps.
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 AI trained the conductor model behind Fugu on Gemma 4 E2B instead of Qwen and reports matching accuracy and cost savings on its internal held-out set. The pool of worker models was already swappable, so the orchestrator itself is now swappable too. A domestic-model conductor is planned.
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 Fugu Ultra, an orchestrator that routes each problem to one to three agents from a pool of public frontier models, is now callable on Vercel AI Gateway as sakana/fugu-ultra, with Gateway usage tracking, failover and no inference markup.
Sakana AI's UnMaskFork lets several masked diffusion language models share the unmasking of one answer, with Monte Carlo Tree Search choosing the handoffs. It beats Best-of-N and tree-search baselines on coding benchmarks without any additional training.
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 shows that closed-form Gaussian discriminants on frozen vision encoder features are often competitive with, and sometimes beat, released AI-image detector heads on Percept-Lens (39 datasets, 7.1M images). Training prior and encoder choice matter as much as the head.