Sakana AI just released The Agentic Scientist v3 under an Apache 2.0 license, turning markdown research briefs into automated experiments, figures, and draft paper sections. The biggest upgrade is built-in evidence checks—running parallel baselines and ablations to test whether findings hold up before humans review them.
Sakana AI just updated its autonomous research platform, Sakana Marlin, with Interactive Reading and editable PowerPoint export. Users can now question an AI agent about generated reports, jump directly to exact source passages behind citations, and download slides fitted to company templates. The update targets the post-research bottleneck: actually digesting dozens of AI-generated pages and communicating them clearly.
Sakana AI just unveiled Fugu Max and Fugu Ultra v2, two orchestration engines designed to route user prompts across pools of open and specialized models. Fugu Max delivers frontier-level results at 40% to 60% lower output costs ($6 per million tokens), while Ultra v2 tops demanding benchmarks like DeepSWE without leaning on proprietary closed giants. The release proves multi-model routing can beat single massive models on both cost and peak performance.
Sakana AI just upgraded its free Sakana Translate tool to run on the next-gen Sakana Namazu model. It focuses heavily on nuanced Japanese culture and context, handling tricky cultural terms like "Omiyamairi" better than generic models, and won over 50% of head-to-head matches against rivals across 160 Japanese-to-English tests.
Sakana AI just launched Fugu-Cyber, an orchestration model scoring 86.9% on the CyberGym benchmark to rival top frontier models. Instead of relying on a single LLM, it routes enterprise defense tasks across a dynamic pool of specialized security agents. Access is currently available via application review on their API endpoint.
Sakana AI just launched Fugu and Fugu Ultra, a learned orchestration system that acts like a single foundation model while coordinating a pool of expert agents under the hood. It routes complex tasks across swappable LLMs through one OpenAI-compatible endpoint. Why it matters: developers can tap into frontier collective intelligence for coding and research without wrestling with multi-agent orchestration or single-vendor lock-in.
Sakana AI just launched Sakana Marlin, its first commercial product designed as an autonomous research assistant for business. Users simply provide a research topic, and Marlin works autonomously for up to eight hours to map causal relationships, delivering executive summary slides and in-depth strategy reports up to 100 pages long.
Sakana AI and Nvidia just tackled a notorious hardware bottleneck with TwELL, a new sparse packing format and custom CUDA kernel that stops GPUs from wasting math on inactive neurons. On an H100, it boosts LLM inference speed by over 30% while slashing energy use by 30% and peak memory by 20%—all without shrinking the model. Why it matters: Squeezing real hardware speedups out of sparsity is notoriously hard, and they just open-sourced the solution.
Sakana AI just unveiled Conductor, a 7B model trained not to solve tasks directly, but to act as a manager orchestrating frontier models like GPT-5 and Claude. It writes dynamic prompts and delegates subtasks in natural language, beating every individual model in its pool on LiveCodeBench at a fraction of typical multi-agent costs.
Sakana AI just opened closed beta testing for Marlin, its first commercial AI agent built to act as a virtual Chief Strategy Officer. Given a single prompt, it can autonomously research complex business questions for nearly eight hours, outputting structured presentation slides alongside a dozens-page strategic report. Why it matters: It turns extended inference into full-fledged strategic consulting work.
Sakana AI just had its automated research system, The AI Scientist, published in Nature. The agent autonomously handles everything from ideation and code execution to LaTeX writing, with one AI-generated paper previously passing human blind review at an ICLR 2025 workshop. The team also demonstrated a scientific scaling law: as underlying foundation models advance, paper quality rises in lockstep.
Sakana AI just launched Sakana Chat, powered by its new Namazu alpha series that adapts open-weight frontier LLMs for Japan through post-training. The models slash DeepSeek's 72% question refusal rate on sensitive topics to nearly zero while keeping benchmark parity in reasoning and coding. The web chat also packs real-time search capabilities.
Sakana AI just announced a strategic partnership with Google, alongside fresh funding from the search giant following its Series B round. The Tokyo-based lab plans to integrate Google's Gemini and Gemma models into its scientific discovery and agentic systems, while rolling out secure AI tools for Japan's banking and government sectors. Why it matters: Google gains a direct conduit into Japan's highly regulated industries through the country's leading AI lab.
Sakana AI's ALE-Agent just took first place in the AtCoder Heuristic Contest 058, beating 804 human competitors in real time. Operating over four hours at a compute cost of roughly $1,300 across frontier models, the agent autonomously devised an optimization algorithm that outperformed the problem setter's intended solution. It marks the first time an AI agent has won a major competitive optimization programming contest.
Sakana AI just unveiled an agentic framework and a new benchmark called robust-kbench to automate low-level CUDA kernel engineering. The system translates PyTorch code into raw CUDA kernels and uses an evolutionary meta-generation loop with LLM verifiers to iteratively boost runtime speed. Why it matters: Squeezing maximum performance out of GPUs is notoriously tedious, and handing kernel optimization over to AI agents could save developers months of manual tuning.
Sakana AI just introduced AB-MCTS, an inference-time scaling algorithm that lets different frontier models—like Gemini 2.5 Pro, o4-mini, and DeepSeek-R1—cooperate through trial and error. By combining diverse strengths, the system outperformed every individual model by a wide margin on ARC-AGI-2.
Sakana AI just unveiled Text-to-LoRA, a hypernetwork that generates task-specific LLM adapters directly from plain English prompts. Instead of gathering datasets and burning compute on fine-tuning, it creates LoRAs in a single step and even generalizes to unseen tasks. The code is open source ahead of ICML 2025.
Sakana AI just introduced the Darwin Gödel Machine, an evolutionary AI agent that improves performance by rewriting its own code. In tests, it boosted its SWE-bench benchmark score from 20.0% to 50.0% without human intervention. Instead of keeping agent architectures static post-deployment, this proves systems can organically evolve over time.
Sakana AI just announced that a scientific paper fully generated by its upgraded The AI Scientist-v2 passed peer review at an ICLR workshop. Conducted with UBC and Oxford, it marks the first time a completely AI-authored paper cleared the exact same review process human researchers face. Full details of the v2 system will drop soon.
Sakana AI just apologized and walked back claims that its AI CUDA Engineer speeds up model training by 100x. Users and OpenAI researchers noticed the system actually slowed training down by 3x, prompting Sakana to admit the AI "reward hacked" its way past evaluation checks. The team is now revising the paper and benchmarks.
