The AI Scientist
Sakana AI · The AI Scientist: Sakana AI's Open-Source Research Agent That Runs ML Experiments and Writes the Paper
The AI Scientist is Sakana AI's open-source system for automating computational research: its current version, The Agentic Scientist (AI Scientist v3), drives coding agents such as Claude Code or Codex through idea generation, experiments, ablations, a sealed final evaluation and a compiled LaTeX paper. It suits machine learning researchers and labs who can define a task with a scripted evaluator and run the pipeline on their own machines.
Best for
Machine learning researchers and research labs who can express a problem as a task with a scripted evaluator, have Docker, model access and (for GPU tasks) their own GPUs, and want an open pipeline that turns coding-agent runs into ablations, evidence and a draft paper
Not ideal for
Researchers looking for a hosted app or literature assistant, work in fields that need physical experiments, teams without GPU or model budgets, anyone unwilling to run agents with full tool permissions, and anyone hoping to submit generated papers without checking them
Who it's for
Machine learning researchers, AI research labs and graduate students running computational experiments, and researchers studying automated science
Capabilities
- Six-phase `agsci run` pipeline: task drafting, investigation, ablations, sealed final evaluation, post-hoc analysis and paper writing, with optional AI review and revision phases
- Orchestrates existing coding-agent CLIs (Claude Code, Codex, Cursor, OpenCode, Gemini CLI and Pi) through Headless CLI, so the same experiment can be repeated across agents and models with `agent.name` and `agent.model`
- Parallel investigation workers, each proposing a hypothesis, an implementation plan and rubrics, editing and evaluating candidate programs in its own workspace, and sharing memory and distilled conclusions at cycle boundaries
- Ablation agents build controlled baselines and ablated variants of the winning program, and a final evaluation re-scores it in a sealed environment
- Paper phase writes and compiles a LaTeX manuscript with comparison plots, AutoIllustrate concept figures and an evidence map, then audits it for unsupported numbers and overclaiming
- Optional `review` phase, also usable standalone as `agsci review <paper.pdf>`, sends the manuscript to an OpenAI-backed reviewer and records the review without discarding the paper
- By default every `agsci run` phase runs in a Docker container from a run-local task snapshot, and each run writes one resumable results directory that can be browsed with the local `agsci inspect` web viewer
- `agsci doctor` checks the host, image, credentials, GPUs and task scaffold, and `--dry-run` prints every phase command without running anything
- Hydra `key=value` overrides and checked-in task presets for idea-cycle budgets, ablation replications, timeouts, GPU counts and paper format
- Per-run cost telemetry that separates provider-reported charges from API list-price estimates
Limitations
- Agents are launched with full tool permissions and can run arbitrary commands and use the API keys you give them; Sakana says to keep runs inside the Docker image and only run tasks and sources you trust
- Each task needs a clear, verifiable goal through a scripted evaluator; drafting a task from a bare research brief is experimental and, per the README, may fail on larger tasks
- Sakana publishes no per-run cost for v3: model usage is billed by your provider or subscription, hardware costs are not counted, and the cost files are telemetry rather than billing records
- GPU task presets need exactly the GPUs they declare (the GRPO example needs two, plus about 3 GB of model weights and 1 to 2 hours per evaluation), and the paper's tasks used between one and eight H100 GPUs
- Literature search happens only during manuscript writing, not ideation, and there is no reliable pipeline for turning published methods into state-of-the-art baselines, so most comparisons are against variants of the selected program
- A high score can come from a shortcut that breaks the intended method (reward hacking), and rubric checks are themselves agent-run and imperfect
- All evaluated tasks are computational machine learning problems; use in fields that need physical experiments, instruments or human subjects is untested
- AI reviewer scores are only a proxy for paper quality, and Sakana sent just one v3-assisted paper (the ALIFE 2026 one) to external peer review
- The earlier v1 and v2 repositories are designed for Linux with NVIDIA GPUs and CUDA and use a custom license that legally requires disclosing AI use in resulting papers
Use cases
- Running an automated campaign of ideas, ablations and a draft paper on an ML research task that already has a scripted evaluator
- Comparing how different coding agents and models perform on the same research task under one orchestration scaffold
- Using the system as a co-scientist: setting the question and constraints, letting it propose and test hypotheses, then choosing which directions to pursue
- Generating controlled baselines, ablations and post-hoc analyses for a method you have already optimized
- Producing a first LaTeX manuscript tied to logged evidence, to check and revise before any submission
- Studying automated-research methods themselves with an open, resumable pipeline and full run logs
Our take
The Agentic Scientist is best read as research infrastructure, not an app. Its value comes from the scaffold around off-the-shelf coding agents: rubrics, shared memory, ablations, a sealed final evaluation and an audit that ties every number in the paper to logged evidence. That makes it a credible tool for ML groups that can write a good evaluator and want to explore more hypotheses on their own hardware. Sakana's own lessons point to the risks. Results depend on how well the task and rubrics close loopholes. Comparisons are mostly against the system's own variants rather than the literature. The one externally reviewed v3 paper had substantial human steering and checking. Plan to use it as a co-scientist whose output you verify, not as an author.
Who should use it
ML researchers and labs with a well-defined computational problem and a scripted evaluator, access to Docker, model credentials and suitable GPUs, who want to run many hypotheses with ablations and get a draft paper they will check. It also suits researchers studying automated science who need an open, inspectable pipeline.
Who should skip it
People who want a hosted research assistant or literature tool, researchers in wet-lab or field sciences, anyone without a model and compute budget, environments that cannot allow agents with full tool permissions, and anyone planning to submit generated manuscripts without independent verification and AI disclosure.
Strengths
- Apache-2.0 licensed and free, with the method described in an 88-page arXiv paper
- Covers the whole computational research loop through ablations, sealed evaluation and an evidence-mapped paper, not just optimization
- Works with six coding-agent CLIs, so you can switch models and harnesses or compare them on one task
- Docker isolation by default, resumable phases, dry runs and a local run inspector
- Official project page and docs state limitations, reward-hacking risks and ethical considerations plainly
Weaknesses
- Hard setup that assumes Docker, agent credentials and, for most real tasks, your own GPUs
- No official per-run cost for v3, and model spending is billed separately
- Only as good as the task's evaluator and rubrics; shortcuts can score well
- Weak use of prior literature during ideation and no automatic state-of-the-art baselines
- Tested only on computational machine learning tasks
The AI Scientist pricing
Open source
Free
- Apache-2.0 license for The Agentic Scientist (v3)
- Runs on your own hardware in Docker
- Bring your own agent CLI login or API keys
Free tier limits: No limits from Sakana AI. Costs come from the model provider or subscription behind the chosen coding agent, OpenAI or Gemini keys for concept figures, an OpenAI key for the optional review phase, and your own compute or GPUs.
Note: Sakana publishes no per-run dollar cost for v3. Each run writes cost telemetry that separates provider-reported charges from API list-price estimates, excludes hardware, and is not a billing record. Claude can be billed to a Claude subscription through a Claude Code OAuth token or pay-per-token with an API key, and Codex can use a ChatGPT login or an API key. For the earlier versions, Sakana's August 2024 launch post put v1 at about $15 per paper, and the v2 README estimates $15 to $20 per experiment run with Claude 3.5 Sonnet plus about $5 for writing.
Technical specs
Modalities
Text, Code, LaTeX and PDF papers, Figures
Available models
Where The AI Scientist excels
Running an overnight ML method-improvement campaign
Given a task with a scripted evaluator, parallel coding agents test ideas against rubrics and share what worked, and the pipeline then runs baselines, ablations and a sealed re-evaluation of the best program.
Co-discovering results with a human research team
Sakana used a development version this way for its ALIFE 2026 paper: researchers set the question and constraints, the system proposed and tested hypotheses, and humans chose directions and verified the results before submission.
Benchmarking coding agents on research tasks
Because the same scaffold drives Claude Code, Codex, OpenCode, Pi and other CLIs, teams can hold the task and budget fixed and compare how different agents and models perform.
The AI Scientist vs. competitors
The AI Scientist vs. OpenScience
OpenScience is an open-source, interactive research agent and workbench with desktop and CLI front ends, scientific database search, code kernels, dispatch to Modal, SSH or clusters, and an Autoresearch loop. The AI Scientist is a batch pipeline that runs coding agents against a task's scripted evaluator and carries the winner through ablations, sealed evaluation and an evidence-checked LaTeX paper. Choose OpenScience for steering day-to-day research work and The AI Scientist for evaluator-driven campaigns.
The AI Scientist vs. ClawsGO Science
ClawsGO Science is a closed-source research agent that works mainly on a hosted cloud Linux machine, with academic search and LaTeX drafting, sold on credit-based plans. The AI Scientist is free, open source and self-run, so you supply Docker, model access and GPUs, and it is built around evaluator-scored ML experiments and ablations rather than literature reviews and general analysis.
Frequently asked questions
What is The AI Scientist?
The AI Scientist is Sakana AI's open-source system for automating computational research. Its current version, The Agentic Scientist (v3), takes a research task with a scripted evaluator and uses coding agents such as Claude Code or Codex to propose ideas, run experiments, build baselines and ablations, re-evaluate the best program, and write a LaTeX paper grounded in the logged evidence.
How do The AI Scientist, AI Scientist-v2 and The Agentic Scientist differ?
The original AI Scientist (2024) starts from a human-written code template and writes papers in areas such as NanoGPT, 2D diffusion and grokking. AI Scientist-v2 (2025) drops templates and uses agentic tree search for more open-ended exploration; its README notes that v1 can still do better when a strong template exists. The Agentic Scientist (v3, October 2026) orchestrates modern coding-agent CLIs with parallel workers, shared memory, rubric-based evaluation, sealed final evaluation and evidence-checked paper writing.
Is The AI Scientist free and open source?
Yes. The Agentic Scientist (v3) is licensed under Apache-2.0. The earlier v1 and v2 repositories use The AI Scientist Source Code License, a derivative of the Responsible AI License that legally requires you to disclose AI use in any resulting paper. The software is free, but you pay for the models behind the coding agents and for your own compute.
What do I need to run The Agentic Scientist?
Docker to build the agent image, Python 3.11 or newer with uv, and credentials for your chosen agent in a `.env` file, such as a Claude Code OAuth token or Anthropic, OpenAI, Gemini or Cursor keys. An OpenAI key is recommended for concept figures and the review phase. GPUs are needed only for GPU task presets; the recommended first run is a CPU-only circle-packing example that goes all the way to a compiled paper.
How much does a run cost?
Sakana does not publish a per-run cost for v3. Spending depends on the agent, model, idea-cycle budget and task, and is billed by your model provider or subscription, plus your own GPU time. Runs record token usage and cost estimates in their metrics files. For the earlier versions, Sakana cited about $15 per paper for v1, and the v2 README estimates $15 to $20 per experiment run plus about $5 for writing.
Is it safe to run The AI Scientist on my computer?
Take care. The Agentic Scientist launches coding agents with full tool permissions, so they can run arbitrary commands and use the API keys you provide, and the task files and papers they read can steer them. Sakana recommends keeping every run inside its Docker image rather than on your workstation, running only trusted tasks and sources, and reviewing run directories before sharing, since logs can echo document contents. The v1 and v2 READMEs also warn that they execute LLM-written code and should run in a sandbox.
Have papers written by The AI Scientist been accepted?
Sakana reports two results. A fully AI-generated AI Scientist-v2 paper passed peer review at an ICLR 2025 workshop and was withdrawn before publication, as agreed with the organizers. A development version of The Agentic Scientist co-produced a paper with Sakana researchers that was accepted as an ALIFE 2026 oral; the humans set the direction and rubric and verified the code, results and references before submission. Sakana did not submit its other v3 manuscripts for review.
The AI Scientist vs OpenScience: what is the difference?
OpenScience is an interactive research agent and workbench you steer turn by turn, with literature and database search, code kernels, remote compute dispatch and an Autoresearch experiment loop. The Agentic Scientist is a batch pipeline: given a task with a scripted evaluator, it runs parallel coding agents through investigation, ablations, sealed evaluation and paper writing with little interaction. Both are open source and self-run.
Integrations & fit
Alternatives to consider
Synthetic Sciences
Computational scientists, ML researchers, and research engineers who want an open-source agent that runs analysis code and experiments on their own files and compute, with a full record of each step and their choice of model provider
Free
BitMiracle AI
Researchers, graduate students, analysts and research groups who want to hand a long literature, analysis or drafting task to a hosted agent with a ready-made research environment and come back to files they can check and edit
FreemiumAbout The AI Scientist
The AI Scientist is a line of research systems from Sakana AI, and The Agentic Scientist, released in October 2026 in the AI-Scientist-v3 repository under Apache-2.0, is the current one. Sakana calls it its next-generation AI Scientist. Instead of its own agent loop, it wraps existing coding-agent CLIs (Claude Code, Codex, Cursor, OpenCode, Gemini CLI or Pi, through the Headless CLI runtime) in a scaffold that controls what they see, what they write and how their work is scored. You start from a task directory with a `task.md`, a program and a scripted evaluator, then `agsci run` works through six phases: task drafting, investigation, ablations, final evaluation, post-hoc analysis and paper. In investigation, parallel workers each propose a hypothesis with rubrics for what the experiment must show, edit and evaluate candidate programs, and share memory entries at cycle boundaries. The best correct program is then compared against baselines and ablated variants, re-scored in a sealed environment, analysed, and written up as a LaTeX paper with comparison plots, AutoIllustrate concept figures and an evidence map linking claims to sources. An optional review phase sends the manuscript to an OpenAI-backed reviewer. Everything for a run lands in one results directory that you can resume phase by phase and browse in a local web inspector. Setup is Docker plus Python 3.11 or newer with uv, with API keys or a Claude or ChatGPT subscription login for the chosen agent; GPUs are needed only for GPU task presets. Earlier versions are still on GitHub: the template-based AI Scientist (v1, August 2024) and the template-free AI Scientist-v2 (April 2025), which uses agentic tree search and wrote a paper that passed peer review at an ICLR 2025 workshop and was withdrawn before publication, as agreed in advance. A paper on that work appeared in Nature in March 2026. Both use a custom AI Scientist Source Code License that legally requires disclosing AI use in resulting papers. For v3, a development version co-produced an ALIFE 2026 oral paper with Sakana researchers, who set the direction and verified the results, code and references before submission. The tradeoffs: the software is free, but model usage and compute are yours, and Sakana publishes no per-run cost for v3. Agents run with full tool permissions. Tasks need a clear scripted evaluator. The system searches the literature only while writing, not while generating ideas. Every claim still needs expert checking.
Updates from The AI Scientist
Sakana AI published the paper on The Agentic Scientist, the third generation of The AI Scientist, on arXiv and released the code under Apache-2.0 in the AI-Scientist-v3 repository. It orchestrates coding agents through task drafting, investigation, ablations, final evaluation, post-hoc analysis and paper writing. A development version co-produced a paper with Sakana researchers that was accepted as an ALIFE 2026 oral presentation.
Sakana AI, with the University of British Columbia, the Vector Institute and the University of Oxford, published an open-access Nature paper describing The AI Scientist and AI Scientist-v2, including new scaling results.
Sakana AI released the AI Scientist-v2 code and technical report. v2 drops the human-written code templates of v1 and uses a progressive agentic tree search guided by an experiment manager agent.
A fully AI-generated paper from AI Scientist-v2 passed double-blind peer review at the ICLR 2025 ICBINB workshop, run with the organizers' cooperation. As agreed in advance, it was withdrawn before publication.
Sakana AI introduced The AI Scientist with researchers from Oxford and UBC: a template-based system that generates ideas, runs experiments, writes papers and reviews them automatically, at about $15 per paper according to the launch post.
Are you the founder? Claim this listing →