OpenScience Review (2026): Pricing, Features & Honest Verdict

TLDR

OpenScience is an open source agent workbench for scientific research: it reads papers, writes and runs code, and reports results with provenance. The software is free under Apache 2.0, but the required account, default trace sharing and hidden model rates need reading. Best for: computational scientists and ML researchers. Price: from $20 Wallet top-up, free plan yes. Rating: 7/10.

What is OpenScience?

We found OpenScience on Product Hunt, where it launched on 27 September 2026 as an open source AI workbench for scientific research. It comes from Synthetic Sciences, an AI research lab, and the repository has been public since 3 July 2026. In under three months it has reached 3,649 stars and 490 forks, with the latest release, v2.0.139, shipped on 26 September 2026.

The pitch is one workspace for literature, code, experiments, compute and results. You describe the work in plain language. The agent plans it, queries scientific databases, runs Python or R, and hands back a report with every step on the record. It runs as a desktop app for macOS, Windows and Linux, as a browser workspace launched from the terminal, or headless in a pipeline.

What Are OpenScience’s Key Features?

Scientific databases and skills as tools

OpenScience ships 42 connectors to scientific databases across six domains, including UniProt, PDB, ChEMBL, PubChem, arXiv, PubMed and Semantic Scholar. It also bundles 371 skills in 18 categories, written procedures for things like a single cell analysis, a docking screen or a DESeq2 contrast. Many are borrowed: 180 come from K-Dense and 79 from a second open project, with attribution files in the repo.

Compute that outlives the conversation

Everyday work runs in local Python and R kernels and a shell, contained by macOS Seatbelt or Linux bubblewrap when containment is on. Heavier jobs go to an SSH host, a Slurm or PBS cluster, or Modal as long-running jobs that keep running after the chat ends. Each dispatch needs your approval, and the Compute pane tracks status, failures and outputs.

Autoresearch studies

Autoresearch is a pane beside Files, Terminal and Compute that runs a hill climbing study over many training or analysis runs. You name one metric and a direction, it establishes a baseline, ranks ideas by expected value, runs exactly one experiment per idea, kills runs that breach your criteria, and keeps a ledger of what was kept or reverted. Studies run locally, on a saved SSH host, or on Modal.

A lab of worker agents

The lead agent can hand bounded pieces of work to seven workers at the same time: an explore scout plus ml, biology, physics, chemistry, data and general specialists. Workers cannot spawn workers or publish anything. A report without a verification section counts as a claim, and when two workers compute the same quantity and disagree, the disagreement is reported as a finding. That is a sensible rule for scientific work.

What Is It Like to Use OpenScience?

Install is one line: npm install -g @synsci/openscience, or curl -fsSL https://openscience.sh/install | bash on macOS and Linux, or npx synsci to try it without installing. Note the package name. The bare npm package called openscience, version 0.0.2, is an unrelated 2020 upload with no description, and there is no openscience package on PyPI at all. Install the wrong one and you get nothing.

First launch walks through four steps, and step one is creating or signing in to a Synthetic Sciences account. The docs call it required for the desktop and CLI setup; only headless openscience run skips it. After that you turn Ace on or skip it, connect your own providers (ChatGPT or Codex sign in, provider keys, Firecrawl, Modal) and open the workspace.

The workspace itself is a chat with a trace: thought, searches, skills loaded, code run, files written, then the answer. Results carry a provenance graph linking each figure to the code and sources that produced it. Sessions can be forked, and several agents can run side by side on the same project. A /plan command lets you agree the method before the agent spends any tokens.

Where Does OpenScience Fall Short?

Trace sharing is on by default while you are signed in, and it covers sessions that ran on your own provider keys or a local model, not just Ace routes. A trace is the full trajectory: your prompts, the model’s reasoning, and every tool input and output. The terms go further and say research content may be used for model training, with no preset deletion date for retained copies. You can switch it off, but you have to know to look.

The #1 claim rests partly on the vendor’s own numbers. OpenScience Bench, where it posts 47.3% pass@3 against 40.6% for Claude Science and 38.1% for Codex, is an internal benchmark. On the public Terminal-Bench Science and Terminal-Bench 4.0 runs, the vendor reports one counted trial per problem while leaderboard entries average three to five. The traces are published, which is good, but the comparisons are not like for like.

Ace pricing is invisible on the website. The Ace page and docs explain the mechanics in detail, yet the actual per-million-token rates only appear inside the client under Customize, Models, Rates and limits. Some models use variable schedules labelled Up to, which the docs admit are estimates. You cannot budget a project from the marketing site.

It is also very young. The repo is under three months old, the release cadence is a version bump most days, and ten open issues on a 3,649 star project suggests either tidy triage or a user base that has not stressed it yet. The desktop app is signed on Windows under InkVell Inc., but the docs warn SmartScreen may still object.

How Much Does OpenScience Cost?

The software costs nothing. It is Apache 2.0 licensed and the FAQ says it comes with free models so you can start without a provider account, although neither the site nor the docs list which models those are. Bring your own provider keys, a ChatGPT Plus or Pro sign in, or a local model through Ollama or LM Studio, and nothing is ever billed by Synthetic Sciences.

Ace is the paid layer: a pay-as-you-go Wallet for 21 reviewed models from OpenAI, Anthropic, Google, xAI, DeepSeek, Qwen and others, plus managed research search. You add a $20 balance, each request settles at the route’s Wallet rate, and direct routes add no fee. Models served through OpenRouter carry its 5.5% funding fee, already folded into the displayed rate. Card processing is charged separately at checkout.

Auto reload is a separate opt in with a saved card: when purchased funds drop below $5 it adds a fixed $20. You can set a monthly charge cap and a managed usage limit, and turn reloads off at any time. Completed usage and Wallet purchases are non-refundable under the terms. The makers say .edu addresses get $5 of Ace credit, but that offer does not appear on the site or in the docs.

Against the field, Claude Code comes with Claude Pro at $17/mo billed yearly or $20 monthly, and Max from $100/mo. Deepnote’s Team plan is $39 per editor per month billed yearly with a free tier. Elicit Plus is $11/mo billed yearly. OpenScience’s model spend is comparable to a provider bill, and the harness costs nothing, so a researcher with existing keys pays only for tokens.

Who is OpenScience Best For?

Use OpenScience if you are a computational biologist, chemist, physicist or ML researcher who already lives in Python or R, has a cluster or a Modal account, and wants an agent that treats reproducibility as the default. The skills library, database connectors and job dispatch are built for exactly that lab. It is also a good fit for groups that want to self host on their own terms, since the whole thing is Apache 2.0.

Skip OpenScience if your research involves protected health information or unpublished data you cannot risk in a shared trace, at least until you have turned sharing off and confirmed it. The terms explicitly say they do not establish a HIPAA business associate agreement. Skip it too if you want a literature tool rather than a code running agent; Elicit or Consensus are simpler for that job.

Best OpenScience Alternatives

Claude Code

Anthropic’s terminal agent, bundled with Claude Pro at $17/mo billed yearly ($20 monthly) and Max from $100/mo. It is a general coding agent rather than a science harness, but it is the strongest rival on the vendor’s own Terminal-Bench 4.0 science subset at 60%, and Claude Science is its nearest competitor on OpenScience Bench.

Deepnote

A collaborative notebook with AI assistance and managed compute. Free tier for individuals; Team is $39 per editor per month billed yearly and includes $39 of AI credits, $280 of CPU and $50 of GPU each month. Pick it if your team wants a hosted notebook with shared projects instead of a local agent.

Elicit

A literature research assistant covering more than 138 million papers. Basic is free with limited usage; Plus is $11/mo billed yearly ($132) and Pro is $39/mo billed yearly. It does not run code or experiments, but for systematic reviews and evidence tables it is far more focused than OpenScience.

Final Verdict: Is OpenScience Worth It?

OpenScience is the most complete open source science agent we have looked at. The database connectors, the skills library, the cluster and Modal dispatch, the worker verification rules and the Autoresearch loop add up to a tool built by people who understand how a lab actually works. Apache 2.0 licensing and bring your own model support mean the price of trying it is an afternoon.

The reservations are about defaults, not capability. A required account, trace sharing on by default across your own keys, training rights in the terms and Wallet rates hidden inside the client are choices a research tool should have made the other way. Install it, turn sharing off in General settings, use your own provider keys, and it earns its 7/10. Leave the defaults and you are contributing your research to someone else’s benchmark.