A quieter fight than the usual model-benchmark horse race is taking shape in scientific computing, and it is not about which system is smartest. It is about whether the software scientists actually work inside is open or owned. This month that contest got a concrete new entry: OpenScience, an Apache-2.0, model-agnostic AI workbench from a startup called Synthetic Sciences, pitched bluntly as the open answer to Anthropic's proprietary Claude Science.
Released on July 5, 2026 and highlighted by the AICell Lab newsletter on July 24, OpenScience is a browser-based workspace backed by a local agent runtime that aims to run the entire research loop end to end: read the literature, form a hypothesis, write and run code, execute experiments on real compute, query scientific databases, and write up the result — all in one continuous session.
What it actually does
The project's own framing is unusually plain-spoken. "Give it a goal," the README says. "It reads the literature, writes and runs code, runs the experiments, and writes up what it found." Installation is a single npm command (`npm install -g @synsci/openscience`, then `openscience`), or `npx synsci` for a no-install run. The command opens a workspace in the browser with a file tree, editor, terminal, session history, and inline rendering for molecules, protein structures, genomes, and plots.
Two numbers anchor the pitch. OpenScience ships 250+ editable skills — covering model training (DeepSpeed, PEFT, TRL), evaluation, dataset work, cheminformatics, molecular and clinical biology, LaTeX and figure generation, and cloud compute — and wires in roughly 30 scientific databases as agent-callable tools, including UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, and Semantic Scholar. Beyond a default `research` agent, it offers `biology`, `physics`, and `ml` specialists plus critique and literature-review sub-agents and a read-only plan mode.
The defining design choice is that it is model-agnostic. Requests are routed per model, chosen from a selector in the workspace, so a user can run the same task on Claude, GPT, Gemini, GLM, Kimi, DeepSeek, or a local fine-tune and switch between them without rewriting anything. Crucially, the keys stay local: "Your keys stay on your machine and requests go straight to the provider," per the documentation. No account is required, and bring-your-own-key usage is free and never gated. An optional managed layer called Atlas offers curated frontier models billed from a prepaid wallet, a persistent research graph, and cloud compute — but the project stresses OpenScience "works with Atlas but never requires it."
The open-versus-closed framing
Synthetic Sciences is explicit about its target. The research team frames OpenScience as an open alternative to Anthropic's Claude Science, which launched in late June 2026, and the argument is ideological as much as technical: scientific AI tooling should not be owned by one vendor. OpenScience keeps the workflow open, the models swappable, and the data local. (The project also notes it is independent and not affiliated with or endorsed by Anthropic; "Claude" is used only to describe compatibility.)
The comparison is genuinely the story. By the company's own accounting, both tools run the full loop, render science inline, and prize reproducibility. The differences are openness and model choice: any provider versus Claude-only; your infrastructure versus lab machines; BYOK-free versus a paid Claude subscription; and 250+ editable skills versus what Synthetic Sciences characterizes as Claude Science's roughly 60 curated skills and connectors. Claude Science, for its part, taps NVIDIA BioNeMo models (Evo 2, Boltz-2, OpenFold3) that OpenScience does not bundle — a reminder that "open" and "most capable on a given task" are not always the same thing.
Analysis: a stack up for grabs
OpenScience lands in a moment when the open path in AI-for-science is looking less like idealism and more like a viable strategy. The AICell Lab newsletter, which flagged the release, situates it alongside Stanford's Biomni — described as the year's most capable general-purpose biomedical agent, shipped fully open in Science — and a wave of open image-analysis agents. The through-line is data sovereignty (records never leave the institute), no vendor lock-in, reproducibility by default, and access for labs without large budgets. For a computational biologist screening protein targets or an ML engineer testing a fine-tuning idea, the ability to keep private datasets on local systems is not a nicety; it is often a compliance requirement.
The caveats are real, and OpenScience does not hide them. The agent is not sandboxed — the permission system keeps users aware of what it is doing but is explicitly "not an isolation boundary," so the docs recommend running inside a container or VM. It is a young project: the GitHub repository showed just three releases, a couple dozen stars, and a handful of contributors at the time of writing, so buyers should expect rough edges next to a polished commercial product. And because it is bring-your-own-key, users own their provider costs, rate limits, and — importantly — the quality tradeoff, since output depends heavily on which model each request is routed to.
A note on sourcing: the most detailed accounts here come from the project's own GitHub README and a MarkTechPost write-up, both of which lean on Synthetic Sciences' framing, plus the AICell Lab digest. Independent, hands-on benchmarks against Claude Science were not yet available, so the head-to-head claims should be read as the vendor's, not a neutral evaluation's.
What to watch
The smartest model of any given month will change. Whether scientists can own, inspect, and run the tools they depend on is a more durable question — and the answer will hinge on adoption, not manifestos. Watch three things: whether OpenScience's contributor base and skill library grow (the editability is only an advantage if a community actually edits), whether independent labs publish reproducible results built on it, and how Anthropic and Google respond as open workbenches begin competing directly with their curated, closed research assistants. If the open side keeps attracting frontier-quality work, the scientific AI stack really is up for grabs.
"Give it a goal. It reads the literature, writes and runs code, runs the experiments, and writes up what it found."— OpenScience project README, Synthetic Sciences