Asta Credits
At a glance
Credits are a way to track your usage of Asta's tools. You're never charged for them. But credits stand for real compute and energy; we cover that cost, and we hope it helps you make informed decisions about what to run.
How do credits work?
When you sign up, you receive 1,000 credits as a one-time grant. Using certain tools consumes credits, but if you ever fall below 120 credits, you can request more.
How are credits used?
Different tools consume credits according to the type of work they perform, and some tools do not use credits at all. Each tool defines its own credit calculation based on the work it performs. An experiment in one tool is not equivalent to a source paper in another.
- Find papers, Generate a report, Analyze data: No credit usage
- AutoDiscovery: 1 credit / experiment
- Theorizer: 1 credit / source paper
What happens if I run out of credits?
If your balance falls below 120 credits, you can request 100 more via the button in the credit dialog.
Analyze data
At a glance
- Do not upload sensitive or regulated data to Ai2’s hosted Asta deployment. You must use a self-hosted deployment for datasets subject to HIPAA, GDPR, or similar requirements.
- All uploaded datasets are automatically deleted from within our systems 7 days after upload.
- No model training, sharing, or redistribution of your uploaded datasets.
- No model training on your interactions without explicit opt-in.
- Explorations remain active for 24h, then become read-only. To continue, start a new exploration.
- You can delete your datasets at the exploration level at any time (signed-in users). Deletion is immediate and permanent.
What datasets can I upload for analysis?
You can upload structured research datasets in the following formats: CSV, Excel (.xlsx), JSON (.json/.jsonl), HDF5, TSV, and Parquet. Do not upload sensitive or regulated data (e.g., PHI, financial account numbers, precise geolocation). If your data is governed by regulatory requirements (e.g., subject to HIPAA, GDPR, or similar frameworks), please refer to Question 6 (“What changes in a self-hosted (institution-managed) deployment”) for more details. Self-hosting is currently the only option for using Asta with regulated data.
Key points
- Uploaded files remain associated with your account.
- We never share, sell, or redistribute your uploaded datasets.
- Uploaded datasets are automatically purged 7 days after upload from within our systems.
- If we detect sensitive or regulated data:
- We will attempt to notify you at the email address you provided
- We may remove the affected files to maintain compliance with our Terms of Use.
- ⚠️ Note: You are solely responsible for ensuring your uploaded datasets complies with relevant laws and institutional policies. Ai2 is not liable for any misuse or unauthorized disclosure resulting from prohibited uploads.
Who can see my uploaded datasets?
- We do not share your uploaded datasets with third parties and we do not make them public.
- Limited, least-privilege access by Ai2 system administrators to your datasets may occur for safety, reliability, or legal reasons. For example, to investigate a bug or system issue.
- If you choose to opt in to allow your de-identified interactions with Asta (e.g., prompts, outputs, analysis steps) to be included in a public research dataset, we will use those interactions only for research purposes; i.e., to improve Asta.
- Regardless of your consent setting, your raw uploaded datasets are never redistributed or used for advertising or profiling.
Do you use my uploaded datasets to train AI models?
- Raw uploaded datasets are never used for model training.
- Prompts and interactions (your questions, feedback, and system events) may be temporarily logged to improve reliability and Asta quality, but they are not used to train models without explicit opt-in consent.
How long do explorations live, and what happens after 24 hours?
Explorations are designed for short research sessions:
- An exploration is active for the first 24 hours. During that window you can ask follow-ups.
- After 24 hours, the exploration becomes read-only: you can view results, but you cannot continue the conversation in that exploration. To keep working, start a new exploration.
- If you uploaded multiple files to an exploration, deletion is at the exploration level (you can’t remove individual files from that exploration).
How do I delete my uploaded datasets, and what’s the retention window?
When you’re signed in
- You can view and delete any of your explorations at any time in the hosted Asta interface.
- Deleting an exploration will remove:
- All datasets uploaded into that exploration
- All analyses, results, and outputs generated in that exploration
- Deletion is immediate and permanent. No recovery of deleted data is possible
- Uploaded datasets are automatically purged 7 days after upload if not deleted earlier. You can continue to access exploration summaries and results, but the source files themselves are deleted.
When you’re using Asta anonymously
- Recent explorations may be visible to you via browser cookies, allowing you to revisit and delete them while the cookie is intact.
- However, if you clear your cookies, you will lose access to those explorations (and the ability to delete them), as they are no longer associated with your browser session.
- To manage storage and maintain system hygiene:
- Uploaded datasets are automatically purged 7 days after upload if not deleted earlier.
- Anonymous explorations with no activity for 30 days will be deleted automatically.
- Once deleted, these explorations cannot be recovered.
- Tip: To ensure you retain access to your data and can manage your explorations directly, we recommend signing in.
What changes in a self-hosted (institution-managed) deployment?
Self-hosting gives your organization full control over storage, retention, access policies, and compliance.
- You decide where data lives, who can access it, how long it’s kept, and how it’s deleted at your institution. You are responsible for setting and enforcing these policies, including any applicable legal or compliance requirements.
- Self-hosting is the only recommended path for regulated data (e.g., PHI under HIPAA, controlled datasets under specific data-use agreements, GDPR residency needs).
- In self-hosted deployments, access to uploaded data is controlled entirely by your organization. Ai2 does not have visibility into datasets stored or processed in your environment.
AutoDiscovery
Preview program
At a glance
- Asta Preview is an early-access program for scientific innovators to test upcoming features and experimental prototypes before public release.
- Open to researchers, scientists, grad students, and research-adjacent professionals across all domains willing to provide critical feedback.
- This is utility testing, not just bug fixing. We need to know if these tools work for your real-world research needs.
- Benefits include influencing the product roadmap, direct access to the development team, and potential for future collaboration.
- Light commitment: use Asta in your actual research, be patient with rough edges, and provide honest feedback through occasional surveys.
- Stay connected via dedicated Discord channel or email. You can opt out at any time.
About the program
What is Asta Preview?
Asta Preview is an early-access program designed for scientific innovators. Members get exclusive access to upcoming Asta features (like DataVoyager) before the public, as well as experimental prototypes directly from our research labs (like Paper+Figure QA). We are releasing these "lab-fresh" experiences to solicit feedback and test whether they can truly accelerate your science in real-world scenarios.
Who is this for?
We are looking for "Enthusiast Early Adopters" across all domains, from bioscience and climate research to computer science.
- Roles: Active researchers, scientists, grad students, and research-adjacent professionals (clinicians, journalists, analysts).
- Mindset: You understand that research is iterative. You are willing to test evolving tech and provide the critical feedback needed to shape it.
Why is Ai2 creating this program?
We know that science is not a single problem to be solved. The workflows of a biologist differ vastly from those of a climate scientist, and AI built in isolation often misses that nuance. We are launching Asta Preview to build a direct bridge to the scientific community— - especially those outside of computer science. We need your domain expertise to ensure our tools work for your reality, rather than forcing you to adapt to generic AI models. We want to close the gap between what is technically possible and what is scientifically useful.
Is this just a beta test?
No. While we want to fix errors, this is primarily about utility testing. We are transparently trying to figure out how to make these agents useful. We aren't just looking for broken buttons; we need to know if the AI can actually deliver value for your real-world research needs.
Can I opt out?
Yes. You can leave the program at any time. Just email us at asta-support@allenai.org. If you opt out, you will return to the standard public version of Asta and lose access to the experimental features.
Benefits & expectations
Why join?
Beyond early access, you get a seat at the table:
- Influence the roadmap: Help us validate use cases. If a feature doesn't work for your data format, tell us so we can prioritize a fix.
- Direct access: Join small-group sessions with the engineers and researchers building Asta.
- Competitive edge: Be the first to integrate agentic AI into your literature review and analysis.
- Potential for collaboration: We view our Preview cohort as our primary pool for future partnerships and deeper collaborations.
What are the expectations?
The commitment is light and flexible:
- Use it: Try Asta in your real research. The best way to help us is to use the tools for your actual work.
- Be understanding: These features are still being built. We appreciate your patience if things look rough or don't work perfectly.
- Be honest: We will send short surveys now and then. Please tell us the truth. If a feature isn't useful to you, we need to know.
How do I stay connected?
- Discord: Join our dedicated channel to chat directly with our team and other participants.
- General feedback: We are always trying to figure out how to make Asta more useful. If you have ideas outside of the specific features, or just want to chat about the platform generally, our inbox (and Discord) is always open.
Generate theories
At a glance
- Generate theories turns a research question into candidate theories. Asta finds up to 100 relevant papers, extracts their findings, and proposes testable claims that hold across them.
- It’s not a literature review. For a summary of what papers say, use the Generate report tool.
- Treat theories as candidates, not conclusions. Expand each hypothesis, read its scope, and confirm the evidence matches the paper it cites.
- Supporting and Conflicting numbers count findings, not papers. Zero conflicting can mean narrow evidence, so check the Evidence tab before you trust a count.
- “Novel” is relative to the papers retrieved for your run, not to all of science. It’s a lead to check, not a claim of priority.
- Runs take 20 to 30 minutes and use Asta credits. Each run can take up to 100 credits, the actual credits used depends on the number of papers used in the theory generation. If a run fails, you get back the credits it didn’t use.
What does Generate theories do?
You pose a research question, and Asta finds up to 100 relevant papers. It then extracts findings from those papers and generates candidate theories about the patterns that hold across them, where they hold, and on what evidence.
How is this different from a literature review?
A review summarizes what papers say. Generate theories proposes claims that cut across papers, stated so you could test them. To perform a literature review, we recommend the Generate report tool.
What are theories, hypotheses, and predictions?
- A hypothesis is a candidate pattern drawn from the papers, such as “X increases Y.” Each hypothesis includes a scope (where it should hold), a novelty note, and the evidence behind it.
- A theory groups a few related hypotheses and explains why they belong together.
- A prediction is a result you would expect if the theory holds. Likely outcomes are where the papers already point; unknown outcomes are questions the papers by themselves can’t settle yet.
What do the Supporting and Conflicting numbers mean?
They count findings, not papers, that support or conflict a theory. Zero conflicting can mean the evidence is narrow rather than strong, so review the Evidence tab before you trust a count.
Are the theories true?
Not necessarily. Treat them as candidates. When predictions could be checked against later-published papers, most held (Jansen et al., 2026, p. 5-6). But only about half had a paper to check against, and riskier predictions held far less often. To verify a theory, expand each hypothesis, read its scope, and confirm that the backing evidence matches the paper it cites.
What does “novel” mean here?
Novel is relative to the papers retrieved for your run, not to all of science. Expert reviewers found most outputs read as strong synthesis of existing knowledge rather than discovery. Treat “novel” as a lead to check, not a claim of priority.
How long does a theory exploration take, and what does it cost?
It usually takes 20 to 30 minutes; comprehensive runs can exceed an hour. Each run can take up to 100 credits, the actual credits used depends on the number of papers used in the theory generation. If a run fails, you get back the credits it didn’t use.
Why didn’t it find the paper I expected?
Generate theories only sees open-access full text it can retrieve. Paywalled papers and gray literature are invisible to it. Search sometimes misses papers, even well-known ones.
Can I export or share results?
Yes: you can share a read-only link or export artifacts as Markdown, HTML, JSON, or a summarized PDF report. If you cite a theory, cite the Theorizer paper.
Where do I start if the page is new to me?
Click “Get to know Theorizer” for a six-step tour of the results page.