Summary: Exa Snapshot, released September 18, 2026, lets users query webpages as they existed on a chosen past date using a snapshotAsOf parameter, backed by more than 400 billion snapshots spanning two decades. The intended uses include preventing answer leakage in agent evaluation and backtesting web-derived signals. The feature remains a research preview, and coverage and rate-limit details require contacting Exa.
Write a batch of tasks in June, then run the eval in September, and you walk straight into a hidden failure mode: the answers may already live on the web. The solution could sit in a paper, a GitHub PR, or a blog post, and an agent with search access can lift it rather than solve anything. Graders cannot tell the two apart, and some may even praise the copier for being efficient.
Exa Snapshot, released September 18, 2026, targets exactly this problem: it lets you search the web the way it stood on a date in the past. Per the Exa announcement, more than 400 billion webpage snapshots back the capability, reaching across two decades. Pass a snapshotAsOf parameter, and your results come from that moment in time—not from the web as it is today.
Why the passage of time reshapes evaluation
Blocking web leakage in reinforcement learning and agent evaluations is the central use case. Train on tasks authored in June, run training in September, and by that point some of the solutions will have found their way online. Agents equipped with search tools locate those answers and submit them, so telling genuine problem-solving apart from retrieval becomes impossible.
Snapshot lets you pin the search date to May—earlier than the publication of those solutions. The agent then sees only whatever existed at that time, which makes a correct answer far more likely to signal real capability. Reproducible evaluation is what’s at stake, and the same concern surfaces when choosing a web search API for agents, where retrieval quality and temporal control both feed directly into agent performance.
Backtesting financial signals minus the months of data collection
Quantitative researchers run into much the same wall. A backtest asks how a prediction model would have performed on a given day, drawing only on information available as of that day. Stock prices and fundamentals come with point-in-time datasets; the web has no comparable equivalent. Testing a trading signal built from web page insights would ordinarily demand months of careful data collection.
Snapshot removes that barrier: backtesting against a versioned web becomes possible immediately. The announcement frames it as a way to test trading signals while skipping the usual data engineering overhead.
How to run a query against the past
Snapshot operates through the /search and /contents API endpoints that already exist. You supply a snapshot_as_of timestamp in ISO 8601 format. Here is a search example taken from the announcement:
from exa_py import Exa
exa = Exa()
result = exa.search(
"latest stable Python release notes",
num_results=3,
contents={
"snapshot_as_of": "2026-05-01T00:00:00Z",
"highlights": True,
},
)
for r in result.results:
print(r.title, r.url)A single URL can likewise be fetched as it appeared on a past date:
result = exa.get_contents(
["https://docs.python.org/3/whatsnew/changelog.html"],
snapshot_as_of="2026-05-01T00:00:00Z",
text=True,
)
print(result.results[0].text[:300])Development is ongoing, and the feature remains a research preview. According to the announcement, details on full index coverage, rate limits, and zero data retention options are only available by contacting the Exa team.
What this means for the eval pipeline you run
Anyone building agents that lean on web search should treat temporal control as a piece of evaluation design. Skip it, and you are measuring retrieval luck as much as reasoning. Snapshot hands you a dial to turn: set the date to before your task’s answer was published, and the signal comes out cleaner.
Coverage is the tradeoff. Two decades is the span, yet the announcement never says how complete the index is for any given date. Recent dates likely enjoy strong coverage, while older ones may have gaps. Test against your own task distribution before you rely on it for high-stakes evals.