Access is limited for now. We're sharing ATLAS with a small group first, so we're approving only a small number of requests and may not be able to reply to each one. We plan to open access more widely over time, and the grader code is public in the meantime. Thank you for your interest and your patience.

📝 Blog post
💻 Evaluation code
🐛 Report an issue

Benchmark

ATLAS (Agentic Tasks for Large Aggregation + Search) is a benchmark designed to grade the accuracy and completeness of agent responses to challenging real-world workflows that heavily utilize web search. It pairs queries grounded in real search demand with verified golden answers, using an automated pipeline that lets us refresh both as the web and models evolve.

Each task asks a system to discover every entity (e.g. a person, place or company) that meets precise conditions, and then to enrich each entity with 2–10 multi-hop attributes. Tasks are generated from seed topics drawn from clustering of anonymized search demand.

A committee of frontier models and search providers spends a median of 8 agent-hours and 1,200 searches per task to find answers, independently verify them against sources, and resolve disagreements or ambiguities. Blind judges then research every disagreement between the golden answers and twelve search systems, and Exa engineers manually adjudicate anything still in dispute. We estimate that under 1% of gold values are incorrect (0.9%, 95% CI 0.6–1.4%).

The challenge ATLAS poses is efficient search: solve each task in under a minute and for under $1, a budget at which no current system reaches even 0.5 row F1.

Example

Here is what an ATLAS task looks like. So that no real task appears online in plain text, this example comes from a separate test run of our pipeline and is not one of the 547 tasks.

Question (abridged; the full task also defines every column and the edge cases):

Which Legend of Zelda video games have been adapted into a reworked version (a product marketed or built as a "remake," "remaster," "HD," "3D," or "definitive/anniversary edition") that actually went on sale on or before 11 October 2026, published and sold by Nintendo for the Nintendo 3DS, Wii U, Nintendo Switch or Nintendo Switch 2? …

Golden table (4 of its 14 rows):

Zelda remake/remaster product title	Critic Metascore	Lead developer founding year	Lead developer headquarters city	Original game's U.S. age rating	Console Japan launch MSRP (JPY)	Hero voice actor birth year	Hero voice actor home prefecture
The Legend of Zelda: Ocarina of Time 3D	94	2006	Tokyo	E	25000	1967	Hiroshima
The Legend of Zelda: The Wind Waker HD	90	1889	Kyoto	E	26250	1973	Kyoto
The Legend of Zelda: Twilight Princess HD	86	1994	Melbourne	T	26250	1973	Tokyo
The Legend of Zelda: Skyward Sword HD	81	1994	Melbourne	E10+	32378	1978	Kanagawa

Dataset Description

547 examples, each with:

id: the task id.
question: the task, including the columns to return and the answer format (encrypted).
gold: the golden table, plus the grading rule for each column (encrypted).
canary: the canary string, which is also the decryption key.

As with BrowseComp, the questions and golden tables are encrypted so they never appear online as plain text. To read them:

import base64, hashlib, json
from datasets import load_dataset

def decrypt(text, canary):
    data = base64.b64decode(text)
    digest = hashlib.sha256(canary.encode()).digest()
    key = (digest * (len(data) // len(digest) + 1))[: len(data)]
    return bytes(a ^ b for a, b in zip(data, key)).decode()

atlas = load_dataset("exa-ai/ATLAS", split="test")
task = atlas[0]
question = decrypt(task["question"], task["canary"])
gold = json.loads(decrypt(task["gold"], task["canary"]))

Evaluation

The grader is in exa-labs/benchmarks. To run it:

uv run bench run --suite atlas --system scout-exa-auto-highlights

We grade returned tables against golden answers using three F1 metrics, computed per task and averaged over tasks:

Discovery F1 counts just row discovery: how many returned rows name a gold entity, and how many gold entities are returned.
Item F1 counts cells, including both discovery and enrichment.
Row F1 counts rows, where a row is correct only when every cell in the row is. This is our headline metric.

The grader aligns returned rows to the golden table by entity name and aliases, falling back to an LLM aligner (GPT-6 Luna) on misses; different versions of a model, product or software count as different entities. It then scores each cell deterministically where possible, with an LLM judge (GPT-6 Luna) adjudicating the remaining 4.5% of cells. Blank cells, where our agents could show that no data exists, earn credit only when left empty, while ungraded cells, which could not be settled, are excluded. Grading one system on all 547 tasks costs ~$0.24.

Limitations

ATLAS is synthetically generated: frontier-model agents built and verified every task and golden table, and we estimate that under 1% of gold values are incorrect. The golden tables reflect the web as of late September 2026. Facts change, so some cells will go out of date.

If you find a clear error, please open an issue with the task id, the row and column, and a link to a source that shows the correct value. Please don't paste the question or the golden table into the issue. We fix confirmed errors in tagged revisions and plan to regenerate ATLAS from fresh demand every 3–6 months.

Citation

@online{exa2026atlas,
  author = {Alexander Goldberg and Joshua Ahn and Scott Langille},
  title  = {{ATLAS: Evaluating Agents on Search-Intensive Tasks}},
  date   = {2026-10-08},
  url    = {https://exa.ai/blog/atlas-benchmark}
}

Downloads last month: 211

Total file size:

10.6 MB