The most honest number Perplexity published this week is the one that makes its own product look worst.
Search as Code, the company's search-orchestration agent, topped the WANDR leaderboard at 0.363 soft F1 and 0.133 hard F1. That hard number means the winning system earned full credit for roughly one in seven of the records it submitted and one in seven the task actually required.
And it won a benchmark it designed, wrote the grader for, and published first.
Hold both of those facts in your head at once, because the tension between them is the whole story. A vendor grading its own contest declared itself the winner—and the winning score is a number most sales decks would bury.
A vendor grading its own contest declared itself the winner—and the winning score is a number most sales decks would bury.
WANDR extends what may be a broader pattern into a different domain—research and knowledge work—with an open evaluation harness. And it arguably does not reveal a slope where models slowly improve. It arguably reveals a cliff.
Start with why WANDR's method is categorically harder than what came before.
Most benchmarks grade against a gold answer key. WANDR does not. Instead it re-fetches every cited page and checks each claim against the underlying evidence the agent actually pointed to.
This matters more than it sounds. A gold-answer test can only ask whether an agent arrived at the right conclusion. WANDR asks whether the agent's stated evidence actually supports the claim it attached to that evidence.
That difference isolates a failure mode that gold-answer grading is structurally blind to: the confident citation that does not say what the agent claims it says. In knowledge work, that is the failure that costs you.
The tasks themselves are built from de-identified production use cases—competitive research, due diligence, literature review, talent sourcing—the exact work buyers are being sold agents to do. There are 500 of them, requiring 170,495 source-backed records across three tiers of difficulty.
WANDR splits the problem into two axes. Wide research means discovering a large, open-ended set of qualifying entities. Deep research means backing every one of those entities with specific, checkable evidence.
A single task might ask an agent to find 70 companies with a named executive appointment, then attach a sourced page proving the appointment and a separate page proving the company's listing status. Every branch of that structure is graded independently against the page and excerpt the agent cited.
Soft scores give partial credit. Hard scores require every component of a record to be complete and correct. The gap between Search as Code's 0.363 soft and 0.133 hard is the gap between "roughly right" and "verifiably complete"—and it is enormous.
Now turn to the scaling data, because this is where the piece stops being a leaderboard recap and becomes a diagnosis.
The failure is not a plateau. It is a collapse.
Perplexity's hard precision fell from 0.235 on the smallest task bin to 0.096 on the largest. Hard recall fell from 0.219 to 0.079. As tasks widen, the agent's grip loosens on both fronts—it finds fewer of the required entities and verifies fewer of the ones it finds.
Depth is worse than width. Going from no intermediate steps in the task hierarchy to three or more, Perplexity's hard precision dropped from 0.392 to 0.019.
Read that again. Add a few layers of required intermediate reasoning, and the winning system's verified precision falls by a factor of twenty, to effectively nothing.
That is not a compute problem you can throw tokens at. Cost across all systems spanned from $0.03 to $324.83 per task, and Anthropic's second-place system spent far more time, money, and tokens per task to get closer on raw quality. Spending harder did not buy structural reliability.
The shape of the curve is the finding. Wide-and-deep research does not degrade gracefully as you scale it—it breaks, and it breaks fastest exactly where the task hierarchy deepens. This is the cliff, not the slope.
Which reframes what a leaderboard position is worth. If the delta between first and last place is 0.363 versus a reported top-out of 0.121 soft F1 for the other systems, and everyone is failing structurally at scale, then the ranking is mostly noise. The value of WANDR is the diagnostic breakdown—whether a failure came from insufficient discovery, weak enrichment, or excerpts that do not support their claims.
Now the tell. In the same window that Perplexity published proof its category is not ready, it also announced SPACE.
SPACE is the sandbox platform behind Perplexity Computer, and it is a bet on a specific thesis: that traditional sandboxes were built for short-lived code execution, while agents need to run code, edit files, and run for hours or days without leaving credentials inside the environment.
The architecture separates the session from the sandbox running it. Each task gets a disposable Firecracker microVM that is destroyed when the work ends, while rolling snapshots preserve live memory and files so a session can pause, resume, or branch across sandboxes.
The numbers are the point. On identical production traffic, SPACE cut median sandbox creation latency from 185ms to 60ms, and P90 from 447ms to 89ms. Last week it handled millions of sandbox creations and tens of millions of reconnects, and it has carried 100% of Computer production traffic since June.
Sit with what this pairing signals.
The company documenting its category's inadequacy is simultaneously pouring engineering into the runtime that makes long-horizon agents operable at all. If Perplexity believed the base model was the constraint, it would arguably be publishing model improvements, not microVM latency curves.
It arguably signals that the frontier has moved. Not to a better model—to two layers a model lab does not own: honest evaluation and runtime orchestration.
Those layers are where a search company, not a model lab, is setting the terms. WANDR's re-fetch-and-verify method reflects how a search company thinks about evidence. SPACE's pause-resume-branch architecture reflects how a company running millions of live sessions thinks about state.
And here is what could be read as a strategic move most readers will miss. Publishing WANDR may not be humility for its own sake. It could be read as an argument to buyers that leaderboard positions are noise and the durable advantage lives elsewhere.
If you are buying a deep-research or agent API to build large, evidence-backed lists—not single-answer lookups—the actionable takeaway is not "buy the top of the leaderboard." It is run your candidate vendors against WANDR's published tasks yourself, because the winning score still means six of seven records fail full verification.
Two caveats keep this honest. WANDR is the only independently repeatable claim here, and it is Perplexity's account of a contest Perplexity ran; the transparency about Anthropic's cost-quality tradeoff is real, but the framing still favors the host. And a benchmark that doubles as a semi-automated factory for training data is a benchmark whose author has an interest in what it measures.
None of that softens the core number. The best agent on the market, graded by the most honest method yet published, verifiably completes one record in seven and falls to near-zero precision the moment tasks get deep.
The winning layer, Perplexity's actions arguably suggest, was never the model. It is the willingness to measure the failure precisely—and the infrastructure to keep an agent alive long enough to have a chance at fixing it.