Gemini 4 Carbon for Deep Research: Perplexity's WANDR Claim, Checked

On the morning of October 31, 2026, Perplexity CEO Aravind Srinivas posted a short note about Gemini 4 Carbon. Perplexity had been running early evals, he wrote, and found it the "strongest model that we've seen so far for long-form research and answer quality, with a new SoTA on the Wide-And-Deep-Research (WANDR) benchmark." It will come to all Perplexity Pro and Max users inside Computer "once we have sufficient token capacity."

The post had about 130,500 views by the time we read it, and the replies were split. This article answers the questions that follow: what is WANDR, who made it, what numbers were actually shown, and should a research workflow change because of it?

TL;DR #

What is the WANDR benchmark? #

WANDR stands for Wide ANd Deep Research. Perplexity released it as an open benchmark and harness in July 2026, and the paper (arXiv 2608.14747) followed in August. It targets a specific failure: agents that answer a narrow question well but collapse when asked to collect hundreds of facts across many entities.

The shape of the benchmark, as the paper reports it:

• 500 public tasks and170,495 source-backed records , drawn from de-identified production use cases such as due diligence and talent sourcing.
• Each task asks for many members (median 100) with several records each (median 3), for a median of 245 records per task. Tasks are key hierarchies, for example company, then employee, then URL.
• Reference-free grading. Instead of comparing to a gold answer, a judge re-fetches every URL the agent cited and checks whether the page really supports the claim.
• Soft and hard F1. Soft F1 gives partial credit. Hard F1 counts only complete members. Both come with precision and recall.

That design is a good idea. Research agents routinely cite pages that do not say what the answer claims, and a verifier that re-opens the source catches that. Our own guide on original-source deep research skills argues for the same habit.

Source citation trail checking each research claim against the page it cites

What did the published WANDR results show? #

The paper's full-set table scores whole research systems, not bare models. Soft and hard F1 on the 500 tasks:

On a 45-task subset, Perplexity at its highest effort setting reaches 0.447 soft F1 and 0.224 hard F1. Two things follow.

First, absolute scores are low. Even the best system's hard F1 is 0.133 on the full set, so WANDR is far from saturated, and a "new SoTA" can be a small move on a tiny number.

Second, the "Gemini" row in the paper is an earlier Gemini research setup scoring the lowest, not Carbon. A system's score depends heavily on the search harness wrapped around the model. That is why the same weights can rank very differently in different products. A Carbon result inside Perplexity's harness tells you about Perplexity plus Carbon.

Why the claim needs a pinch of salt #

Perplexity is the author of the benchmark, the harness and the product being scored. That does not make the claim false. The grader is open, the tasks are public and the paper is detailed. It does mean the claim is vendor-run on a vendor-built test, and the post shows no scores, no harness settings, no cost and no comparison table.

Replies to the post reflected that. One reader called it "benchmaxed" with "no release date." Another noted that a new state of the art on a benchmark called WANDR is "very on brand for research models" and said they were curious how it does outside evals. A third said their own experience put Argon below Sol 6.1 and Opus 5.5 and they would wait to try Carbon. We quote these as reader reactions, not evidence either way.

There is also an access puzzle. Google says Carbon went first to a subset of cyber defenders and that broader access runs through the Fairwind Program, with no date for developers. Perplexity's post says it ran early evals but does not say how it got access, and we have not seen Google or Perplexity explain it. Until one of them does, the practical point is simple: you cannot reproduce the result.

What independent evidence points toward research strength? #

Setting WANDR aside, a few third-party results are relevant to research work, though none of them test deep research directly.

• Intelligence Index v5. Artificial Analysis scores Carbon59 , Opus 5.5 (max)58 and GPT-6 Astra (max)57 . A one-point lead is a close race, as we covered in theCarbon vs Opus 5.5 vs Astra vs Grok 4.7 comparison .
• AA-Omniscience. Carbon scores 54 against Opus 5.5's 48, with accuracy at 58% and a hallucination rate of 11%. Research tools live and die on not inventing facts, so this is the most relevant independent line.
• Vals AI. Carbon leads the Finance Index (66.35%) and the Legal Research benchmark (55.77%, first of 87), both document-heavy research tasks. Vals also reports Legal Research runs averaging over 50 minutes per task.
• Agent Arena. Carbon ranks second across 21,900-plus sessions, behind Opus 5.5.

Search source citations with threads linking an answer back to cited cards

The cautions go the other way. Hallucination rose from 4% on Argon to 11% on Carbon, and Artificial Analysis found Carbon guesses more when unsure. A "wide" research task that lists 100 entities multiplies the cost of every confident wrong answer. Read our Carbon vs Argon breakdown for the full trade-off.

What about Perplexity Computer? #

Computer is Perplexity's agent product, and the post says Carbon will arrive there for Pro and Max subscribers once capacity allows. Capacity is the honest constraint: Carbon used about 2.3 times Argon's reasoning tokens in Vals AI's runs, and a deep research task already burns many tool calls. Token capacity is the constraint Perplexity itself names.

If you already use Computer, nothing changes today. For how Perplexity handles local and cloud work, see our post on running Perplexity Computer on a Mac.

How to test a research claim yourself #

You do not need WANDR to evaluate a research model. When Carbon, or any new model, reaches your account:

1. Pick ten real questions your team asked in the last month, including two where the answer is "not publicly known."
2. Require citations for every claim and open each cited page. Count claims the page does not support. That is the same idea as WANDR's re-fetch check.
3. Score recall on a known list. Ask for all members of a set you can verify, such as the vendors on a published directory, and measure how many it finds.
4. Log cost and time per task. A better answer that takes 50 minutes and many more tokens may not fit your product.
5. Run it side by side against Claude Opus 5.5 and GPT-6 Astra with the same search tools.

Should you change anything today? #

No. Carbon is not callable by most people, Perplexity has not shipped it, and the claim has no published scores. Keep research agents on a live model, keep a source-verification step, and add Carbon to your bake-off list for when access opens. If you run research agents with tool access, a guardrail layer such as AgentBeam from the explainx.ai team can stop an agent before it takes a dangerous action, which matters more as agents browse more of the open web.

What to watch next #

Watch for three things. Perplexity or Google publishing the actual WANDR numbers, with harness settings and cost. The date Carbon reaches Perplexity Pro and Max. And independent deep-research tests from groups that do not own the benchmark.

Related reading #

Details are accurate as of October 31, 2026 and come from Perplexity's CEO post, the WANDR paper, Google's announcement and Artificial Analysis, Vals AI and Arena as reported that week. They may change as Carbon access widens.