Gemini 4 Carbon scores 0.857 on WANDR, and commenters zero in on $5.86 first

Perplexity CEO Aravind Srinivas posted a set of early evaluations on X: Gemini 4 Carbon scores 0.857 on the WANDR benchmark, making it the highest-scoring model so far. According to him, the model performs best on long-form research and answer quality, and will later come to the Computer feature for Perplexity Pro and Max users, provided there is sufficient token capacity. He congratulated Google DeepMind.

WANDR stands for Wide-And-Deep-Research; it measures a model’s ability to conduct broad, multi-step research and provide an answer. It is not the same as everyday chat, nor the same as writing code. For ordinary users, it can be understood as a task that requires looking up a lot of material and then writing a report.

In the chart, the horizontal axis is per-task cost, and the vertical axis is WANDR score. Carbon scores 0.857 at a cost of $5.86; Gemini 4 Argon scores 0.802 at $7.88; GPT-6 Astra scores 0.735 at $13.34; Claude Fable 5.1 scores 0.601 at $12.76. On the cheaper side are GPT-6 Sol at 0.566/$2.23, Claude Opus 5.5 at 0.647/$3.69, and DeepSeek V4 Pro 0813 at 0.359/$0.75. Grok 4.7 is 0.528/$9.30.

From the chart, Carbon is not the cheapest, but within the high-score range, its cost is clearly lower than Astra and Fable 5.1. X user Moriarty spelled this out: Astra gets 0.735 for $13.34, Fable 5.1 gets 0.601 for $12.76, and Carbon gets 0.857 for $5.86.

The comments did not stop at congratulations. Tanish Mehta asked how much of the WANDR improvement comes from the model itself and how much comes from the surrounding search harness. Running the same harness with the previous generation of Gemini as a control would be more convincing. This question hits the key issue in evaluation: the scores of research-oriented products are often determined jointly by the model, retrieval, orchestration, and citation strategy.

Raghav Mehta and Max Petrusenko pressed on what is being scored. Does WANDR evaluate only the final report, or also whether it found the right sources and whether the citations hold up? If only the finished draft is considered, a report that reads smoothly but cites incorrectly could still score well. For long-form research, source quality is harder than fluency and also more important.

Pedro G’s reply was somewhat sarcastic: with a newly named benchmark accompanying every model release, will it say "I didn't find it" instead of confidently citing a 2019 Medium article? Hershal Rao also worried that models hallucinate with extremely high precision and confidence. These are old problems for research-oriented products. A high score does not eliminate hallucination; it only shows good performance on this particular test.

Availability also drew scrutiny. Aravind wrote 'once we have sufficient token capacity'; Chandu asked directly, days or weeks? Harley Lewis Foote noted that Google currently locks the model to cyber defenders and trusted partners, and considered 'soon available to all Pro and Max users' to be overly optimistic. This is a commenter's claim, not an official Perplexity timeline. What is certain is that Aravind did not give a date.

There was also skepticism at the experience level. Mikhail Rogov asked how long a typical WANDR query takes on Carbon, since long research can easily become slow. Gonzalo Ruiz de Villa said that in his experience Argon ranks below Sol 6.1 and Opus 5.5, so he remains skeptical until he tries it himself. ChrisUniverse, by contrast, praised Argon 4’s very low token usage. Taken together, the two kinds of feedback show that benchmark scores and day-to-day feel are not the same thing.

Natural Stupidity’s reply was more like self-deprecation: broad-and-deep research is great; I’m narrow and shallow, but usable today. This line explains why 'coming soon' makes people impatient. Model capability is one layer; whether you can call it reliably, how long you have to wait, and how much one task costs are another.

What is most worth looking at in this chart is that three variables are put side by side: score, cost, and availability. 0.857 is only one point among them. Carbon offers a very strong combination of score and cost, but the benchmark’s construction, the search harness’s contribution, citation quality, latency, and release timing all remain unanswered. Aravind says it will be opened to Pro and Max users, with timing depending on token capacity. The commenters have already asked all the questions for the next round.