# Gemini 4 Carbon Tops WANDR Benchmark At A Cost Lower Than GPT-6 Astra, Claude Fable 5.1 > Source: > Published: 2026-10-31 11:30:15+00:00 Google’s Gemini 4 Carbon continues to top a wide variety of benchmarks. Google’s Gemini 4 Carbon has posted a new state-of-the-art score on Perplexity’s WANDR benchmark, and it’s done so while costing noticeably less per task than two of its closest rivals, GPT-6 Astra and Claude Fable 5.1. The result was shared by Perplexity CEO Aravind Srinivas, who called Gemini 4 Carbon the strongest model Perplexity has evaluated so far for long-form research and answer quality. Srinivas said the model sets a new SoTA on WANDR and will be rolled out to all Perplexity Pro and Max users inside Perplexity Computer once the company has sufficient token capacity to support it. He capped off the announcement with a congratulatory note to Google DeepMind. ## What Is The WANDR Benchmark WANDR, short for Wide ANd Deep Research, is an open benchmark and evaluation harness built by Perplexity to test AI research agents on realistic, high-volume knowledge work. It’s the wide counterpart to Perplexity’s earlier DRACO benchmark, which measures whether an agent can produce a single accurate, complete long-form report. WANDR instead measures whether an agent can build a large, evidence-backed collection of records. The benchmark is built around two demands that have to be satisfied together: discovering a broad, often open-ended set of qualifying entities (the “wide” part), and digging into each one deeply enough to back every claim with evidence (the “deep” part). Tasks are modeled on real research work such as competitive analysis, due diligence, literature reviews, market research, product comparisons and talent sourcing, where a handful of good examples isn’t enough and a polished answer built on incomplete research still fails. Rather than scoring against a fixed answer key, which goes stale quickly for research questions whose answers change over time, WANDR grades every claim an agent makes against the evidence it actually cites, re-fetching cited pages to verify them. Models are scored on Soft F1 and Hard F1, with Hard F1 requiring every component of a result to be fully correct and complete. Even the best-performing systems in Perplexity’s own evaluations have landed well below full marks, underscoring how far wide-and-deep research still has to go before it’s considered solved. ## Gemini 4 Carbon’s Scores On the WANDR leaderboard, Gemini 4 Carbon posted a score of 0.857 at a cost of $5.86 per task, the highest score among all models tested and comfortably ahead of the next closest competitor, Gemini 4 Argon, which scored 0.802 at $7.88 per task. The gap becomes more pronounced against the other major labs’ flagship models. GPT-6 Astra scored 0.735 but at a cost of $13.34 per task, more than double what Gemini 4 Carbon costs to run. Claude Fable 5.1 trailed further behind at 0.601, at $12.76 per task. Grok 4.7 scored 0.528 at $9.30 per task. Further down the cost curve, GPT-6.1 Sol scored 0.654 at $3.70, Claude Opus 5.5 scored 0.647 at $3.69, GPT-6 Sol scored 0.566 at $2.23, and Claude Sonnet 5.5 scored 0.550 at $2.73. DeepSeek V4 Pro 0813 was the cheapest model evaluated at $0.75 per task, but also scored lowest overall at 0.359. Taken together, the results put Gemini 4 Carbon in a league of its own on the score-versus-cost curve, delivering the highest accuracy on WANDR while undercutting the two next-best scoring frontier models, GPT-6 Astra and Claude Fable 5.1, on price by a wide margin. Google has been on something of a hot streak with Gemini 4 Carbon recently, with the model also [topping an updated version of Artificial Analysis’ AA-Omniscience benchmark](https://officechai.com/ai/gemini-4-carbon-tops-updated-aa-omniscience-v1-1-benchmark-with-big-lead-over-competitors/) by a wide margin over its competitors. Its predecessor, Gemini 4 Argon, had also been topping the charts, having recently led on the newly launched [HLE-Diamond benchmark for closed-ended research questions](https://officechai.com/ai/gemini-4-argon-tops-new-hle-diamond-benchmark-on-closed-ended-research-questions/). These back-to-back benchmark wins come at a time when the [performance gap between US and Chinese open models has been widening](https://officechai.com/ai/gap-between-us-and-open-chinese-models-is-now-among-the-largest-in-last-2-years/), with American labs pulling further ahead. For now, Gemini 4 Carbon isn’t broadly available through Perplexity Computer just yet, but with Srinivas confirming it’s coming to all Pro and Max users as soon as token capacity allows, it could soon become the default choice for users running research-heavy workloads on the platform.