SIAgentNews.com
← All news

Google Gemini 4 Carbon enters Perplexity with lower costs and high research scores

The new model delivers better performance than Anthropic's Fable 5.2 on research benchmarks while costing a fraction of the price per task.

· First reported by OfficeChai

$5.86Cost per task for Gemini 4 Carbon

The short version

  • Gemini 4 Carbon scores 0.857 on the WANDR benchmark, slightly beating Anthropic's Fable 5.2 score of 0.851.
  • The model costs $5.86 per task, which is roughly 40% of the $14.51 cost for Fable 5.2.
  • Perplexity CEO Aravind Srinivas says the model is 2.5x cheaper than Fable 5.2 for similar agentic research capabilities.
  • Subscribers can now use the model for persistent Teammates or one-off sessions in Perplexity's Computer product.

Why it matters

For users and businesses, this shift highlights a move toward 'agentic' AI where cost-efficiency is as important as raw intelligence. It makes it more economically viable to run AI agents that perform high-volume research tasks continuously rather than just answering single questions.

A new benchmark for research

Perplexity recently released its WANDR benchmark to evaluate how AI agents handle large-scale research work. Unlike simple question-answering tests, WANDR measures a model's ability to pull together broad, evidence-backed information from across the internet. It is designed to simulate the high-volume knowledge work that AI agents perform in real-world settings.

The benchmark is distinct from Perplexity's earlier DRACO test, which focuses on the depth of a single long-form report. WANDR instead looks at the breadth of an agent's performance across many research-heavy tasks. This distinction is important because it aligns more closely with how users interact with persistent AI teammates.

The WANDR results show that Gemini 4 Carbon is currently a leader in this specific category. While other models might lead on general intelligence indices, Carbon is positioned as a high-performer for complex, multi-step research. This makes it a primary tool for users who need to synthesize large amounts of data quickly.

The economics of AI agents

The primary headline regarding Gemini 4 Carbon is its cost-to-performance ratio. According to data shared by Perplexity, the model costs $5.86 per task on the WANDR benchmark. In comparison, Anthropic's Fable 5.2 costs $14.51 for a nearly identical score of 0.851.

This means that Gemini 4 Carbon delivers a slightly better result than Fable 5.2 while running at roughly 40% of the price. Perplexity CEO Aravind Srinivas noted that Carbon is 2.5x cheaper than Fable 5.2. This pricing gap is significant for companies looking to deploy agents that run continuously rather than occasionally.

The cost curve is a major factor in how AI products are scaled for public use. By lowering the price per task, developers can offer more complex features without making the service unaffordable. For Perplexity, this efficiency allows for more frequent and deeper interactions within their platform's ecosystem.

Availability for subscribers

Perplexity has officially rolled out Gemini 4 Carbon to all Pro and Max subscribers. This move follows a period where the model was only available to a small group for early testing. Now, a broader audience can use it to power persistent Teammates or one-off sessions in the Perplexity Computer product.

Users can now choose to spin up these Teammates to handle ongoing research projects. The integration allows the model to act as an agentic assistant that can perform many tasks over time. This is a step toward more autonomous AI behavior where the model follows through on complex instructions.

The rollout marks a transition from experimental testing to practical application for the subscriber base. By providing access to Carbon, Perplexity is giving its users a tool that balances high-level reasoning with economic feasibility. It allows for a more flexible way to handle heavy research workloads without hitting high price ceilings.

The competitive landscape

The release of Gemini 4 Carbon highlights a fast-moving cycle in the frontier AI market. While Fable 5.2 recently took the top spot on the Artificial Analysis Intelligence index, Carbon is now challenging that lead on specific research benchmarks. This back-and-forth shows how quickly new models can leapfrog previous state-of-the-art results.

Other models also appear on the cost-versus-score plot provided by Perplexity. For example, Gemini 4 Argon scores 0.802 at $7.88 per task, while GPT-6 Astra scores 0.735 at $13.34. These numbers help illustrate where different models sit in terms of both intelligence and price points.

It is important to note that different labs may configure effort and token budgets differently when testing. Because WANDR is a Perplexity-built benchmark, the reported numbers naturally favor the models the company is currently promoting. However, the data provides a clear picture of the current trade-offs between high-end performance and production costs.

What to watch

  • How other labs respond to the cost-efficiency of Gemini 4 Carbon.
  • Whether Fable 5.2 receives an update to improve its WANDR scores.
  • How the cost-per-task metrics change as models become more optimized.

Try it

If you are a Perplexity Pro or Max subscriber, you can now select Gemini 4 Carbon to power a persistent Teammate for a research project.

Sources and links

First reported by OfficeChai. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news