Perplexity Adds Gemini 4 Carbon Matching Claude's Research Score at 40% the Cost

Perplexity has added Google’s Gemini 4 Carbon to its multi-model agent platform. Perplexity reports a score of 0.857 at $5.86 per task on its WANDR research benchmark, matching Anthropic’s Claude Fable 5.2 at about 40% of the cost. The result gives developers a reason to test Carbon for web research, although production spending will depend on Perplexity’s routing and credit system.

Availability depends on the subscription tier. Pro and Max subscribers can use Carbon in Perplexity Computer, while access through standard Search is limited to Max subscribers.

Carbon’s price-performance pitch

Google describes Carbon as its “next frontier model.” The reported launch results placed it first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, narrowly ahead of Claude Opus 5.5 and GPT-6 Astra. Published workflow benchmarks also ranked Carbon first on CursorBench 4.0, WANDR, and Zapier’s AutomationBench, with less than half the cost per task of the next-best external model.

A third-party coding evaluation cited in the launch material found that Carbon completed tasks through many small steps. At medium reasoning effort, it averaged 56 steps, compared with six for Fable 5.2 and 10 for Opus 5.5. Most of Carbon’s input tokens came from cache, which reduces the price of repeatedly processing the same context. That pattern can lower token costs while increasing wall-clock time and tool calls.

WANDR rewards broad, sourced research

Perplexity created the WANDR benchmark , short for Wide ANd Deep Research, to evaluate agents that collect large sets of evidence-backed records from the web. Its earlier DRACO benchmark evaluates a single long-form research report. WANDR instead measures coverage, factual support, and an agent’s ability to find many verifiable items across multiple sources.

WANDR’s rankings have shifted as new models and results arrived. Carbon initially took the lead, and Fable 5.2 later moved ahead. Perplexity’s latest comparison reports the same 0.857 score for both models, with Carbon delivering the lower benchmark cost. An equal aggregate score does not establish identical behavior, since the models may differ in coverage, citation quality, latency, and failure patterns.

A 19-model router changes the math

Perplexity Computer coordinates 19 models to complete complex, long-running workflows in the background. It decomposes a request, assigns individual stages to selected models, and bills the resulting usage in credits. A single job may invoke Carbon for one stage and other models elsewhere, so Carbon’s $5.86 WANDR result does not represent the total cost of every Computer task.

The published plan figures do not support a direct conversion from the $5.86 benchmark cost to a fixed number of subscriber runs. That calculation requires the credits consumed by a Carbon task, along with any credits used by other models in the workflow. Under WANDR’s benchmark conditions, the available comparison indicates that Carbon matched Fable 5.2 while reducing model cost by about 60%.

What developers should measure

Model selection for agent systems depends on task completion, evidence quality, latency, and cost. Carbon’s WANDR result supports testing it on research-heavy workloads, particularly those that reuse substantial context and can benefit from cached input. Its many-step execution pattern may be less suitable for latency-sensitive applications.

Anthropic reports that Fable 5.2 at medium reasoning effort matches Opus 5.5’s top WANDR score at roughly 40% of the cost, while leading GPT-6 Astra and Carbon on most benchmarks Anthropic publishes. Perplexity’s narrower WANDR comparison gives Carbon the cost advantage at the same reported score, illustrating how model rankings change with the task, evaluator, and configuration.

Teams comparing Carbon with Fable 5.2 or Opus 5.5 should record metrics from representative production tasks:

• Task completion rate and evidence coverage
• Citation accuracy and unsupported claims
• Median and tail latency
• Tool calls and reasoning steps per task
• Input, output, and cached token usage
• Computer credits consumed across all routed models

Numbers still missing

The announcement details cited here do not specify several figures needed for deployment planning:

• Standalone API availability and pricing
• Context-window and output limits
• Rate limits and concurrency quotas
• Computer credits consumed by a typical Carbon task
• Median and tail latency on WANDR
• The source and full configuration of the third-party coding evaluation

Carbon’s practical appeal rests on a specific, testable claim: Perplexity measured Fable-level WANDR performance at about 40% of the model cost. Its production value inside Computer will depend on whether those economics persist after routing, credit consumption, latency, and workload-specific failure rates are included.