Perplexity has made Claude Opus 5.5 the Standard effort level in Perplexity Computer, changing the default coordinating model used for a broad class of everyday agentic work. The company says Opus 5.5 scored 0.610 on its proprietary WANDR benchmark at an average cost of $4.13 per task, slightly outperforming Claude Fable 5.1 while costing 67.6% less per task.
The rollout was confirmed in an official Perplexity company update. The benchmark numbers come directly from Perplexity and should be treated as vendor-reported results rather than independent evaluation, but they explain why the model has been placed in one of Computer’s most consequential default tiers: Standard is designed to balance reasoning quality and cost for normal work rather than maximize either extreme.
The change matters beyond benchmark rankings because Perplexity’s effort settings select the model responsible for orchestrating an assignment. According to the company’s official Effort Mode documentation, the orchestrator decides how to approach the work and delegates parts of the assignment to supporting agents. Changing that orchestrator can therefore change the research path before the final answer, sources or citations are produced, although Perplexity has not published evidence showing that Opus 5.5 specifically improves or alters citation quality in production.
Opus 5.5 is now the Standard effort model
Perplexity Computer exposes four effort presets: Light, Standard, High and Ultra. The user chooses how much effort a task deserves, while Perplexity chooses the orchestrator model and reasoning depth behind the preset.
Standard is positioned for everyday assignments that require a balance between cost and reasoning. Perplexity gives the example of combining account notes and company updates into a sales brief: more involved than a simple clerical task, but not necessarily worth the maximum cost and reasoning allocated to an open-ended strategic investigation.
Claude Opus 5.5 now occupies that Standard slot. Unless a user changes the effort level or uses Custom controls to select another model, Opus 5.5 becomes the coordinating model for tasks routed through the Standard preset.
This is different from merely adding Opus 5.5 to a model picker. A default effort tier determines which model many users will encounter without making an explicit provider-level decision.
Perplexity reports a 0.610 WANDR score at $4.13 per task
Perplexity says Claude Opus 5.5 achieved a score of 0.610 on WANDR at an average cost of $4.13 per task. According to the company, that slightly exceeds Claude Fable 5.1 while reducing per-task cost by 67.6%.
The numbers are useful because Perplexity is explicitly optimizing Computer around the relationship between intelligence and cost. Its effort-mode documentation says the company tests models from different providers at different reasoning levels, comparing work quality with the cost required to produce it.
When a cheaper combination provides comparable performance, Perplexity can move that model into a lower-cost or default tier and preserve user credits for assignments where deeper reasoning creates more value.
WANDR remains Perplexity’s own benchmark, so the 0.610 score and cost comparison should not be generalized into a universal claim that Opus 5.5 is superior to Fable 5.1 across all agent workloads. The result describes performance under Perplexity’s evaluation framework.
A 67.6% cost reduction can matter more than a small benchmark gain
The reported score improvement over Fable 5.1 is described by Perplexity as slight. The larger operational difference is the claimed 67.6% reduction in average cost per task.
Agentic systems can generate substantial inference expense because a single user assignment may trigger planning, multiple supporting agents, web research, tool calls and synthesis. The orchestrator itself can also reason repeatedly as new information arrives.
A model that produces similar or slightly better benchmark performance at roughly one-third of the task cost gives Perplexity more flexibility in how much work it can allocate to Standard assignments without exhausting user credits as quickly.
That economic advantage is particularly relevant to a default tier. Ultra can justify expensive reasoning because users explicitly ask for maximum effort. Standard has to work economically across a much larger set of routine professional tasks.
The orchestrator is the agent brain that decides how research happens
Perplexity’s architecture makes the Standard-model change more consequential than a simple response-model swap. The company says the orchestrator determines how an assignment should be approached and delegates portions of the work to supporting agents.
For research, those planning choices can include how a question is decomposed, which branches deserve investigation, when additional evidence is necessary and how supporting agents should divide the work.
The supporting agents can use models from different providers. That means a Computer task coordinated by Claude Opus 5.5 does not imply that every subtask or every generated token comes from Opus 5.5.
The orchestrator is better understood as the model making the high-level decisions about the workflow. Its influence can therefore extend across a multi-model research process even when other models execute individual steps.
Changing the orchestrator could change source selection
Research agents do not retrieve sources in a vacuum. Before a source appears in a final answer, the system has to decide what to investigate, which queries to issue, whether the evidence is sufficient and whether another search branch should be opened.
A different orchestrator can make different decisions at those stages. One model might decompose a question into several narrow searches while another uses fewer, broader investigations. One might request additional verification when sources conflict, while another may decide the available evidence is sufficient.
Those differences can propagate into the final source set. If the search path changes, the domains retrieved and ultimately cited can change as well.
Perplexity has not published a controlled study demonstrating that the switch to Opus 5.5 changes source diversity or citation accuracy. The possibility follows from the documented role of the orchestrator, but the magnitude and direction of any real-world citation effect remain unmeasured.
AI visibility tracking needs to record the effort tier
This architecture creates a measurement problem for marketers tracking visibility in Perplexity. Recording only that a brand appeared in “Perplexity Computer” is increasingly incomplete because different effort settings can use different coordinating models and reasoning depths.
A Standard task after this rollout can be orchestrated by Opus 5.5, while Light now defaults to GPT-6 Sol. Higher tiers can use different configurations again, and users can override the defaults through Custom controls.
Two visibility tests using the same prompt but different effort settings may therefore travel through different research paths. Even if the final answer looks similar, the retrieval process underneath can differ.
Longitudinal GEO and AI-search studies should capture the date, product surface, effort level and visible model configuration whenever possible. Otherwise, an apparent visibility change may partly reflect a platform-routing change rather than a change in the brand’s underlying web presence.
Perplexity is making model choice an invisible optimization layer
The effort system represents a broader product shift away from asking ordinary users to understand model benchmarks. Instead of choosing between model names, users can choose how much effort and budget a task deserves.
Perplexity then selects the model and reasoning configuration it believes provides the appropriate balance. Advanced users retain Custom controls, but the default experience increasingly hides model selection behind the task-level setting.
This allows Perplexity to replace the model behind Standard without redesigning the interface. The meaning of Standard remains “balanced reasoning and cost,” even if the model that best satisfies that target changes as new frontier releases arrive.
Opus 5.5 is therefore competing not only as an Anthropic model but as a component inside Perplexity’s routing system.
The timing shows how quickly frontier models become infrastructure
Claude Opus 5.5 was released by Anthropic only days before Perplexity moved it into the Standard tier. That short gap illustrates how quickly new frontier models can move from vendor announcement to default infrastructure inside another company’s product.
Perplexity does not need to rebuild Computer around one provider. Its model-agnostic harness is designed to evaluate new models and adjust orchestrator and subagent combinations as the frontier changes.
The company made that strategy explicit when announcing effort controls: it plans to continue refining the model combinations behind each effort level as newer models improve.
For users, this means a familiar product setting can become materially different underneath without requiring a migration or even a conscious model-selection decision.
Standard sits between GPT-6 Sol on Light and higher-effort configurations
Perplexity’s latest changes create a clearer cost-capability ladder. GPT-6 Sol has become the default Light option for straightforward everyday work, while Claude Opus 5.5 now occupies Standard for balanced reasoning and cost.
Higher effort settings can allocate more expensive models and deeper reasoning when the task justifies it. Perplexity’s September 21 changelog also added GPT-6 Astra for eligible Computer users and supporting agents, illustrating the range of models available inside the same orchestration framework.
This is less like choosing a single search model and more like a cloud-compute scheduler choosing the appropriate resource class for the workload.
The user chooses the desired effort. Perplexity decides which model currently provides the economics and capability needed to satisfy that tier.
The benchmark is about task economics, not citation quality
The WANDR figures should not be stretched beyond what Perplexity actually reported. A score of 0.610 at $4.13 per task supports the company’s decision to position Opus 5.5 as an efficient Standard orchestrator under its own benchmark.
It does not independently establish that Opus 5.5 retrieves better sources, cites more accurately or produces more diverse evidence than Fable 5.1 in Perplexity Computer.
Those outcomes require separate measurement. Citation quality depends not only on the orchestrator model but also on Perplexity’s retrieval systems, source-ranking logic, supporting agents and synthesis pipeline.
The model can influence the research plan without controlling every layer that determines which source ultimately appears in the answer.
Default-model changes can become hidden SEO variables
For publishers and brands, model routing adds another source of volatility to AI visibility. Traditional search ranking changes are usually attributed to the search engine’s own systems. In an answer engine built from interchangeable frontier models, visibility can also be affected by changes in the model that plans the research.
A site does not need to change for the agent’s search strategy to change. Perplexity can update the orchestrator, reasoning depth or subagent configuration and produce a different retrieval path from the same web corpus.
This does not mean every model update will cause measurable citation movement. But it means analysts should include platform configuration changes among the hypotheses they test when citation patterns shift suddenly.
The default model is becoming part of the search environment, even when users never see or select it directly.
Perplexity is optimizing the agent stack, not pledging allegiance to one model
Claude Opus 5.5 becoming Standard should not be interpreted as Perplexity standardizing its entire product on Anthropic. The company’s own architecture is deliberately multi-model.
GPT-6 Sol can orchestrate Light, Opus 5.5 can orchestrate Standard, GPT-6 Astra is available for eligible higher-end work, and supporting agents can use other models depending on the assignment.
Perplexity’s competitive layer is increasingly the harness that decides how those models work together. The frontier providers supply intelligence; Perplexity decides where that intelligence is economically useful inside the task.
That makes benchmark cost almost as important as benchmark quality. A model does not need to be the strongest available option to become the most influential default. It needs to provide enough intelligence at the right price for the workload Perplexity expects to run most often.
Opus 5.5 now sits in the path of everyday research by default
The immediate change is straightforward: Claude Opus 5.5 is now the Standard effort level in Perplexity Computer, backed by a Perplexity-reported WANDR score of 0.610 at $4.13 per task and a claimed 67.6% cost reduction versus Claude Fable 5.1.
The broader consequence is architectural. Standard is not simply a quality label; it selects the orchestrator that decides how Computer approaches and delegates everyday work.
That means a model replacement can potentially alter research planning before users ever see the final answer. Different planning can lead to different retrieval branches, and different retrieval branches can produce different sources and citations.
Whether Opus 5.5 measurably changes citation quality still needs evidence. But Perplexity has already changed the model making the default planning decisions for Standard tasks, making the agent brain itself a moving variable in AI search and research measurement.