Perplexity Integrates Anthropic's Fable 5.2, Topping Web Research Benchmarks at $14 per Task

Perplexity has added Anthropic’s Claude Fable 5.2 to its agent-driven products. Pro and Max subscribers can use the model in Perplexity Computer, while Max subscribers also receive access through the core Search product. Anthropic positions Fable 5.2 as an agent-focused model that improves research accuracy, task cost, and generation speed over Fable 5.1.

Perplexity’s headline result comes from WANDR, its benchmark for web research agents. Fable 5.2 scored 0.851 at a reported cost of $14.51 per task, placing it on the benchmark’s Pareto frontier. In this context, that means no measured alternative offered a higher score at the same or lower cost, or a lower cost at the same or higher score. The $14.51 figure represents model spending during the evaluation; subscription and API prices are separate.

The result extends a rapid shift in WANDR’s rankings. In earlier WANDR results, Grok 4.6 matched Fable 5’s 0.496 score while reducing task cost from $20.30 to $7.58. Gemini 4 Carbon subsequently took the top position before Fable 5.2 surpassed it.

WANDR grades the evidence

Perplexity built WANDR, short for Wide ANd Deep Research, to test agents that collect large sets of evidence-backed records from the web. Its earlier DRACO benchmark evaluates a single long-form research report. WANDR focuses on breadth, testing whether an agent can find, structure, and support many results across a larger research task.

WANDR evaluates submitted claims against their cited sources and fetches those pages again during grading. This design accommodates research questions whose answers change as web pages and datasets are updated.

• Soft F1 awards partial credit when a result contains some correct fields.
• Hard F1 awards credit only when every required component is correct and complete.
• Citation verification checks factual claims against the pages the agent supplied as evidence.

The benchmark therefore rewards coverage, field-level accuracy, and source support. Fluent prose alone contributes little when the underlying records are incomplete or the citations fail to substantiate them.

Inside Fable 5.2

Anthropic describes Fable 5.2 in its model release as its highest-capability agent-focused model, launched alongside the smaller Mythos 5.2. Fable supports a one-million-token context window and offers adjustable reasoning effort, allowing applications to trade additional computation for stronger results.

Anthropic says Fable 5.2 at Medium effort matches Opus 5.5’s best WANDR result for about 40% of the cost per task. At Low effort, it reportedly matches Opus 5.5’s top Humanity’s Last Exam score for less than one-tenth of the task cost. Actual spending will vary with prompt length, output volume, reasoning effort, tool calls, and retries.

Anthropic’s evaluation table reports several additional results:

• CVE Rediscovery: 89.6% on a task that tests whether a model can identify known software vulnerabilities.
• Protein design: 85.1% on a 24-hour de novo protein-binder task.
• WANDR: 82.1% under the company’s relaxed scoring configuration, ahead of Gemini 4 Carbon in the cited comparison.
• AutomationBench: 51.4%, trailing Gemini 4 Carbon’s 53.9%.
• Other listed tests: Anthropic reports higher Fable 5.2 scores than GPT-6 Astra across every benchmark in its table.

Perplexity’s 0.851 WANDR result and Anthropic’s 82.1% relaxed result use different labels and reported configurations. A direct comparison requires the underlying run settings, including reasoning effort, scoring mode, tool access, and task-cost calculation.

Vals independently ranks Fable 5.2 at 71.31%, followed by Claude Sonnet 5.5 at 70.85% and Gemini 4 Carbon at 70.45%. Fable’s margin of error is approximately 0.96 percentage points, so the evaluation does not establish a clear statistical lead over those nearby scores. Anthropic’s strongest results also sometimes use Xhigh effort, which consumes more tokens and raises task cost.

A new planner for Computer

Perplexity Computer uses an orchestrator to divide a request into subtasks, choose tools or models, and combine their outputs. Fable 5 already served in that role for workflows such as multi-source research, evidence collection, and report generation. Fable 5.2 can occupy the same position while improving Perplexity’s reported accuracy and cost per completed benchmark task.

Research workflows stand to benefit when they involve many searches, structured records, or repeated source checks. Better orchestration can reduce missed fields, unsupported claims, and unnecessary tool calls. The published speed gain concerns model output generation, so it does not guarantee a 30% reduction in end-to-end workflow time; browsing, external tools, and page retrieval can remain the dominant sources of latency.

Perplexity’s announcement covers access through Computer and Search. It does not provide API availability, model identifiers, rate limits, context limits within Perplexity, or details about whether Fable 5.2 becomes the automatic default for eligible tasks. Developers integrating through Perplexity’s API will need separate documentation before assuming equivalent access.

The leaderboard needs more runs

Anthropic and Perplexity supplied the main benchmark claims, making independent replication important for assessing reliability across different prompts and tool configurations. Vals provides an early external result, though its narrow score gaps and stated uncertainty limit firm model rankings.

WANDR’s citation-based design gives researchers a useful way to inspect failures, but leaderboard placement still depends on scoring configuration, reasoning effort, retrieval tools, and cost accounting. Comparable public runs with those settings disclosed will show whether Fable 5.2’s reported advantage persists beyond the launch evaluations.