Developers Run 50,000 Requests In One File Faster

Perplexity launched a Batch API that lets developers send up to 50,000 questions to its AI search in a single file and collect the answers later, with files kept for 30 days.

Key Takeaways

Perplexity Batch API launched Thursday supporting up to 50,000 requests per file to extend its Agent API into large asynchronous workloads for developers
Perplexity founder Aravind Srinivas posted the release on X
Each request in a batch can run the same full agent loop available in the live API, including web_search and fetch_url
Files are kept for 30 days and the feature is available to every API organization except those on zero data retention

Perplexity founder Aravind Srinivas posted the release on X. The announcement gives operational details but leaves open questions about license terms, model weights, compute requirements, and protocol support that independent builders need to evaluate before adopting the service at scale.

The post says each request in a batch can run the same full agent loop available in the live API, including `web_search` and `fetch_url`, and that batch requests do not count toward Agent API request rate limits, though token limits still apply.

That distinction matters because agent workloads often fail on economics before they fail on model quality, live agents can suit interactive customer support or research copilots, while bulk classification, enrichment, evaluation, and back-catalog analysis usually need different cost and latency profiles that batch processing addresses.

Also Read: OpenAI's 10,000 Agents Run 36 Hours Straight With Zero Human Checks

Batch Integration And Flow 50,000 Requests

H2 Perplexity Batch API integration and flow The new flow is simple, developers upload a JSONL file that can contain 50,000 Requests, create a batch from the file ID, poll for status, and download results when the run is complete.

Perplexity validates files, runs jobs asynchronously, and returns a result file when the batch finishes, mirroring the pattern many teams already use for offline inference while applying it to agents that can browse and fetch.

Perplexity says files are kept for 30 days and that the feature is available to every API organization except those on zero data retention. That carveout is a practical limit for enterprises with stricter compliance needs, because retention rules often decide whether a batch system can be used at all.

Perplexity also says the Batch and Files endpoints follow the shape of OpenAI’s Batch API, which lowers integration friction for teams that already built tooling around JSONL jobs, file uploads, and polling loops.

For developers, the core appeal is operational, packaging many agent calls into one uploaded file that supports 50,000 Requests avoids opening 50,000 separate live requests from an app server, and the 50,000-request cap gives data teams a concrete scale target rather than a vague enterprise promise.

Batch systems matter when the job is large, repeatable, and not urgent, product catalog cleanup, lead scoring, web extraction, QA sweeps, or running thousands of research prompts against the same internal process, because throughput, retries, and predictable billing matter more than latency.

Perplexity’s design also hints at a competitive layer in agent platforms, if search, retrieval, and URL access are core tools, vendors that make those tools cheaper to run in bulk can win workloads that never appear in a consumer app.

Operational And Licensing Considerations

Independent builders should read the fine print before shifting significant workloads to Batch.

The announcement leaves open several operational questions that determine real-world feasibility, license terms for using Perplexity’s tools, the model weights and compute footprint that drive cost, and which protocols and integrations (for example, specific retrieval or browsing behaviors) are supported in large runs.

Those are the metrics that decide whether a given pipeline is cheaper and more reliable as a batch job versus a scaled fleet of live agents. Evaluating the Perplexity Batch API requires understanding retention constraints, model access conditions, and the compute profile behind bulk runs.

This release follows a broader shift in AI infrastructure during October, as labs and tool vendors moved from single-turn chat products toward systems that can perform many small tasks in sequence.

The center of gravity is no longer only the answer on screen but the workflow behind it.

For teams focused on web-heavy workloads, a batch job lets browsing and fetching happen in the background, better for audits, enrichment pipelines, or overnight research, while the retention rule and zero-retention exclusion determine which organizations can actually adopt the feature.