Perplexity

Overview

Perplexity is an AI search company best known here for its search infrastructure, agentic products, and the launch of Perplexity Decider / Decisions API. Across recent mentions, the company appears as both a consumer AI brand and a platform provider: it operates search products, builds retrieval and ranking systems such as Photon, publishes SDKs and APIs, and open-sources decision models like `pplx-decider-27b` and `pplx-decider-v1-27b`. For AI Product Managers, Perplexity matters because it sits at the intersection of search, agents, retrieval quality, latency engineering, and low-cost decision inference.

The recent signal around Perplexity is especially relevant for PMs building AI features that need fresh web information, grounded answers, or fast action selection. The company’s updates highlight practical product levers: sub-250ms search latency claims for its Search API, aggressive token pricing for decision models, open-source contributions, and research on improving computer-use agents from real user sessions. That combination makes Perplexity a useful reference point for PMs evaluating search-backed assistants, agent APIs, and cost/performance trade-offs in production AI systems.

Key Developments

• 2026-08-29: Aravind Srinivas shared that Decagon powers customer support for Delta Airlines, Ticketmaster, Deutsche Telekom, and American Airlines, with Perplexity search supplying online information for support-related questions.
• 2026-09-05: Aravind Srinivas shared a deep dive on how Perplexity serves search results at scale, covering embeddings for ranking, GPU-based model inference, batching, inference servers, and latency/throughput trade-offs.
• 2026-09-17: Aravind Srinivas announced a `pplx-search-sdk` coding-agents cookbook, including patterns for parallel documentation searches, official-source filtering, and detailed snippet extraction.
• 2026-09-23: Srinivas shared research on training the Perplexity Computer agent from real user sessions using rejection sampling fine-tuning and hint-guided self-distillation.
• 2026-09-25: Srinivas said Perplexity’s Search API, powered by the Rust-based retrieval and ranking engine Photon, achieves sub-250ms p95 latency and ranks highly on relevance and accuracy benchmarks; he also said Photon was built by a small team using hundreds of auto-research loops powered by a more capable internal version of Perplexity Computer.
• 2026-10-02: Srinivas announced that `pplx-decider-27b`, described as a multimodal Decision Model, is being open sourced and exposed through a new Decisions API at 4 cents per million input tokens with free output tokens.
• 2026-10-03: Srinivas recapped seven recent open-source contributions from Perplexity across models, benchmarks, and tools, including `pplx-decider-v1-27b`, which he described as a state-of-the-art multimodal decision model.
• 2026-10-05: A one-shot Pokémon FireRed Elite Four and Champion run was demonstrated using Perplexity’s Decisions API; across 137 live API calls, reported median latency was 592 ms, p95 was 987 ms, 96.4% of responses were under one second, and estimated inference cost was $0.028.
• 2026-10-07: Srinivas announced that Perplexity cut Decision API pricing in half, to 2 cents per million input tokens.
• 2026-10-09: Srinivas shared that Perplexity Decider ranked first on DecisionBench, with quoted results of 93.9% accuracy and the lowest cost.

Relevance to AI PMs

1. Evaluating search-backed product experiences
Perplexity is relevant if you are choosing between plain LLM answers and retrieval-backed answers. Its Search API and Photon-related claims give PMs a benchmark for what “fast enough” and “relevant enough” can look like in production search-assisted AI.

2. Designing low-cost decision or agent layers
The Decisions API and open-source decider models are useful reference points for products that need rapid action selection rather than long-form generation. PMs can use these updates to think about routing, fallback logic, and whether a specialized decision model can reduce cost and latency.

3. Operationalizing agent performance improvements
Perplexity’s work on Perplexity Computer and post-training from real user sessions is relevant to PMs building browser, workflow, or desktop agents. It suggests a concrete product learning loop: capture real usage, generate preference signals, and convert them into better task execution over time.

Related

• Aravind Srinivas: The CEO and most visible spokesperson in the supplied mentions; many of the announcements about Perplexity’s search, SDKs, and decision models come through him.
• Perplexity Computer / Perplexity Computer Agent: The company’s agentic/computer-use efforts, referenced in research on learning from user sessions and in internal tooling used to optimize Photon.
• Perplexity Decider / Decisions API / `pplx-decider-27b` / `pplx-decider-v1-27b`: Perplexity’s decision-model stack, positioned around multimodal decision-making, open source releases, benchmark performance, and aggressive pricing.
• Photon: Perplexity’s Rust-based retrieval and ranking engine powering the Search API, central to its low-latency search claims.
• Search API / `pplx-search-sdk` / `search-sdk`: Developer-facing search tools and SDKs that make Perplexity relevant beyond its consumer app.
• Decagon: An enterprise customer support platform that, according to the mention, uses Perplexity search for support-related online information.
• OpenAI, Anthropic, Google, Meta, Microsoft, Amazon, Apple, Mistral AI, xAI/Grok, Gemini, Claude, ChatGPT: Adjacent AI ecosystem entities that form the competitive and integration context in which PMs may compare Perplexity’s search, agent, and model offerings.
• Wikimedia Foundation, Reddit, Samsung: Related entities in the broader information, distribution, and ecosystem landscape referenced around Perplexity.