LoginSign Up
← Back to all sparks
Perplexity logo

Perplexity

AI-ASSISTANTS
Velocity8.8

Conversational answer engine combining LLMs with live web search.

www.perplexity.ai ↗

Perplexity is selling access to other people's models, not just its own answers.

api-platformmodel-routingmcppricingagents
◆Current state
The recent changelog is almost entirely API-side. A Gateway API fronts open-weight models behind one endpoint that speaks both OpenAI Chat Completions and Anthropic Messages, a remote MCP server exposes Perplexity to outside agents, and the Agent API keeps absorbing new models. Consumer-facing notes — preset tuning, inline citations for research presets — read as maintenance beside that.
◆Where it's heading
The product is splitting in two: an answer engine for end users and an inference-and-routing layer for developers. Price moves in the same window, a GPT-5.6 cut and a faster low-cost mode, put Perplexity in a cost-per-token argument rather than an answer-quality one. Building the gateway to mimic the two dominant API dialects makes the switching cost it removes its own.
◆Prediction
Expect more hosted open-weight models behind the gateway and firmer pricing tiers, with the remote MCP server moving from a listed feature to a documented, permissioned surface.

◆Recent moves

  1. 10d ago

    Low preset updated

    The Agent API low preset now runs openai/gpt-5.6-luna with minimal reasoning effort and a 32,768-token output cap. Dynamic requests inherit it automatically; frozen configurations must be updated by hand — the recurring cost of Perplexity's preset abstraction.

  2. 10d ago

    GPT-5.6 price cuts and Sol Fast mode

    GPT-5.6 Luna and Terra get lower per-token rates and Sol gains a Fast mode at 2x standard pricing via service_tier: priority. Perplexity is passing upstream economics through and adding a latency tier on top.

  3. 11d ago

    Remote MCP Server

    The Perplexity MCP Server is now hosted at api.perplexity.ai/mcp for any Streamable HTTP client, authenticated with the existing API key. Tools and behavior are unchanged from the local server, so this removes installation and upgrade work rather than adding capability.

  4. 11d ago

    New: Gateway API

    ⚡ SPARK

    The Gateway API opens one endpoint onto frontier models from Anthropic, OpenAI, Google, xAI and Perplexity, with OpenAI and Anthropic wire compatibility, health-based routing and failover, and no per-request fee. It reframes Perplexity from answer engine to model broker.

  5. 14d ago

    Agent API: New Models

    The Agent API adds Claude Opus 5, the GPT-5.6 family, Gemini 3.6 Flash and 3.5 Flash Lite, Grok 4.5 and Kimi K3 at first-party token pricing, and retires a Gemini preview model. Routine breadth work, though it is the inventory the Gateway API now fronts.

  6. 14d ago

    Inline citations for research presets

    Search-backed presets now emit inline citations — numbered for fast, source-typed such as [web:1] for low, medium and high — with at least one citation guaranteed after a successful tool call. Provenance moves from a separate field into the answer text where agents can act on it.