The latest AI news and updates of Perplexity — an AI-powered answer engine delivering real-time, cited answers to any question. Covering Perplexity's latest product updates, launches, and research from the past 90 days.
PerplexityOct 23Perplexity released Decider for Mac, a menu-bar app that runs the 4B pplx-decider-flash-v1.2 model locally. It serves the Decisions API at localhost:8080, enabling text and image classification without an API key or data leaving the machine. The app requires an Apple silicon Mac with macOS 13 or later and 16 GB of memory for the 5.8 GB model.
PerplexityOct 22Perplexity launched a Batch API for the Agent API, supporting up to 50,000 requests per JSONL file. Each request executes the full agent loop, including built-in tools like web_search and fetch_url. This enables asynchronous processing for tasks like dataset enrichment, where the API handles the batch and provides results upon completion.
PerplexityOct 21Perplexity published JOLT, a post-training method that jointly trains one model as both a teacher and a student. By training the teacher to provide feedback within the student's reach, JOLT matched GRPO's accuracy on MATH with 6.5x fewer processed training tokens. JOLT+ outperformed GRPO across five benchmarks, including a 9.9-point gain on the AppWorld task.
PerplexityOct 20Perplexity Computer now runs fully local on supported Apple silicon Macs, keeping model reasoning and agent execution on-device. This mode processes sensitive files and native apps without sending data to the cloud or consuming Computer credits. Available to Pro, Max, and Enterprise subscribers, the local-first stack recommends 36 GB of unified memory for optimal performance.
PerplexityOct 16Perplexity released its next-generation open-weight multimodal decision models, pplx-decider-v1.2 and pplx-decider-flash-v1.2. Both models improve text and vision scores and achieve up to 5x faster inference by processing all questions in a single pass. The 27B v1.2 model ranks #2 on the Hugging Face Decision Index, while the 4B Flash model leads its tier.
PerplexityOct 16Perplexity added Google's Gemini 4 Argon to its platform for Pro and Max subscribers. The model sets a new record on Perplexity's WANDR benchmark, achieving a score of 0.802 at a cost of $7.88 per task.
PerplexityOct 15Perplexity added NVIDIA’s Nemotron 3.5 Super VL to its platform for Pro and Max subscribers and the Agent API. The model includes a dedicated Thinking toggle within the model picker, enabling activation of its reasoning capabilities for complex tasks.
PerplexityOct 15Perplexity open-sourced SPACE, the in-house sandbox platform powering Perplexity Computer, under Apache 2.0. Each agent task runs in an isolated Firecracker microVM, with credentials and network egress managed externally. In adversarial testing, nine model configurations made 108 escape attempts, and none reached the host. The platform supports stateful, branching work and is now available for developer integration.
PerplexityOct 14Perplexity added Ultrafast speed versions of GPT-6 Astra and SOL 6.1 to Perplexity Computer for Max subscribers. These versions generate tokens faster, completing long Computer tasks sooner. Access these models by selecting them under the Custom menu and setting the speed to Ultrafast.
PerplexityOct 12Perplexity built a Git-forge filesystem for Projects, allowing agents to work in ephemeral sandboxes while committing to a shared S3-backed repository. The system serializes concurrent submits to prevent overwrites and enables agents to search via standard file tools. On financial-document benchmarks, this approach achieved 0.627 answer accuracy, significantly outperforming the 0.37 accuracy of traditional semantic search methods.
PerplexityOct 6Perplexity released pplx-decider-v1.1-27b, an updated multimodal decision model that tops the Hugging Face Decision Index 0.3 benchmark. The model supports text and image input with a 250k context window. It is now available via the Perplexity Decisions API and Hugging Face at a price of $0.02 per million input tokens.
PerplexityOct 1Perplexity launched the Decisions API, powered by the pplx-decider-v1-27b multimodal model. The model outputs a probability distribution over a fixed set of choices rather than text, scoring 85.71% across benchmarks. Priced at $0.04 per million input tokens, the API automates tasks like support ticket triage by routing escalations to the Agent API.
PerplexityOct 1Perplexity Computer now generates interactive charts and visualizations directly within conversation threads. For financial data, the agent uses TradingView Lightweight Charts to render candlesticks, volume, and moving averages. These interactive elements allow for direct manipulation and exploration of data, expanding the agent's ability to deliver rich, visual outputs alongside text-based answers.
PerplexitySep 30Perplexity released pplx-embed-v2-context-9b-preview, a contextual embedding model trained by distilling relevance from a query-aware context compression model. The model achieves state-of-the-art results on ConTEB and turbopuffer’s context-bench, outperforming voyage-context-4 while using 8x less storage per vector. It is available now on Hugging Face.
PerplexitySep 29Perplexity launched Automations in Perplexity Computer, enabling agents to execute tasks on a schedule or in response to event triggers. These agents maintain memory of prior runs to pick up work without starting from scratch. Automations integrate with Slack, Gmail, Outlook, Linear, and GitHub to monitor updates and flag deadlines, with credits consumed only during active task execution.
PerplexitySep 29Perplexity published its security engineering framework for AI agents, featuring layered defenses including SPACE microVM isolation, BrowseSafe content screening, and Numbat endpoint monitoring. In adversarial testing, nine models given root access failed to escape the SPACE sandbox in 108 runs. The company also disclosed network-policy weaknesses found in eight of ten third-party platforms and requires human approval for rule changes.
PerplexitySep 28Perplexity updated its Agent API to support reusable, versioned agent configurations. Profiles store models, instructions, and tools under a single ID, while versioned Skills and managed connectors—currently including GitHub, Slack, and Datadog—allow teams to share resources across applications. Project administrators configure these assets in the API Console for consistent, repeatable agent workflows.
PerplexitySep 28Perplexity tested its SPACE sandbox by giving nine AI models root access to attempt VM escapes and network-policy bypasses. While no model breached VM isolation, four models exploited network-policy vulnerabilities through DNS spoofing and shared-IP routing. Perplexity implemented defenses to block these bypasses and found similar network-policy weaknesses in eight of ten third-party sandbox platforms tested.
PerplexitySep 25Perplexity launched Fast Search, a new Search API preset powered by its in-house Rust-based retrieval and ranking engine, Photon. The engine returns 95% of search results in 230 ms or less and reduces task costs by 68% across agent benchmarks. Photon now powers all retrieval and ranking across Perplexity, replacing the previous open-source engine.
PerplexitySep 24Perplexity now supports AMD Ryzen AI Max Series processors for its Portable Computer agent. The local-first stack runs the agent harness and models on-device, allowing Pro and Max subscribers to automate tasks across local files and apps like Slack and Gmail without consuming Computer credits. Cloud models remain available for complex steps upon user approval.
PerplexitySep 22Perplexity post-trained its Computer model using hint-guided self-distillation to learn from real-world errors. By imitating successful steps and correcting mistakes with grounded hints, the model achieved a 21.2% relative reduction in tool-call failures during live A/B testing. This approach improves agent reliability by focusing on decision-level corrections rather than just final task outcomes.
PerplexitySep 22Perplexity added OpenAI's GPT-6 Sol to its answer engine and Computer agent platform. The model now serves as the default selection for the Light effort tier in Computer's model selector. This tier is one of four presets designed to manage task efficiency and cost during agentic operations.
PerplexitySep 22Perplexity integrated Anthropic's Claude Opus 5.5 as the Standard effort level in Perplexity Computer. On Perplexity's WANDR benchmark, the model scored 0.610 at $4.13 per task. This configuration outperforms Claude Fable 5.1, which scored 0.601, while reducing task costs by 67.6%.
PerplexitySep 21Perplexity Computer now generates videos using MiniMax H3 and ByteDance Seedance 2.5. Requests for campaign clips, product demos, or social assets result in finished video delivered alongside copy and creative in the same thread. This capability is available now for Pro and Max subscribers.
PerplexitySep 17Perplexity is rolling out effort controls in Perplexity Computer, adding a slider to the model selector with four presets: Light, Standard, High, and Ultra. These presets combine the orchestrator model and reasoning depth to manage task efficiency and cost. The update is available now on the web, with mobile and desktop support arriving soon.
PerplexitySep 16Perplexity released a Search SDK cookbook featuring a Search as Code recipe for coding agents. The Python script automates focused web searches, filters results to official documentation, extracts relevant passages, and compiles them into a source-linked Markdown brief. This process provides agents with verifiable evidence to improve the accuracy of dependency migrations and code edits.
PerplexitySep 15Perplexity published research on CobbleDB, a key-value database built by two engineers and hundreds of AI agents in two months. Replacing DynamoDB, CobbleDB reduces median batch-read latency by 5x and lowers infrastructure costs by at least 20%. The company plans to open-source the database for other AI search teams.
PerplexitySep 15Perplexity is pre-installing its Computer agent on the HP ZBook Ultra G3a mobile workstation. The integration includes Portable Computer, a local-first version that runs tasks on-device without cloud credits. The platform also connects to Autodesk Revit via a new MCP server, with the rollout expanding to additional HP Windows devices in the coming months.
PerplexitySep 14Perplexity has launched Portable Computer for Windows PCs equipped with NVIDIA RTX GPUs featuring at least 24GB of VRAM. The local-first agent stack runs the harness, orchestrator, and models on-device, supporting local file interaction, app connectors, and scheduled tasks. Cloud model escalation is available for complex tasks while local control of the agent stack is maintained.
PerplexitySep 10Perplexity updated its API MCP server to support OAuth authentication. Agents now connect by adding the server URL to a client, signing in with a Perplexity account, and approving the organization. This integration provides agents with real-time web search, deep research, and advanced reasoning capabilities, all powered by the underlying Agent API.
PerplexitySep 9Perplexity introduced Q2D-Web, a benchmark and public leaderboard for evaluating retrieval in agentic RAG systems. It features 190 million web documents and 69,721 agent-reformulated queries across ten languages. Initial results show pplx-embed-v1-4b leading Web Ranking and Combined metrics on Recall@1000, while Nemotron-3-Embed-8B leads in Citation-based retrieval.
PerplexitySep 9Perplexity integrated its Search API into the Hermes Agent platform, providing real-time web search results across an index of more than 400 billion URLs. The integration delivers relevance-ranked snippets directly within agent workflows. New accounts receive $10 in free API credits to begin building with the search-enabled agent backend.
PerplexitySep 5Perplexity published research on its in-house inference stack, Ivy, Tulip, and ROSE, which powers its embedding and ranking models. The infrastructure uses CUDA graphs, LazyTensors, and ragged attention to overlap CPU and GPU work. These optimizations reduce latency and improve throughput for both live-search and batch-indexing workloads, delivering faster search results at a lower cost than off-the-shelf solutions.
PerplexitySep 4Perplexity evaluated OpenAI's GPT-6 Astra on its WANDR benchmark, where it scored 0.682 at $11.98 per task. This result marks the highest score of any model tested. Compared to Fable 5.1, Astra scored 13.5% higher at 6.1% lower cost, and it outperformed Opus 5 by 27.0% at 3.3% higher cost.
PerplexitySep 3Perplexity has integrated its API into Stripe Projects, enabling terminal-based provisioning of API projects, keys, and prepaid credits. The flow supports auto-provisioning for new users and linking existing projects to maintain current billing. Provisioned keys provide access across the entire Perplexity API platform, including the Agent and Search APIs, with Stripe managing spend limits and key rotation.
PerplexitySep 2Perplexity open-sourced Lily, a local inference engine built for hybrid compute on Apple silicon. Specialized for Qwen3.6-35B-A3B, Lily maps model operations directly to hardware, achieving 1.23× higher prefill and 1.35× higher decode throughput than MLX-LM on an M5 Max MacBook Pro. The engine uses a Rust runtime and custom Metal kernels to minimize data movement during inference.
PerplexitySep 2Perplexity’s Agent API now supports 42 models, adding Z.AI’s open-weight GLM 5.3 reasoning model. Developers can now configure fallback chains to ensure high availability if a model becomes unavailable, and switch between models by changing a single request field. A new cookbook example demonstrates pairing GLM 5.3 with the people_search tool for customer record enrichment.
PerplexitySep 1Perplexity launched hybrid compute for its Mac app, splitting tasks between cloud models and a local model to keep sensitive data on-device. An on-device PII classifier detects personal information before it leaves the machine, ensuring sensitive steps remain local. This feature is available today for Pro, Max, and Enterprise subscribers on macOS 15+.
PerplexityAug 29Perplexity launched managed Connectors for its Agent API, providing integrations for GitHub, Slack, Google Drive, and Datadog. API Group admins connect each service once in the API Portal, allowing agents to reference them by ID without per-request tokens. The GitHub connector integrates with the Sandbox, enabling agents to clone private repositories, modify code, and open pull requests autonomously.
PerplexityAug 29Perplexity integrated Z.ai's GLM 5.3 into its Computer agent platform. Designed for long-context and multimodal workloads, the model scored 0.278 on Perplexity's WANDR benchmark, outperforming the previous GLM 5.2 version by 0.019 points. The model is available now for agentic research tasks.
PerplexityAug 26Perplexity updated Brain, its self-improving memory system for Computer, showing a 9.3-point gain in correctness, 8.0 in currentness, and 8.9 in recall while reducing token usage by 15%. Brain organizes context as a linked Markdown filesystem, using background Dream agents to synthesize information and verify claims through citation edges, ensuring agents maintain an accurate, navigable knowledge wiki.
PerplexityAug 25Perplexity launched Portable Computer, a fully local version of its agent platform running on NVIDIA DGX Spark. The system executes the orchestrator, subagents, and agent harness locally, eliminating cloud dependencies for on-device tasks. It supports PPLX 27B and Qwen 3.8 27B models, with user-gated cloud escalation to frontier models for complex tasks. Available today for Pro and Max subscribers.
PerplexityAug 22Perplexity released the Agent API, providing access to 41 frontier models across nine providers through a single endpoint. The API includes built-in tools for web search, finance search, URL fetching, and code execution. Presets range from fast to wide-research, supporting up to 100 steps. Pricing follows provider-direct rates with no markup, and tool fees apply only upon invocation.
PerplexityAug 22Perplexity released its API MCP server as an Agent Plugin 1.0.0 package. This portable format allows compatible clients like Cursor, Codex, and VS Code to install the server directly without per-client adaptation. The plugin provides agents with real-time web search, reasoning, and research capabilities through the Perplexity API.
PerplexityAug 22Perplexity integrated DeepSeek V4 Pro into its Computer agent platform with U.S. hosting. Internal WANDR benchmark evaluations show the model scoring 0.359 at $0.75 per task. This performance-to-cost ratio is 62 percent lower than the next model on the cost-performance frontier.
PerplexityAug 22Perplexity now enables Computer tasks directly through email. Sending, forwarding, or cc’ing computer@perplexity.com initiates a full Computer session, maintaining the same audit trail and memory as web-based tasks. The platform verifies sender permissions and returns deliverables as attachments, making the agent available to all Computer users today.
PerplexityAug 15Perplexity released an agent-first Python SDK that implements its Search as Code architecture. The SDK provides composable retrieval primitives, including ranking, filtering, and query fan-out, for direct orchestration by coding agents. It includes a dedicated pplx-search-sdk agent skill to integrate real-time web search and page snippets directly into agent workflows.
PerplexityAug 13Perplexity is migrating its Sonar model to the Agent API, which bundles multi-step research, code execution, and multi-model access into pre-configured presets. The new API outperforms Sonar models on BrowseComp and WideSearch benchmarks. Existing Sonar users have 45 days to migrate their integrations to the new platform.
PerplexityAug 13Perplexity integrated xAI’s Grok 4.6 into its answer engine and Perplexity Computer platform. On the WANDR benchmark, the model matches the performance of Anthropic’s Fable 5 while reducing task costs by over 60%. This addition provides a high-efficiency option for research and agentic workflows.
PerplexityAug 13Perplexity updated its Search as Code system, increasing WANDR benchmark scores by 9.2% while reducing per-task costs by 9.7%. Additionally, updates to the Search SDK within the Computer platform raised action execution reliability from 81.9% to 92.6%. These improvements enhance the efficiency and dependability of autonomous research workflows.
PerplexityAug 10Perplexity has integrated Moonshot AI’s Kimi K3 model into its Agent API, hosted exclusively on US-based servers. The integration supports long-horizon research and code execution, allowing agents to generate source-linked, interactive HTML data stories. The API combines live web search with K3’s reasoning capabilities to produce automated, inspectable reports.
PerplexityAug 8Perplexity integrated OpenAI’s GPT-5.6 Terra and Luna models into its Computer agent platform. Terra now serves as the default model for subagents and an orchestrator, while Luna handles scheduled automations. These models reduce subagent costs by 28 percent and automation costs by 72 percent, with Terra outperforming Claude Sonnet by 11 points on the WANDR benchmark.
PerplexityAug 5Perplexity has integrated Anthropic’s Claude Fable 5 into its Agent API, bundling the model with native search, retrieval, and code execution capabilities. The integration targets complex agent workflows, with pricing set at $10 per million input tokens and $50 per million output tokens. The API requires a max_output_tokens parameter for all Anthropic model requests.
PerplexityAug 3Perplexity released the pplx CLI, providing coding agents with direct access to live web search and query-relevant page snippets from the terminal. The tool supports automated web research and content extraction for non-interactive agent workflows. Agents authenticate using the PERPLEXITY_API_KEY, and the CLI outputs structured JSON data for integration into existing development environments.
PerplexityAug 1Perplexity launched a hosted remote MCP server, enabling direct integration with Claude Code, Cursor, and VS Code. The server connects using an API key without requiring local installation or maintenance. It exposes Perplexity’s search, reasoning, and research tools, providing AI assistants with real-time web context and advanced analysis capabilities directly within existing coding workflows.
PerplexityJul 30Perplexity launched Projects, an evolution of its Spaces feature, as a collaborative hub for ongoing work within its Computer agent platform. Projects provide a shared file system and persistent memory through Brain, where agents retain context across tasks and sessions. The update is available to all users today, supporting team collaboration and tool integration.
Frequently asked questions