Is Perplexity AI Down?

Quickest ways to check the current status of perplexity.ai plus recent known issues and working alternatives if it's out.

Last editorial review: 2026-10-02

How to check right now

Step 1 -- Load the site
Open perplexity.ai→

The simplest check. If it loads and you can log in, the tool is up for you.

Step 2 -- Crowdsourced reports
DownDetector: Perplexity AI→

User-reported outages aggregated across regions. Good signal for partial or regional outages.

Step 3 -- Live social
Search X for “Perplexity AIdown” →

Live tweets from other users. Fastest signal for a fresh outage -- often beats status pages by 10-20 minutes.

Step 4 -- Official status page
Find Perplexity AI's status page →

Most vendors publish a status page at status.[domain] or on Statuspage.io. Google-search to find the current URL.

Known issues we've tracked

AUTOMATIONS REPLACE SCHEDULED TASKS IN COMPUTER -- EVENT TRIGGERS, MEMORY ACROSS RUNS, MISSED-DEADLINE FLAGS -- AND AN AMERICAN EXPRESS SKILLS BUNDLE FOR ENTERPRISE SUBSCRIBERS (2026-09-29 and 2026-10-01, vendor-primary): **Automations (9/29):** 'bring Computer's capabilities to long-running agents that work proactively on a schedule or in response to conditional events'; each automation uses Computer's connected files, tools and Projects and **'remembers prior work, picking up where it left off'** -- a weekly competitor-pricing report compares against last week's findings, a customer-service automation drafts replies with earlier emails in mind. **'Automations replace Scheduled Tasks, adding event-based triggers and memory of prior runs'**; triggers are a schedule or a specified event in **Slack, Gmail, Outlook, Linear or GitHub**, narrowed by conditions (e.g. an email from a named sender asking for a decision). Automations can also flag absence -- 'Tell me if the Tuesday deployment hasn't shipped by Wednesday morning'. Setup from the omnibar or Automations > New Automation; users specify which actions Computer may take alone and which need review ('Ask for review before sharing an update in Slack'); admins scope connector permissions; run history records the trigger and the output. No credit pricing is stated in the post. **AmEx Skills (10/01):** 'ready-to-use Skills for Perplexity Computer specifically curated for U.S. American Express Business Card Members with a Perplexity Enterprise subscription' -- ten pre-built workflows across finance, marketing, operations and hiring (cash flow forecasting, tax-season prep, vendor benchmarking, scenario planning, reconciliation and more), reached by opening Personal CFO from the connectors page and linking the Amex Business Card through Plaid; forecasts pull the operating bank account via Plaid and optionally an accounting tool. **'Running Skills requires Computer Credits, and usage varies with the complexity of the task.'** Both items are Computer features; the Perplexity Computer page carries the same note.
2026-10-01•Perplexity (perplexity.ai/hub/blog/computer-adds-automations-for-ongoing-work, 'Sep 29, 2026'; perplexity.ai/hub/blog/perplexity-and-american-express-make-ai-easier-for-growing-businesses, 'Oct 1, 2026'; dates from the hub index 'Recently published' list via Firecrawl, bodies via curl) -- fetched 2026-10-02
AGENT API NOW SUPPORTS REUSABLE AGENTS -- VERSIONED PROFILES, SKILLS AND MANAGED CONNECTORS (2026-09-28, vendor-primary): 'Today we're making these agent workflows easier with updates to Perplexity Agent API that give teams a reusable, versioned configuration layer for agents. We're adding custom Skills, managed connectors, and Profiles.' **Profiles** 'assemble an agent by saving its model, instructions, tools, Skills, connectors, and run settings under one versioned ID' -- configure once, make it available to authorized developers across a project, invoke the same procedure from an internal tool, CI workflow or service automation. **Versioned Skills** package instructions, procedures and supporting files. **Managed connectors** let project administrators configure approved integrations once instead of every developer wiring the service; users can also build custom connectors for internal systems. Availability: 'Profiles and custom Skills are available now for Agent API projects. Managed Connectors are available in preview, with **GitHub, Slack, Google Drive, Datadog, Linear, and Notion** currently supported', configured in the API Console and referenced with a project API key. Perplexity reminds that last month the Agent API became 'one programmable surface for web search, URL fetching, code execution, MCP, and specialized search tools'. No pricing change stated.
2026-09-28•Perplexity (perplexity.ai/hub/blog/agent-api-now-supports-reusable-agents; hub index date 'Sep 28, 2026' via Firecrawl) -- fetched 2026-09-28
PORTABLE COMPUTER ON AMD RYZEN AI MAX -- AND PERPLEXITY FINALLY NAMES THE LOCAL MODELS: PPLX 27B AND QWEN 3.8 27B (2026-09-24, vendor-primary): 'Today we're expanding Portable Computer to systems powered by AMD Ryzen AI Max Series processors including the Ryzen AI Halo developer platform.' The line this page has waited for since 8/25: **'Available models include Qwen 3.8 27B and PPLX 27B, Perplexity's post-trained model'** -- so the Qwen 3.8 27B attribution we carried from NVIDIA's 9/14 blog is now Perplexity-primary, and PPLX 27B is a named Perplexity checkpoint (the model selector screenshot shows 'PPLX 27B selected as the local model, Qwen 27B highlighted, and cloud models listed below'). Economics: **'local inference doesn't consume Computer credits'**; recurring local tasks reuse the same local model; the local agent can still escalate to the cloud for advanced research and reasoning; local MCP servers reach apps on the same PC; connected apps only through user-configured connectors; an administrator decides whether it is allowed on enterprise plans. **Requirements:** an AMD Ryzen AI Halo developer platform or any Ryzen AI Max Series system with **at least 24 GB of GPU-accessible memory**, Windows 10 or 11, ~20 GB free disk for setup; **available today to Pro and Max subscribers on individual and enterprise plans through the Perplexity app for Windows**. That matches the NVIDIA-side 24 GB floor recorded 9/14 and extends Portable from RTX to AMD's unified-memory APUs.
2026-09-24•Perplexity (perplexity.ai/hub/blog/portable-computer-comes-to-amd-powered-agentic-pcs; hub index date 'Sep 24, 2026' via Firecrawl) -- fetched 2026-09-28 via curl with browser UA
ESCAPING SPACE, PART I -- PERPLEXITY RED-TEAMED THE SANDBOX BEHIND COMPUTER AND FOUND NETWORK BYPASSES, NOW FIXED (2026-09-23, vendor research): 'We tested two containment boundaries in SPACE, the sandbox platform behind Perplexity Computer: VM isolation and network confinement. Nine models received root access inside a guest VM and attempted to obtain a host-side secret or reach a blocked network destination. **No VM-to-host escape was observed in 108 runs.** With limited network access, however, **four models bypassed network restrictions through DNS spoofing or shared-IP routing.**' Part I covers those bypasses, the defenses implemented and 'a reevaluation in which neither mechanism produced a verified bypass'; it also reports results for nine of ten other sandbox platforms and summarizes disclosures to affected vendors (one vendor asked for more time and is omitted). Part II will cover the models' attack strategies. Relevance: Computer runs real software with credentials, and this is the vendor's own evidence that network confinement, not VM isolation, was the weak boundary -- consistent with the sandbox-escape behaviors Anthropic and OpenAI described in their September alignment posts.
2026-09-23•Perplexity (perplexity.ai/hub/blog/escaping-space-part-i; hub index date 'Sep 23, 2026' via Firecrawl) -- fetched 2026-09-28 via curl with browser UA
WHAT ACTUALLY RUNS PERPLEXITY COMPUTER: GLM 5.2, POST-TRAINED BY PERPLEXITY ON YOUR SESSIONS (2026-09-21, vendor-primary research post): Perplexity's 'Learning from Real-World Mistakes' is a training write-up, but it settles a question this page has carried since Computer launched -- **which model is underneath.** Verbatim: the pipeline 'draws from a training-eligible subset of real-world Computer sessions **served by GLM 5.2**', the trainer 'runs the same GLM 5.2 checkpoint twice' (teacher with a hint, student without), and '**the post-trained student then becomes a model option in Computer**'. So at least one of Computer's selectable models is a Perplexity fine-tune of Z.ai's open-weight GLM 5.2 -- the same lineage as the Qwen 3.8 27B it post-trained for Portable Computer. **The method, briefly:** rejection-sampling fine-tuning on successful sessions plus hint-guided On-Policy Self-Distillation on marked mistakes, where a 'hint' is a short correction grounded only in information the model already had (tool schema, earlier turns); user complaints and tool errors are traced back to the responsible turn by LLM judges. **The numbers, with Perplexity's own caveats attached:** offline tool-error rate 2.79% for stock GLM 5.2, 1.35% for an RFT-only checkpoint, 0.87% with RFT plus self-distillation ('not a controlled ablation'); a first live A/B against stock GLM 5.2 was not statistically significant; a second A/B between two trained checkpoints (about 100,000 users each) cut tool-call failures from **2.24% to 1.77%, a statistically significant 21.2% relative reduction**, with **no significant change in user dissatisfaction** (2.58% vs 2.54%). Perplexity states the training excludes sessions with PII and from users who opted out. **Why this belongs on the page:** buyers evaluating Computer against ChatGPT Work or Claude Cowork should know its default agent brain is an open-weight Chinese-origin model tuned in-house, not a frontier API -- that is the source of both its credit economics and its ceiling. (Also relevant to glm.ts: GLM 5.2 now has a named production deployment at scale.)
2026-09-21•Perplexity (perplexity.ai/hub/blog/learning-from-real-world-mistakes, hub index date 'Sep 21, 2026'; contributors listed incl. Denis Yarats) -- rendered via Firecrawl 2026-09-21
COMPUTER GETS EFFORT CONTROLS -- A FOUR-STEP SLIDER THAT PICKS THE MODEL AND REASONING LEVEL FOR YOU (2026-09-17, vendor-primary): Perplexity Computer now has an **effort slider in the omnibar with Light, Standard, High and Ultra settings**; each setting 'selects the model responsible for coordinating an assignment and how deeply it reasons', and the orchestrator delegates sub-tasks to supporting agents that 'can use models from different providers'. Perplexity's examples: Light for collecting vendor invoices into a spreadsheet, Standard for a sales brief, High for prioritising product work, Ultra for evaluating a new market. **Cost rule as stated: 'Lower settings use less expensive models. Credit usage depends on the work Computer performs'** -- no credit multipliers are published, so the slider is a cost lever without a price list. Custom controls (pick a specific model and reasoning level) remain. **Web today; Android and iOS 'soon'.** The screenshot shows Ultra mapping to GPT-6 Astra at maximum effort. This is the same shape as GitHub Copilot's efficiency/balance/intelligence auto-selection tiers (9/14) and Claude Code's effort levels: vendors are converging on abstracting model choice behind a spend dial.
2026-09-17•Perplexity (perplexity.ai/hub/blog/computer-adds-effort-mode-for-model-selection, hub index date 'Sep 17, 2026') -- rendered via Firecrawl 2026-09-17

Issues here are sourced from our editorial sweeps, not real-time telemetry. Newer issues may exist.

What to use if Perplexity AI is down

Top AIToolTier-ranked alternatives in the same category, ordered by our overall score.

A

ChatGPT

8.8

**Oct 16: ChatGPT Slides began rolling out to Pro, Business and Enterprise users with Space access -- build a deck from a chat, files or an outline, co-edit it with teammates in real time, and export to PowerPoint, PDF or Google Slides -- and GPT-6 Pro in the Chat tab got faster responses, better plugin and tool use, and parallel agents for Pro users.** **Oct 15: GPT-Live-1.1 began rolling out in ChatGPT Voice, Plus and Pro lost the fixed cap on how many Sites you can create, and US academic labs can apply for the Rosalind Academic Plan (up to 30 free Business seats for 12 months); on 10/14 OpenAI gave API developers notice that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini snapshots leave the API on 2027-01-14.** **GPT-6 reached the Chat tab on 2026-10-07 with Intelligent UI: answers can now be built from interactive components (diagrams, maps, calculators, forms, small games), GPT-6 can start answering while it keeps thinking, and the rollout runs GPT-6 Sol for Plus, Pro, Business and Enterprise and GPT-6 Luna for Free and Go (from 10/08). Audio file uploads for transcripts and summaries arrived 10/06 on paid plans.** **Oct 6: OpenAI expanded its Atlassian partnership (GPT-6 family models power Rovo and Atlassian agents; Jira, Confluence and Teamwork Graph plugins for ChatGPT and Codex) and published its first computer-use research partnership, with Ironclad, where GPT-6 Astra scored 55.0% on 11 contracting-software tasks against 41.6% for GPT-5.6 Sol in roughly half the time.** **Custom GPTs are being retired: Dec 11, 2026 for all plans (Feb 11, 2027 for Enterprise workspaces with an approved deferral), with new-GPT creation ending Oct 26 -- migrate them to plugins, and note that custom actions do not transfer.** A new Enterprise rate card replaces per-message Chat billing with per-token rates (GPT-6 Pro $10/$50, GPT-5.6 Sol $4/$20 per 1M) and publishes first prices for Daybreak Red and Astra for Law. On 10/05 OpenAI also added a visual ad format inside image generation (US test later this month) and said EU ChatGPT and Codex output will carry an invisible text watermark within weeks, with worldwide API opt-in from today. **DevDay 2026 (2026-09-29) added a model, an agent class and a $500 plan in one day: GPT-6.1 Sol at $2/$10 per 1M with cached input cut to $0.10, dots (always-on agents with their own cloud computer, powered by GPT-6 Astra), and Pro 500 with Astra Ultrafast** -- plus Codex Cloud, Pages and Space, the Meetings plugin, Sign in with ChatGPT across 16 partner apps and a reopened Pro 200. GPT-6.1 Sol nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth of Astra's token prices, but like GPT-6 Sol it runs in ChatGPT Work, Codex and the API and is 'not yet available in Chat'. Then on 9/30 OpenAI said it had disrupted a coordinated model-distillation campaign it attributes to individuals associated with Moonshot AI, and on 10/01 it put tts-1, tts-1-hd and gpt-4o-mini-tts on a 2027-01-06 API removal list and GPT-5.3-Codex, GPT-5.1 and GPT-5.4-Nano on a 2027-04-01 list. Virtual try-on shopping (10/01) and Finances for Free and Go (10/02) round out the window.

A

Alexa+

8.3

Amazon's reborn consumer AI assistant -- rolls out to every Alexa device with memory, multi-turn agent tasks, and Fire TV navigation

B

Gemini Notebook (formerly NotebookLM)

7.8

Google's free research assistant that turns your documents into an AI you can query -- and a podcast you can listen to. **Renamed from NotebookLM to Gemini Notebook on 2026-07-16**, still a standalone product, now with in-notebook code execution and syncing across the Gemini app and Search. **Expert Intelligence (2026-08-27) makes purchased Play Books ebooks a first-class source**; compute-weighted usage limits took effect 2026-09-02 -- they refresh every 5 hours **but sit under an undisclosed weekly cap**, with AI Plus at 2x standard, AI Pro 4x and AI Ultra 5x or 20x above Pro (Google's help pages still read prospectively and still publish the old daily table -- no delay announced)

B

Microsoft Copilot

7.5

Microsoft's answer to ChatGPT -- free GPT-4 access with Bing search built in, but it feels like a corporate product

B

Poe

7.5

One subscription, every major AI model -- Quora's multi-model chatbot lets you switch between GPT-4, Claude, and more

About Perplexity AI

Tier A (8.5/10). **Perplexity Computer gained Automations on 2026-09-29 -- long-running agents that run on a schedule or when an event fires in Slack, Gmail, Outlook, Linear or GitHub, remember prior runs instead of starting from scratch, and replace Scheduled Tasks** -- followed on 10/01 by a curated set of ten business Skills for US American Express Business Card Members on a Perplexity Enterprise subscription (cash flow forecasting, tax prep, vendor benchmarking; Skills consume Computer Credits). Earlier: **Perplexity names its local models at last: Portable Computer reached AMD Ryzen AI Max PCs on 2026-09-24 running PPLX 27B (Perplexity's own post-trained model) or Qwen 3.8 27B**, 24 GB of GPU-accessible memory minimum, Windows 10/11, ~20 GB disk, Pro and Max on individual and enterprise plans -- and local inference does not consume Computer credits. On 2026-09-28 the **Agent API gained reusable agents**: versioned Profiles (model, instructions, tools, run settings under one ID), versioned Skills and managed connectors (GitHub, Slack, Google Drive, Datadog, Linear, Notion in preview). Its 9/23 'Escaping SPACE' report red-teamed the sandbox behind Computer: no VM-to-host escape in 108 runs, but four models bypassed network confinement via DNS spoofing or shared-IP routing before defenses were fixed. The 9/21 GLM 5.2 disclosure stands.

Read the full Perplexity AIreview →