Perplexity October Update Adds Automations and Open-Source SPACE

A month of agent features

In the four weeks from late September to mid-October 2026, Perplexity released more than many software companies do in a quarter. Most of it pushes in one direction. The product is becoming an agent platform that happens to start from a search box, and the company is also building developer infrastructure it wants other teams to use.

The most visible consumer change came on October 5. Perplexity added Automations to Computer, its agent product, so it can run work on a schedule or when something happens in a connected app. The same release added interactive charts and 3D visualizations inside threads and gave eligible paid subscribers access to OpenAI's GPT-6.1 Sol, which also powers the new "Light" effort level.1 Two weeks earlier, on September 21, Perplexity had introduced Effort Mode, which offers four tiers: Light, Standard, High and Ultra. The same day it added GPT-6 Astra support, launched Portable Computer for Windows and Linux PCs, and opened a Skills Marketplace.2

The developer side moved just as fast. Perplexity released an open-weight 27-billion-parameter model behind a new Decisions API, followed by two multimodal embedding models.24 On October 15 it open-sourced SPACE, the sandbox that runs every Computer session.36 Taken together, these releases describe a company working on two fronts at once. It wants to be where knowledge workers get things done, and it wants to supply the plumbing for other people's agents.

Automations and Effort Mode

Automations are the feature with the most practical effect on daily use. Perplexity's examples show what it is aiming for. One automation follows up when a priority customer has gone quiet for four days by summarizing the account and drafting a next step.1 A UK training provider that checked the feature against the release notes says Automations can use Memory, Skills and connectors, can build on earlier runs, and can stop for human review before taking a consequential action.2

That review step matters. Agent products have struggled to show they can be trusted with ongoing work. A tool that only answers questions when asked is a research assistant. A tool that acts on its own when events happen is a different kind of product. By building in approval points, Perplexity is admitting that autonomy with no checks is not yet something businesses will accept.

Effort Mode deals with a less exciting but equally real problem: cost. Users can match how hard the system reasons to the job, so a light task like moving invoice data into a spreadsheet doesn't use the same compute as a market-entry analysis. The same reviewers say this makes cost and speed more predictable. Effort Mode is on the web for personal accounts first, with mobile and desktop to follow.2 This fits a pattern visible all year. Perplexity has kept adding controls over credits, including per-member credit limits for enterprise admins and usage analytics broken down by model.1

Finance features

The October 5 visualization release has a clear finance angle. Perplexity's own example asks Computer to chart Nvidia's past year of trading with candlesticks, volume, and 50- and 200-day moving averages, using TradingView Lightweight Charts.1 On its own that is a small addition. It matters more as the latest step in a finance push that has defined Perplexity's 2026.

The timeline shows how deliberate that push has been. In April, Perplexity linked personal checking accounts, credit cards and brokerage accounts through Plaid. The company said 75% of paying users tried the Finance features within 48 hours.11 In early May it launched Computer for Professional Finance. That product brought in licensed data from Morningstar, PitchBook, Daloopa and Carbon Arc, and shipped with more than 40 live tools and 35 prebuilt analyst workflows.1314 It also lets finance teams connect data subscriptions they already pay for instead of buying a bundle, an approach Perplexity calls bring-your-own-license.13

Since then the release notes have added more. Personal CFO now supports up to 30 linked accounts, up from five. Computer can research private companies, covering fundraising history, implied valuation and secondary pricing.1 Commentators have repeatedly compared the product to the Bloomberg Terminal and its roughly $30,000-per-seat annual price.1215 That comparison flatters Perplexity. Still, finance is the area where the company's mix of licensed data, citations and agent workflows makes the strongest commercial case.

Anthropic: partner and competitor

Perplexity's relationship with Anthropic has two sides. In finance they compete directly. On May 5, both companies launched institutional finance products on the same day. Anthropic offered ten configurable agents built on Claude Opus 4.7, and Perplexity offered workflows that trace back to source filings.18 Both are pitching hedge funds, investment banks and asset managers.18

On the model side, Perplexity sells a lot of Claude. Its Agent API lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, and Sonnet 5.5 at $2 and $10.22 Those prices match what Anthropic charges directly.21 Perplexity also recently switched its top "xhigh" API preset from OpenAI's GPT-5.6 Sol to Claude Opus 5.5, while moving the low, medium and high presets to the GPT-6 family.24 One analysis concluded that Perplexity now "partly functions as a router that sells you Claude."25

The two companies are very different in size. One comparison reported Anthropic's annualized revenue at about $65 billion in August 2026, against roughly $750 million for Perplexity.28 That gap explains Perplexity's strategy. It can't outspend the frontier labs on training, so it competes on orchestration, retrieval and price. Search is the clearest example. Perplexity's Agent API charges $0.0025 per web search call. Claude's API search tool costs $10 per 1,000 searches, though the two companies don't define a search call the same way.21 Perplexity also uses competitors' models to promote its own platform. Its release notes say xAI's Grok 4.6 matches Claude Fable 5's score on a long-task benchmark at 62.7% lower cost per task.1 That is a vendor's own benchmark, run by a company with a stake in the result, and it should be read that way.

Developer tools: SPACE and open-weight models

For developers, the most consequential release is SPACE, short for Sandboxed Platform for Agentic Code Execution. It is written in Rust, released under Apache 2.0, and runs each agent task inside its own Firecracker microVM.68 Network access and credentials are handled outside the virtual machine, so agents can reach approved services without ever seeing the secrets that grant access.58 CoreWeave, Akamai and Crusoe are already offering it on their platforms.57 Perplexity says every Computer session has run on SPACE since July.6

The coverage agrees on the main points but not on every detail. The red-team results are an example. MarkTechPost and BigGo Finance report 108 escape attempts, none of which reached the host.78 SiliconANGLE puts the figure at 118.5 All of them agree on the more important finding. Models did not break out of the VM, but some got around network restrictions: 11 of 54 runs before Perplexity fixed the weakness, and zero in retesting afterward.8 The speed figures look inconsistent but probably aren't, because they measure different things. One outlet reports that 90th-percentile sandbox creation time dropped from 447 to 89 milliseconds.8 Another reports the median falling from 185 to about 60 milliseconds.9

SiliconANGLE places the release in the context of recent incidents in which autonomous agents escaped their containment.5 It also fits a series of open-source security tools from Perplexity, including BrowseSafe, Numbat and Bumblebee.5 The likely strategy is to establish Perplexity's infrastructure as a standard and earn credibility with enterprise security teams before rivals do. Releasing it for free gives up little, since the hosting partners do the distribution.

The open-weight models point the same way. The Decisions API runs on pplx-decider, a 27B model for classification and routing that is licensed under Apache 2.0 and was updated on October 6.2 On October 7 Perplexity released pplx-embed-v2-late in 0.6B and 9B sizes. The models search text, images and PDFs without needing OCR, and because they share an embedding space, a corpus indexed with the large model can be queried with the small one.410 Meanwhile, Perplexity ended support for its legacy Sonar Chat Completions on September 27 and now routes those requests to the Agent API.21

What it means

Perplexity's business numbers depend on who is reporting. Estimates of its valuation run from about $20 billion to $22.6 billion, and estimates of annualized revenue run from $450 million to $750 million.161728 The figures don't line up, but they all describe a company that is small next to the labs it builds on.

The October releases address that weakness from both directions. Automations, Effort Mode and finance features give paying users a reason to stay even when the underlying models are Anthropic's or OpenAI's. SPACE, the embedding models and the Decisions API give developers reasons to depend on Perplexity's infrastructure. One reviewer has pointed out the risk: the products change so often that teams should check what's current before training staff on them.2 Even so, more of what Perplexity ships each month is software other companies run, not just a search interface on top of other labs' models.