Perplexity Claims Its Agent Product Computer Launched 15 Major Improvements in September
Perplexity claims its agent product Computer launched 15 major improvements in September, including Portable Computer landing on Linux, Windows, and Intel PCs, Hybrid compute on Mac, and Effort mode.
The Portable Computer integrates orchestrators, models, tools, local search, and task queues onto the device. Targeting Pro and Max, the hardware is NVIDIA DGX Spark or NVIDIA RTX Windows/Linux computers with at least 24GB of VRAM. Work completed locally does not consume Computer credits, and permission is required before moving to the cloud. The Windows version entered the Perplexity app on September 14, with the local model primarily being PPLX 27B.
Hybrid compute landed on Apple chip Macs on September 1, requiring macOS 15 or higher, with unified memory of at least 24GB, and 32GB recommended. Tasks are initiated in the cloud first, with sensitive files and local steps handled by the on-device model; it can be initiated via iPhone, with the powered-on Mac reading local files. Locally available options include Gemma 4 E4B, Qwen3.6 35B-A3B, and Perplexity's fine-tuned version, with local steps not deducting from cloud credits.
Effort mode offers four levels: Light, Standard, High, and Ultra, with the system matching models to inference intensity. Other updates include access to GPT-6 Astra and Skills Marketplace. In September, there were also HP pre-installed Windows applications and email-triggered tasks running complete jobs at computer@ addresses. The local version for DGX Spark/Linux was launched in August, connecting to over 20 authorized financial data sources including Dun & Bradstreet.
The local path is aimed at paid users who do not want files to leave their domain but need to run long tasks. Buyers are Pro/Max accounts with 24GB-level graphics cards or large memory Macs, while sellers are competitors selling pure cloud agents by token. Funding shifts from cloud credits to a one-time purchase of GPUs and subscriptions; beneficiaries are NVIDIA RTX and DGX hardware, while those under pressure are users without high VRAM workstations and free users. The cloud still handles retrieval, browsing, and cutting-edge inference, while local handles private files and scheduled tasks.
Source: Public Information
ABAB AI Insight
Perplexity has broken down the cloud-based Computer launched in February into three execution fronts: fully local on Windows/Linux, cloud then local on Mac, and Effort slider controlling costs. NVIDIA's branding on the Windows release effectively sets a 24GB VRAM threshold for gaming cards to find daytime loads. The Mac route accommodates unified memory, avoiding competition with NVIDIA for local orchestrators, focusing only on unloading sensitive steps. Effort, Astra, and Skills appearing in the same week indicates that the product is addressing model gaps through tiered usage and plugin markets, rather than just releasing a larger cloud model.
The capital path shifts billing from token-based to device sunk costs. Local tasks incur zero credits, cloud upgrades require permission, and billing changes from "pay per step" to "pay only when out of domain." HP pre-installation expands Windows entry, and the email address integrates Computer into existing workflows. Financial authorized data was connected in August, allowing local agents to reference external libraries without file transfer.
This targets computer usage agents against Claude and ChatGPT, as well as Ollama's self-built local stack. Perplexity is in an expansion phase transitioning from search box to persistent agent operation: first capturing personal workstations with discrete graphics, then enabling phones as remote controls. The industry is shifting from who has the strongest cloud model to who can complete tasks without file transfer.
The structural change is a shift in pricing power. The cost of agents is no longer solely determined by laboratories but jointly priced by user VRAM and whether domain transfer is authorized. The mechanism is: 24GB becomes the ticket for agents to enter offices, and the permission gate becomes the switch for privacy and billing. Whoever controls the orchestrator on which silicon determines whether credits can still be charged per conversation.
ABAB News · Cognitive Law
Files must stay in-house for credits to be preserved
Agents must pass the VRAM threshold to enter offices
The slider sells not intelligence, but how much you are willing to spend on inference