Perplexity Portable Computer, Checked: Local Runs Really Cost Zero Credits, but the Floor Is a 24GB GPU

On 2026-08-25, Perplexity and NVIDIA launched Portable Computer — the Perplexity Computer agent, repackaged to run entirely on your own machine. The headline is easy to remember: work completed locally consumes no credits.

That sounds like the end of running out of quota. Once you lay out the conditions, though, the sentence turns out to be heavily qualified. This piece does not restate the press release. It answers one question: what this actually means for your bill.

The news in three sentences

1. Portable Computer is the local edition of Perplexity Computer: the model, the agent harness, the conversation, and the execution trajectory all live on your device and stay there by default.
2. Local work does not draw down credits. Only steps that need the cloud — web search, a frontier model call — get billed, and each one requires your explicit approval first.
3. It is not a free tier. The official announcement opens it to Pro and Max subscribers, and the hardware bar is high: first release supports NVIDIA DGX Spark and RTX GPUs on Linux only.

The hardware bar: 24GB of VRAM is the floor, not a suggestion

This is the part most coverage skips, and it decides whether you can use the thing at all:

Perplexity described the 24GB VRAM number in interviews as the floor where they want to be sure they can deliver. In plain terms: most consumer laptops and mainstream GPUs are out.

As for the DGX Spark itself, market listings put the 2026 price at roughly US$4,699 per unit (it launched in October 2025 at US$3,999 and rose on memory supply constraints). We only found that figure in third-party roundups and retail listings, so we log it as second-hand evidence, not confirmed against NVIDIA’s own pricing page.

What “zero token cost” actually zeroes out

This is the part worth getting right, because it is easy to read as “no more monthly fee.”

• ❌ It does not save you the subscription. Portable Computer requires you to already be a Pro or Max subscriber. Our checked list prices are Pro US$20/month (about US$17/month billed annually) and Max US$200/month (about US$167/month billed annually) — our last direct read of Perplexity’s own site was 2026-08-08, and the site returns 403 to automated requests as a rule, so that read used no-cache headers and was repeated to confirm.
• ✅ It does save the credits consumed by Computer agent tasks. Cloud Computer burns credits across every step of a multi-step job, and high-volume repetitive work is exactly what drains a quota. The local edition pins that counter at zero — in a demo that reporters watched, the credit counter sat motionless at zero throughout a financial-document analysis.

So this is a product that trades a one-time hardware purchase for a recurring usage bill. It is not a replacement for the monthly fee, which you still owe.

It also keeps a cloud escape hatch: when a task needs more capability, you can escalate to a cloud model such as Claude Opus 5. In the reported demo, one terminal coding task was estimated at about US$0.415. That part still gets billed.

The trade-off: model choice drops from 19 to 2

Running locally is not free of cost in the other sense. Portable Computer currently offers two models:

• Qwen 3.8 27B (open source)
• PPLX 27B (Perplexity’s own post-trained variant)
• NVIDIA Nemotron 3.5 Lightning (a 30-billion-parameter mixture-of-experts model) is listed as coming soon

The cloud version of Computer offers 19 model options by comparison. Two more things you will hit in practice:

1. Nominal context and usable context differ. The stated window is roughly 256K–260K tokens, but reporting indicates it starts to struggle past about 100K tokens. Perplexity’s answer is context compaction, which automatically summarizes over-long prompts — meaning long-document workflows need you to verify nothing important got summarized away.
2. It still trails the frontier. Coverage states plainly that these local models lag frontier models on hard reasoning tasks, and that the published benchmark results come from Perplexity’s own evaluations, not independent third-party testing.

The privacy side is doing real work

If your concern is data leaving the building rather than saving money, this matters more than the credits:

• The whole stack runs locally by default; every task starts on the device, and escalating to the cloud requires user authorization.
• There is an OS-level sandbox that blocks unauthorized system access by the agent.
• Outbound content passes through a PII classifier.

For client data, financial documents, and internal source code — the material that should never have been pasted into a cloud chat box — this architecture addresses a real problem.

Who should care now, and who can wait

Worth evaluating today: anyone who already owns a DGX Spark or an RTX 3090-class Linux workstation (24GB VRAM and up), already pays for Pro or Max, and runs high-volume repetitive jobs that keep exhausting credits.

Wait for September: Windows users. The stated timeline is September 2026. There is no reason to buy hardware ahead of that.

Not your problem: Mac users (not on the roadmap) and anyone on the free tier or doing light Q&A. Buying a US$4,699 machine to save credits makes no arithmetic sense unless your cloud usage bill is already near that order of magnitude.

If what you actually want is simply to stop hitting quota walls, the far cheaper route is to master the free tiers first — see our OpenRouter free models daily limit guide and the Google AI Studio free tier breakdown.

What we could not verify, and therefore did not claim

Per our house rules, the following is flagged as unverified rather than written as fact:

• Whether Pro and Max differ inside Portable Computer (available models, concurrent tasks). The announcement says only “Pro and Max” and lists no differences — unverified.
• Whether Enterprise Pro / Enterprise Max get access at the same time. Only one outlet mentions it; the official wording names Pro and Max — insufficient evidence.
• The exact Windows release date and its hardware requirements. We have the month, September 2026, and no date — unverified.
• The current official DGX Spark list price. US$4,699 comes from third-party roundups and retail pages, not confirmed on the vendor’s own pricing page.
• Perplexity’s announcement and product pages return 403 to automated fetching (blocked, which is not the same as absent). The official wording quoted here comes from search-engine-indexed copy of those official pages, cross-checked against three outlets.

Sources checked

• Official announcement: Introducing Portable Computer for local-first AI (Perplexity Blog)
• Official research paper: A Local-First Agent for Private and Cost-Effective Knowledge Work
• SiliconANGLE: Perplexity AI launches Portable Computer on-device AI agent (2026-08-25)
• VentureBeat: Perplexity partners with Nvidia to launch Portable Computer
• Gizmodo: Perplexity Launches Local AI Model That Will Run on Your GPU
• Subscription list prices: our own Perplexity tool page verification log (last direct read 2026-08-08)

Related reading

• Claude Sonnet 5 API price lock, checked — if you plan to take the cloud-escalation route
• OpenRouter free models daily limit guide — cutting your credit burn with zero hardware spend

This is news verification and commentary, not investment or purchasing advice. Prices and specifications change; check the official pages before deciding. Last checked: 2026-08-26.