Perplexity Computer’s July 2026 Update for Commercial Real Estate

Perplexity’s July 13, 2026 changelog groups several Computer updates into one release: Brain, a source-linked memory system; Opus 4.8 Fast mode on the web; Claude Fable 5 as an orchestrator; mid-task orchestrator switching; website publishing; model-level usage analytics; account switching on Mac; and private-company financial research sourced from Forge Global. Perplexity says the private-company feature covers fundraising history, implied valuation, secondary pricing, public marks, and company profiles for all Computer users without a separate Forge license.

For Brain, Perplexity says Computer builds a private context graph from prior tasks, connectors, files, and decisions, refreshes that graph overnight, links memories to their source, and lets a user remove retained items under Customize. The provider reports gains in correctness, recall, and cost on tasks with prior context. Those are Perplexity’s test results, not CRE benchmarks; PSV did not independently test this release and does not convert those figures into an underwriting, diligence, or productivity claim.

Availability is feature-specific. Opus 4.8 Fast mode is live on the web with mobile described as coming later. Website publishing is available on the web to users who can create sites with Computer. The release page does not state a regional restriction, but it also does not promise identical availability in every plan, workspace, or geography. No release-specific system card is linked.

Computer is credit-based. Perplexity’s July 10 billing guide says 100 credits currently equals $1, while warning that pricing and task ranges can change by plan, promotion, or region. It lists Computer access for Pro and Max plans, including Enterprise Pro and Enterprise Max; the published starting monthly allocations are zero for consumer Pro, 10,000 for consumer Max, 500 for Enterprise Pro, and 15,000 for Enterprise Max. The new Computer Analytics API is limited to Enterprise organizations, uses an organization-scoped analytics-only key, and does not consume API credits. The cited materials do not announce API access to Brain or private-company research.

• Provider and date: Perplexity, July 13, 2026.
• Material CRE capabilities: persistent source-linked context, private-company financial research, model switching, usage analytics, and controlled internal publishing.
• Important limits: feature-by-feature availability, credit consumption, no linked system card, and no announced API for Brain or Forge-backed research.

The release is relevant to commercial real estate because recurring work is unusually context-heavy. A tenant review, acquisition screen, weekly asset-management update, or broker pursuit rarely begins with a clean packet. The analyst has to recover prior decisions, identify the current file, remember which assumption was rejected, and reconstruct the format expected by the reviewer. Source-linked memory can reduce that restart burden, but only if the workflow treats memory as a lead back to evidence, not as evidence itself.

Private-company research is also operationally relevant, particularly for brokerage, leasing, and tenant or sponsor diligence. Fundraising and valuation records can help an analyst frame questions about a private company, but they do not establish tenant credit, liquidity, lease capacity, guarantor strength, or sponsor net worth. The output should remain a research brief with dated sources and open questions, not a credit conclusion.

• Tenant and leasing research. Inputs: the approved requirement record, prior tour notes, public company sources, and permitted private-company data; output: a dated tenant brief and unresolved-question list; reviewer: broker or leasing manager; approval gate: before any credit characterization, proposal term, or client communication.
• Acquisition diligence. Inputs: the data-room register, prior issue log, current rent roll, operating statements, and specialist reports; output: a refreshed exception queue with source links and changed assumptions; reviewer: acquisitions lead with counsel, engineering, environmental, or tax specialists as applicable; approval gate: before closing an issue or changing the investment recommendation.
• Asset management. Inputs: approved prior report, current actuals, budget, rent roll, capital plan, and property-manager updates; output: a variance draft that distinguishes remembered context from current-period evidence; reviewer: asset manager; approval gate: before forecasts, business-plan changes, lender messages, or investor reporting.
• Brokerage and market research. Inputs: prior pursuit decisions, licensed market data, ownership research, call notes, and approved positioning; output: a source-backed pursuit brief in the team’s standard format; reviewer: deal lead; approval gate: before pricing, representations, outreach, or CRM writes that affect the relationship record.

A persistent context graph introduces a new category of operating risk: a statement can be accurately remembered and still be wrong for today’s deal. A prior asking rent may have expired, a tenant’s funding round may no longer describe its condition, or an old committee preference may not apply to a new strategy. Every reused item should carry a source, an as-of date, a deal or portfolio identifier, and an owner. If any of those are missing, the system should ask for confirmation instead of silently carrying the item forward.

Firms also need a boundary between matters. Memory from one client, fund, property, or transaction should not leak into another workflow merely because the documents look similar. Before connecting live data, an administrator should verify connector permissions, file-sharing controls, model-provider settings, retention, audit coverage, and the process for deleting an incorrect memory. Perplexity documents enterprise controls in these areas, but some advanced retention and audit features require either an organization with at least 50 seats or at least one Enterprise Max user.

Source links improve inspectability, but they do not resolve conflicts. If Brain recalls one decision while the current source packet shows another, the workflow should preserve both records, highlight the conflict, and route it to a named reviewer. The current document and authorized decision record, not the most fluent summary, must control.

Test the release on a bounded recurring workflow with representative but permitted material. A strong example is a four-week tenant-review or asset-management cycle using fictional or approved property files: a current rent roll, two operating reports, a tenant roster, prior meeting notes, a dated company-research packet, and the firm’s normal output template. Do not begin with confidential live-deal documents until the account, retention, connector, and sharing configuration has been approved.

Create a baseline run without inherited context, then repeat the workflow with Brain enabled. Plant controlled changes: replace one rent roll, reverse one prior decision, expire one market source, remove one tenant from the roster, and add a contradictory note. The test is not whether the second report sounds more consistent. It is whether the system retrieves the right prior context, links it to the right source, recognizes what changed, and refuses to treat a stale item as current.

• Measure retrieval precision: which remembered items were relevant to this property, period, client, and workflow?
• Measure source integrity: can a reviewer open the exact session or file behind each consequential remembered statement?
• Measure change handling: does the system prefer the current authorized record and surface contradictions rather than smoothing them over?
• Measure review burden: count corrections, unresolved exceptions, and reviewer minutes; do not use provider benchmark claims as a substitute for your own test.
• Measure economics: record credits by run and model, then compare the total cost of a reviewed output, not the cost of an unverified draft.
• Require a final approval record naming the reviewer, the sources checked, the exceptions left open, and any memory deleted or corrected.

Do not let persistent memory approve an investment assumption, close a diligence exception, characterize tenant credit, interpret a lease, make a legal or engineering conclusion, alter a deterministic underwriting model, publish a property or investor site, or send an external message. Those actions carry financial, legal, privacy, operational, or relationship consequences and need an authorized human reviewing the current record.

The most likely failure modes are stale memory, cross-deal contamination, an outdated private-company mark, source links that point to a superseded file, silent model changes during a long task, credit exhaustion that pauses a workflow, and a publishing setting that exposes material to the wrong audience. The control response is explicit scope, current-source precedence, visible model and usage logs, exception routing, and a separate publish or send approval.

Website publishing deserves its own gate. Perplexity-hosted sites can be private, shared with named people, organization-visible, or public; organization sites default to organization-only, and admins can disable public publishing. Sites deployed through Vercel are governed in Vercel rather than by Perplexity’s visibility controls. A CRE team should verify the destination, audience, source permissions, and absence of confidential or licensed material before any site goes live.

This release is worth testing where the same team repeats a research-heavy workflow and spends meaningful time reconstructing prior context. It is not a reason to delegate professional judgment or to connect every data source at once. Start with one recurring output, one approved source boundary, one accountable reviewer, and one deletion or correction procedure for memory.

The strongest signal is not that Computer can remember more. It is that remembered context can link back to a source and that organizations can inspect model usage and publishing controls. If those controls survive a deal-shaped test, the update may improve continuity. If they do not, a clean current-period packet and a disciplined human review remain the safer system.