Revision: Perplexity Computer October 5 feature update

Our verdict

Perplexity remains strongest when the job begins with research and source inspection. Computer is expanding that foundation into longer workflows, reusable context, connected APIs, and governed team deployment.

What changed

Perplexity added custom enterprise roles with granular permissions, SCIM group synchronization, and monthly group credit limits for eligible annual Enterprise contracts. Computer can now authenticate to third-party APIs through personal or project credential vaults while keeping raw secrets out of its activity and sandbox.

Brain, Perplexity’s self-improving memory system, is now available globally to Max subscribers. Perplexity says it creates wiki-style pages about important people, projects, and preferences in 15 languages, then makes that catalog available to new Computer sessions.

The same release added composable Skills for Agent API agents, Computer handoff from the Comet Assistant, claim-level source checking, a side-by-side sources panel, and natural-language tools for organizing, forking, pinning, archiving, and coordinating sessions.

Where it works best

Perplexity is a good fit for source-led research, competitive analysis, market scanning, and workflows where a cited answer must become a structured deliverable. Teams gain the most when administrators can define narrow roles, budgets, and connector boundaries.

How we would evaluate it

• Run a real research task and open every source behind its most important claims.
• Use Check Sources to identify weak support, then independently verify high-stakes conclusions.
• Test one read-only API credential before allowing write access or shared project credentials.
• Measure complete-task cost, correction rate, and human review time instead of relying on model benchmarks alone.

Who should use Perplexity?

Choose it if research quality, visible sources, and evidence-to-artifact workflows are central to your work. If your priority is broad creative collaboration or a deeply integrated office suite, compare Perplexity vs. ChatGPT Search and our AI research tools guide.

Bottom line

Perplexity is evolving from an answer engine into a governed research-and-action environment. The new controls make team deployment more credible, but source quality, credential scope, and consequential actions still need human verification.

Primary source: Perplexity changelog, July 27, 2026.

August 2026: Agent API and Router updates

Perplexity’s official changelog lists several August changes that matter to developers. Agent API presets now use stable prompt-cache keys automatically, which can reuse shared system prompts and tool definitions when requests repeat. The changelog also updates the fast preset to GPT-5.6 Luna with minimal reasoning and priority processing, and adds Gemini 3.7 Flash, Grok 4.6, NVIDIA Nemotron models, and DeepSeek V4 Flash 0731 to the Agent and Router APIs.

These entries include launch or token prices and a note that priority processing costs twice standard token rates. Prices, model availability, and preset behavior are time-sensitive, so pin a model and configuration when reproducibility matters. Before migrating, compare cache hit rates, latency, output quality, tool-call behavior, and full request cost on your own workload. This update reports Perplexity documentation; it is not an independent benchmark of the listed models.

Primary source: Perplexity API changelog.

August 2026 update: Computer in Email and expanded model access

Perplexitys official changelog now highlights Computer in Email, GPT-5.6 Terra and Luna, Grok 4.6 access, and new agent and search tools. The release direction is clear: more work can start from an email request, while users can choose or access a broader set of models inside Perplexity workflows.

Who should care: research and operations teams that already use Perplexity for source discovery may benefit from fewer context switches. However, email-triggered or agentic actions should be treated as a permissions decision. Define which mailboxes, files, and external services can be reached, and keep consequential actions behind a human approval step.

Our assessment and verification limits

The changelog confirms the announced capabilities, but availability, plan requirements, regional rollout, rate limits, and model access can vary. Perplexitys release page is not an independent benchmark, so we do not treat the announcement as proof of better answer quality. Check the cited changelog, your accounts controls, and current pricing before changing a team workflow.

Compare research workflows in Perplexity vs. Gemini and follow our source-verification guide.

September 2026 update: GLM 5.3 Flash joins Perplexity APIs

Perplexity’s official API changelog now lists perplexity/glm-5.3-flash in both the Agent API and Router API. The published rates are $0.15 per million uncached-input tokens, $0.03 per million cached-input tokens, and $0.50 per million output tokens. This gives developers a lower-cost, faster-oriented option inside the same API surface used for other routed models.

The practical implication is portability: teams can test an open-weight model through Perplexity’s endpoint while retaining the Agent API or Router API integration pattern. That does not mean every workload will be cheaper or equally reliable. Cache hit rates, output length, routing behavior, tool calls, latency, availability, and the model’s own quality profile affect the real bill and result quality.

Checks before adoption

• Confirm the exact model identifier and current pricing in the model catalog.
• Pin or log the selected model so routing changes are visible.
• Measure cache utilization and output-token growth.
• Review data handling, rate limits, and provider availability for your region.

September 2026 update: structured finance search and safer API operations

Perplexity’s official API changelog now documents Finance Search in the Agent API. The finance_search tool can return structured market and company data such as quotes, financials, earnings history, analyst estimates, segment KPIs and peer context. The model chooses the fields needed for the request, which is useful for an analyst workflow that needs numbers and source links rather than a narrative-only answer.

This is an API capability, not a promise of investment accuracy. Perplexity’s documentation says availability can depend on the API key and data coverage, and the tool is billed separately from model tokens. Use it for a first-pass research brief, then verify figures against filings or the linked source. It is not investment advice, and the output should not be treated as a trading recommendation.

What changed for builders and administrators?

• Model and tool selection: the Agent API can combine Finance Search with web search, fetch_url and third-party models for a more controlled research chain.
• Security: API keys are now shown only once at creation. Save the token immediately, use a descriptive token_name, keep it in a secret manager, and rotate before revoking the old key.
• Integrations: official documentation lists native n8n and OpenClaw integrations, AWS Marketplace billing, and a no-auth /v1/models endpoint for dynamic model selection.

Suitable users: developers building research assistants, finance teams creating reviewable briefs, and operations teams that need consolidated API procurement. Less suitable: anyone needing guaranteed real-time coverage for obscure securities, unsupervised investment decisions, or client-facing numbers without a second-source check.

Our decision rule is to start with a single ticker and a short output schema. Log the timestamp, tool calls, returned source URLs, model, token cost and missing fields. Require a human to verify every material number before it reaches a client, dashboard or order system. See the Perplexity review and comparison guides for alternatives and workflow trade-offs.

Evidence note: the Finance Search, key-management, integration and billing statements above come from Perplexity’s official documentation. AI Tools Manual has not independently tested market-data latency, coverage or forecast accuracy.

Primary sources

• Perplexity API changelog
• Agent API quickstart and Finance Search
• API key management and one-time reveal
• AWS Marketplace documentation

September 21, 2026 update: Effort Mode, GPT-6 Astra and Skills Marketplace

Perplexity’s official changelog adds Effort Mode to Computer, a Light, Standard, High or Ultra setting that lets the user choose how much effort to spend while Perplexity selects the model and reasoning level. Custom controls remain available. The release says Effort Mode is on the web for personal accounts, with mobile and desktop availability to follow, so access should be checked in the account rather than assumed from the announcement.

Eligible Pro and Max subscribers can use GPT-6 Astra for Computer tasks and supporting agents, subject to plan and data-retention eligibility. Perplexity also describes Portable Computer for supported NVIDIA DGX Spark and Windows or Linux PCs with NVIDIA RTX GPUs and at least 24 GB of VRAM; local work does not use credits, while cloud escalation requires approval. The changelog additionally introduces a public Skills Marketplace, organization-level sharing and installation controls for Enterprise teams, and Side Chat for read-only questions alongside a running Computer task.

Who benefits and who should be cautious

Effort Mode is useful for teams that want a visible cost/quality dial instead of manually choosing a model for every task. Portable Computer may fit privacy-sensitive analysis that can stay on an approved local machine, while Skills Marketplace is relevant to organizations that can govern reusable workflows. These features are not a fit when the required plan, GPU hardware, retention eligibility or admin approval is unavailable. Local-first does not automatically mean every action stays offline: cloud escalation and connected apps need their own review.

Our assessment and decision rule

Start with a low-risk task and compare Light, Standard and High on completion quality, runtime, credits and correction effort. For a local deployment, verify model downloads, hardware support, update path and whether cloud escalation is disabled by default. For Skills, inspect the package, owner, requested permissions and update process before installing it for a team. We have not independently tested Effort Mode, Portable Computer, GPT-6 Astra access or Skills Marketplace governance.

Check the official Perplexity changelog entry and the broader release index. Continue with our Perplexity profile, Perplexity vs Gemini comparison, and AI privacy checklist. The official pages support availability and feature descriptions; the fit and limitations are editorial synthesis.

September 2026: Perplexity’s API changes matter to developers, not just search users

Perplexity’s September changelog records a set of changes to its Agent API and Router API. The audience distinction matters: these are API-platform updates, not a claim that every Perplexity consumer plan exposes the same model choices or controls. The Agent API added support for GPT-6.1 Sol and Claude Sonnet 5.5; the xhigh preset now uses Claude Opus 5.5, while low, medium, and high presets use models in the GPT-6 family. Perplexity also documents changes to the fast preset and a new Fast Search path.

Migration and cost decisions

The most time-sensitive notice is that the Agent API and Router API are scheduled to retire several older OpenAI model IDs on October 24, 2026 at 00:00 UTC. Applications that pin those identifiers or use them in fallback chains should migrate before the cutoff and test the replacement behavior; the changelog says the IDs will no longer be accepted or returned by model-list endpoints after retirement. Perplexity’s September entry also prices Fast Search at $1 per 1,000 Search API requests or Agent API web_search invocations, with model tokens billed separately for Agent API requests. The older standard search path remains a distinct option. Verify current billing rates, model availability, and deprecation timing in the live docs before deployment.

Connectors and sign-in

The changelog says the remote Perplexity MCP server now supports OAuth, allowing supported clients to sign in with a Perplexity account instead of pasting an API key; users select which API organization is billed. Custom remote MCP connectors can store a credential in Perplexity so an application need not transmit it on every request. These capabilities can reduce secret-handling friction, but they also make organization selection, credential custody, connector scopes, and audit logs important review points. OAuth support depends on the MCP client, and API keys remain an option for clients without it.

Our decision rule

For production integrations, inventory pinned model IDs, preset assumptions, search modes, connector credentials, and fallback chains. Run a staging test against every replacement model, compare cost and output quality, then update monitoring and rollback procedures before removing the old route. Do not treat the new presets as performance guarantees; they can change the selected model and therefore cost, behavior, and output. For reader research workflows rather than API engineering, see our source-verified AI research guide and Perplexity vs. ChatGPT comparison.

Evidence disclosure — verify before approval

The primary source is Perplexity’s official API changelog. It supports the model IDs, preset changes, scheduled retirements, search prices, and MCP authentication changes summarized here. AI Tools Manual did not run API requests or independently validate billing, behavior, or connector security. Prices and the October 24 cutoff require rechecking before publication.

October 2026: Perplexity adds Decisions, document input, image search and search-cost changes

The Perplexity API changelog now lists several October changes. The Decisions API returns probabilities for text, JSON, or image inputs rather than generated prose. Agent API requests can include PDF, DOC, DOCX, TXT, and RTF files inline or by public HTTPS URL. A new image-search tool returns image results with source-page URLs and filters for limits, domain, format, and safe search. The changelog also updates usage reporting for sandbox extraction calls so token and cost details appear in response-level fields.

Two changes are especially important for developers because they affect cost and model behavior. Perplexity says the low, medium, and high Agent API presets now use Fast Search for web search, at a listed invocation rate reduced from $2.50 to $1.00 per 1,000 calls, with search about 800 ms faster; xhigh remains on standard web search, and callers can explicitly set search_type: "web". The same changelog schedules retirement of several older OpenAI model IDs in the Agent and Router APIs for October 24, 2026 at 00:00 UTC. These dates and price figures are transcribed from the current official changelog and require human confirmation against the live pricing and model references before publication.

What existing API users should do before switching

Inventory any explicit model IDs, fallback chains, and search presets in production. Test the new preset behavior against representative prompts, confirm whether faster search provides enough source coverage for your use case, and record the actual invocation and token costs from your own responses. If you need standard web search, explicitly select the documented setting and test it. Test document and image inputs with non-sensitive files first, validate returned source URLs, and check file-handling and retention terms before submitting confidential material.

Our assessment: the October changes broaden the Agent API from answer generation toward structured decisions and multimodal/document workflows, but the preset change means teams may see different search behavior and billed tool usage without changing their prompt. Do not assume lower total cost: model tokens, other tools, retries, and workload mix still affect spend. We have not independently benchmarked latency, decision calibration, or cost. Compare this API with the Perplexity vs. ChatGPT workflow overview and review our AI privacy checklist before sending private files.

Primary source: Perplexity API changelog, October 2026 entries. Human review required: re-confirm the $/1,000 call figures, exact effective dates and retired model IDs, preset defaults, data handling for uploaded files, and the current model/pricing documentation before applying this revision.

October 5, 2026: Computer adds recurring automations and inline visualizations

Perplexity’s October 5 changelog adds scheduled or event-triggered Computer automations that can use Memory, Skills and connected apps, build on earlier runs, and pause for review before consequential actions. It also describes interactive charts and 3D visualizations inside Computer threads, Wiley journal/book search with plan-dependent premium allowances, video generation with MiniMax H3 and Seedance 2.5 for Pro/Max personal accounts, and Portable Computer for specified AMD Ryzen AI Max Windows PCs. Full details are in the official changelog.

The most consequential distinction is between a one-off research answer and a recurring agent that can access connected services. Before enabling an automation, define its schedule, data sources, output destination and human-approval point; test on low-risk tasks first. Inline charts may help inspect data without leaving the thread, but they do not validate the underlying source or calculations. Wiley access is not the same as a full publisher subscription, and premium allowances vary by plan.

The same changelog lists GPT-6.1 Sol and Claude Opus 5.5 for eligible paid Computer users and ten Amex-curated Skills for eligible U.S. card members on Enterprise Pro/Max, with Computer Credits consumed when Skills run. These model, plan and account restrictions should be confirmed against current subscription terms before choosing a tier. The release note alone does not establish privacy retention for each connector, guarantee event-trigger delivery, or prove that generated financial analysis is accurate. We did not test the feature. For alternative sourced research, compare ChatGPT and the model ecosystem on Hugging Face. Next review: November 16, 2026.

Human review required: verify paid-plan/model eligibility, Pro/Max video limits, Wiley allowances, Amex country/card and Enterprise eligibility, credit consumption, connector permissions and any data-handling implications before publication.nn

What changed in Perplexity Computer and the API

Perplexity’s October 19 changelog groups several distinct updates. Computer on the web added an Ultrafast speed setting for GPT-6 Astra and GPT-6.1 Sol, available to Max subscribers. The changelog explicitly says Ultrafast uses more credits than Normal or Fast, so users should check their current allowance before assigning recurring work to it. This is a speed-versus-usage choice, not a general recommendation to leave every task on the fastest setting.

The same release updates the Decisions API with the pplx-decider v1.2 standard 27B and Flash 4B options. Perplexity describes probability-based classification and scoring for text, JSON, and images, including multiple questions in one pass. It also advertises up to 5× faster inference for the update. These are provider claims; teams should validate calibration, error costs, and latency on their own labeled examples before routing support tickets, moderation, or business decisions automatically.

Enterprise connector controls and decision guidance

For sales-assisted Enterprise plans, administrators can assign connector access to all members, admins only, or selected custom roles; newly added connectors can remain disabled until approved, and role changes appear in audit logs. This is a meaningful governance improvement, but it does not replace least-privilege review of each connected system or an organization’s own access-control process.

Choose Ultrafast only when the value of lower wait time justifies the extra credit use. Consider Decisions API when software needs a structured score rather than prose, but retain a human escalation path for uncertain or consequential classifications. Enterprise admins should inventory connectors, identify data owners, and test role changes before broad rollout. For the local Mac execution announcement, see our separate Perplexity Computer local-first review.

Evidence note: Feature availability, plan eligibility, credits, and performance statements above come from Perplexity’s changelog and have not been independently tested by AI Tools Manual. Next review: November 11, 2026.