The July 2026 Buyer's Guide to AI Worker Agents

TL;DR

• The category split in 2026 into three shapes: cloud-native agents that build finished documents (ChatGPT Work, Copilot Cowork), desktop agents that touch your local files (Claude Cowork, Perplexity Personal Computer), and developer/self-hosted platforms (Gemini Enterprise Agent Platform, Devin, OpenClaw). Pick by where your work lives, not by which model benchmarks highest.
• For most UK marketing agencies and SMBs, the sweet spot is Claude Cowork or ChatGPT Work at roughly £16-20/month per user ; both now ship agent features on their ~$20 entry plans. Enterprise stacks (Copilot Cowork at £24.70-30/user/mo on top of M365, Gemini Enterprise from $21-50/seat, Perplexity Computer at $200/mo) cost far more and suit larger, governed rollouts.
• Treat every agent run like a paid meeting, not a free chat message. Usage is metered on all of them, single long tasks can burn most of a weekly allowance, and the real cost is deployment and review, not the licence.

Key Findings

• ChatGPT Work launched July 9, 2026, merging Chat, Work and Codex into one desktop app, powered by the new GPT-5.6 family (Sol, Terra, Luna). It ships finished spreadsheets, slides, docs and web apps (“Sites”). Atlas browser is being retired August 9, 2026.
• Claude Cowork (launched January 12, 2026, GA April 9) is the most mature desktop knowledge-work agent, now on macOS and Windows, web and mobile, running on Sonnet 5 (new default since June 30) with Opus 4.8 and the Mythos-class Fable 5 available.
• The Anthropic model lineup shifted materially: Fable 5 and Mythos 5 launched June 9, were suspended June 12 under a US export-control directive, and Fable 5 was restored July 1. Current generally available models are Sonnet 5, Opus 4.8, Haiku 4.5, and Fable 5.
• Pricing has converged around $20/month for entry agent access across ChatGPT, Claude, Manus and others ; but credit/usage metering makes real costs unpredictable.
• Notable additions to the category: Manus, Genspark (with its “Claw” AI employee), plus the developer-focused Devin and self-hosted OpenClaw (now ~350k GitHub stars, stewarded by a non-profit after creator Peter Steinberger joined OpenAI).

Why This Category Matters Now

For three years, AI tools mostly answered questions. In 2026 they started doing the work. The shift has a name inside every major vendor’s marketing ; “agent,” “cowork,” “digital worker” ; and a shared mechanic: you describe an outcome, the tool breaks it into steps, works for minutes or hours, and hands back a finished artifact rather than a chat reply.

This matters for agencies and SMBs for one blunt reason: the work these tools now complete ; competitor research, deck building, data cleanup, campaign briefs ; is the work junior staff and freelancers bill for. The market reaction to Anthropic’s January 12, 2026 Claude Cowork launch made the stakes concrete: what Jefferies strategist Jeffrey Favuzza dubbed the “SaaSpocalypse” saw the S&P 500 Software & Services Index shed roughly 25% between the launch and February 23, with Bloomberg describing a $285 billion rout across software, financial-services and asset-management stocks ; Thomson Reuters fell 18% (a record single-day loss) and RELX dropped 14% (its biggest since 1988). Microsoft, meanwhile, reported in its 2026 Work Trend Index that active agents across the Microsoft 365 ecosystem grew 15 times year over year, and 18 times in large enterprises. The category is no longer experimental.

But the hype outruns reality. A Gartner press release dated June 25, 2025 ; authored by Senior Director Analyst Anushree Verma and based on a January 2025 poll of 3,412 webinar attendees ; predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing “escalating costs, unclear business value or inadequate risk controls.” Gartner separately estimated that only about 130 of the thousands of vendors claiming “agentic AI” are genuine. And a Gartner note dated February 5, 2026 pushed back on the sell-off narrative, calling Cowork and its plug-ins “potential disrupters for task-level knowledge work but…not a replacement for SaaS applications managing critical business operations.” These tools burn usage budgets fast, hallucinate confidently, and need human review on anything that feeds a decision. This guide is written to help you buy the right one and avoid that failure rate.

Answer-first definition: An AI worker agent is software that takes a goal, plans the steps, uses tools and files, and completes a multi-step task with minimal supervision ; returning finished work (documents, spreadsheets, code, web apps) rather than just an answer. That is the line between a chatbot and an agent.

Comparison Table

Individual Write-Ups

ChatGPT Work (OpenAI)

What it is: An agent inside ChatGPT, launched July 9, 2026, that takes an outcome, gathers context across connected apps and files, breaks the job into steps, and works for hours to produce finished spreadsheets, slides, documents and shareable web apps (“Sites”). It launched alongside the GPT-5.6 model family ; Sol (flagship), Terra (balanced), Luna (fast/cheap) ; and a merged desktop app that unifies Chat, Work and Codex. OpenAI is retiring its Atlas browser on August 9, 2026, folding browser-agent capability into ChatGPT and Codex.

Primary use case: Cloud-first research, synthesis and document generation ; competitor analysis into a deck, data gathered from multiple sites into a formatted spreadsheet.

Ideal user: Solo operators, marketers and SMB teams already living in ChatGPT who produce documents and analyses all day. The free tier now includes a real agent running GPT-5.6 Terra.

Pricing: Work is included on all paid plans and even Free (with hard caps). Plus is $20/month (about £16 + 20% VAT, roughly £19 all-in). The two Pro tiers are $100 (5x limits) and $200 (20x limits, unlocks “ultra” four-parallel-agent mode). Business is $25/seat monthly or $20 annual. No annual discount on Plus.

Security/data model: Cloud execution in a virtual environment; built on ChatGPT Enterprise security with an “auto-review” step where advanced models check important connected-tool actions before they run. OpenAI’s launch materials claim adversarial red-teaming saw auto-review block 100% of attempts to extract protected data (a vendor claim). Business/Enterprise plans don’t train on your data by default.

Integrations: Launched with a directory of over 1,400 connectable apps you @-mention mid-task ; HubSpot, Gong, Slack, email, calendar, CRM, Drive/SharePoint. Replaced the App Directory with a Plugin Directory.

Standout differentiator: Multi-agent parallelism (GPT-5.6 spawns concurrent subagents), “Plan mode” to approve steps before execution, and genuine free-tier agent access.

Key limitation / criticism: ChatGPT Work shares one usage pool with Codex, and launch-week users watched single substantial tasks eat most of a usage window; “ultra” burns roughly four times as fast. The strain was real enough that OpenAI’s Codex lead temporarily removed the 5-hour usage limit on Plus, Pro and Business on July 12 after GPT-5.6 Sol demand doubled traffic in 48 hours. A ZDNet hands-on test (“I let ChatGPT Work and Claude Cowork loose on my files ; only one made me nervous”) concluded: “My testing reveals similar results, similar strengths, and one major reason Claude currently feels considerably safer right now” ; a trust gap, not a proven vulnerability. Public publishing of “Sites” is unavailable in the UK, EEA and Switzerland at launch ; directly relevant for a London agency.

Claude Cowork (Anthropic)

What it is: Anthropic’s knowledge-work agent, launched January 12, 2026 as a research preview and generally available since April 9. It runs the same agentic architecture as Claude Code, pointed at non-coding work. On desktop it reads, edits and organises local files directly; it works across connected apps like Slack and Google Drive, runs scheduled tasks, coordinates subagents, and produces polished Excel, PowerPoint and formatted documents with citations back to source files.

Primary use case: Local-file and cross-app knowledge work ; point it at a folder of drafts and downloads, get a structured report or renamed/sorted files back.

Ideal user: Solo professionals through enterprise teams who work in documents, data and files. Especially strong for writing-heavy work.

Pricing: Included in all paid plans. Pro is $20/month (about £16; $17/month billed annually). Max is $100 (5x) and $200 (20x). Team Standard is $25/seat, Team Premium $125/seat ($100 annual) ; note Cowork and Claude Code are excluded from Team Standard. Enterprise from $20/seat plus usage. Current models: Sonnet 5 (default since June 30), Opus 4.8, Haiku 4.5, and the restored Fable 5 (credits-based billing after July 7).

Security/data model: Desktop-first with explicit per-folder permissions and confirmation for destructive actions. Team/Enterprise exclude training at the contract level. Anthropic has a stronger privacy-by-default reputation; the M365 permission model borrows your credentials, can only see what you can see, and requires admin consent before write access.

Integrations: Slack, Notion, Asana, Linear, Jira, Monday, ClickUp, Microsoft 365 (read since February, write added July: draft/send email, manage calendar, edit OneDrive/SharePoint), Google Drive, Zoom connector, plus 11 open-sourced knowledge-work plugins.

Standout differentiator: Direct local file access, task scheduling, Skills auto-loading, and the most mature agent in the category with six months’ live enterprise use before ChatGPT Work shipped.

Key limitation / criticism: New capabilities roll out to $100+ Max plans first (the July mobile/background features went Max-first). Heavy per-task usage ; one user reported a single presentation task consuming roughly $45 of usage, and a five-hour session eating ~15% of a Pro weekly allowance. Prompt-injection and destructive-action risks remain, as with all agents. Cowork’s deepest local features are desktop-only.

Copilot Cowork (Microsoft)

What it is: An agent inside Microsoft 365 Copilot, announced March 9, 2026 as the centerpiece of “Wave 3,” generally available via the Frontier program from late March. Built in collaboration with Anthropic (it incorporates the technology behind Claude Cowork), it turns a request into a plan and executes it across Word, Excel, PowerPoint, Outlook and Teams, grounded in “Work IQ” ; a context layer that reads your emails, files, meetings and org graph.

Primary use case: Multi-app, long-running work inside Office ; calendar triage, meeting prep, deep research into Excel workbooks, product-launch planning across apps.

Ideal user: Enterprises already standardised on Microsoft 365 with governance requirements.

Pricing: Requires an M365 Copilot licence ; about £24.70/user/month on annual commit (list) or £30 monthly (Microsoft’s global $30 headline), on top of a qualifying M365 base subscription. It is not sold standalone. Microsoft 365 E7 (“Frontier” suite) bundles Copilot, Agent 365 and E5 security at $99/user/month. Note UK M365 base prices rose 8-33% from July 1, 2026.

Security/data model: Runs within M365’s security and governance boundaries ; identity, permissions and compliance policies apply by default; actions are auditable and reversible. Agent 365 provides a governance control plane.

Integrations: Deep native M365 (Outlook, Teams, SharePoint, OneDrive, Excel), plus Dynamics 365 and Power Apps; multi-model routing across Anthropic Claude and OpenAI models.

Standout differentiator: Work IQ’s organisational-graph context ; it resolves “email my leadership team” without you naming names ; plus enterprise governance.

Key limitation / criticism: You can’t buy it without M365 and a Copilot licence, so the effective cost is high. UK IT commentators flag a “governance nightmare” and a training gap ; untrained users save near zero; a February 2026 issue saw Copilot ignoring sensitivity labels. The product story (Copilot, Cowork, Agent Mode, Agent 365, Work IQ, Frontier, E7) is dense and hard to explain to buyers.

Perplexity Computer & Personal Computer

What it is: Perplexity Computer, launched February 25, 2026, is a cloud-based “digital worker” that orchestrates roughly 19-20 frontier AI models (Claude, Gemini, GPT and others), spins up subagents, and runs multi-step workflows in an isolated cloud sandbox. Personal Computer, announced at the Ask 2026 dev conference (March 11) and rolled out April 16, extends this onto a user’s Mac ; an always-on local agent (recommended on a Mac mini) with access to local files and native apps, activated by pressing both Command keys.

Primary use case: Recurring multi-source research, competitive intelligence and automated “research → build → document” workflows; always-on background execution for operators.

Ideal user: Power users, founders and ops teams who orchestrate complex, multi-hour workflows and want model-agnostic execution.

Pricing: Computer launched exclusive to Max ($200/month, $2,000/year, 10,000 monthly credits). Personal Computer opened to Pro ($20/month) and Enterprise on May 7, though Pro users burn credits quickly. Enterprise Max adds security controls; a per-seat enterprise tier around $325/seat/month has been reported. Claude Opus 4.7 was the default orchestrator as of April, with GPT-5.5 added as an option.

Security/data model: Cloud sandbox per task; the Mac version runs actions in an isolated sandbox with on-device authorisation, kill switch, and auditable/reversible actions. SOC 2 Type II, SAML SSO on enterprise. Reviewers note the “local” privacy story still routes data to Perplexity’s servers and third-party model providers, and no external audit has been cited.

Integrations: 400+ business apps including Snowflake, Databricks, HubSpot, Salesforce, SharePoint, GitHub, Linear, Gmail, Outlook, Notion; @computer inside Slack; MCP connectors; plus Agent, Sandbox, Search and Embeddings APIs and “Perplexity Finance.”

Standout differentiator: True multi-model orchestration (Perplexity builds no foundation model, so it acts as a neutral router) plus the always-on Mac mini architecture no competitor matches at the consumer level.

Key limitation / criticism: $200/month is steep, and per-task credit costs are opaque ; no confirmation before a task runs. Reviewers report no clarifying questions (it makes assumptions and sprints), integration gaps, and note that OpenClaw offers similar capability for API costs alone. Most users get ~80% of the value from Pro at one-tenth the price.

Gemini Enterprise Agent Platform (Google)

What it is: Announced at Google Cloud Next ‘26 (April 22, 2026) and generally available the same day, it’s Google’s full-stack platform to build, scale, govern and optimise agents ; the explicit evolution of Vertex AI (all future Vertex roadmap ships here). It provides Agent Studio (low-code), an upgraded Agent Development Kit, Agent Runtime, Agent-to-Agent Orchestration, Agent Identity, Registry, Gateway, Observability, Simulation and Evaluation. The Gemini Enterprise app is the employee “front door,” with Agent Designer, Inbox, Projects/Canvas and long-running agents.

Primary use case: Enterprises building and governing fleets of custom agents across business systems.

Ideal user: Developers and IT teams at mid-to-large organisations, especially Google Cloud/Workspace customers.

Pricing: Gemini Enterprise editions are Business $21, Standard $30 (annual; $35 flex), Plus $50 (annual; $60 flex) per user/month, and Frontline (custom, 150-seat minimum). Custom-agent usage (model tokens, compute) is billed separately on top. A free trial is available; existing Vertex customers see the rebrand automatically. Google also announced a $750M partner innovation fund.

Security/data model: Single governance control plane; Agent Identity gives every agent a trackable identity; secure-by-design sandboxed workspaces; customer data excluded from model training under enterprise terms. Agentic Defense combines Google threat intelligence with Wiz.

Integrations: 200+ models via Model Garden (Gemini 3.1 Pro, Gemini 3.1 Flash Image/Nano Banana 2, Lyria 3, Gemma 4) plus Anthropic’s Claude Opus 4.7/Sonnet/Haiku; partner agents from Oracle, Salesforce, ServiceNow, Box, Workday and others; A2A protocol, MCP, BYO-MCP.

Standout differentiator: Google’s full-stack argument ; it owns silicon (TPUs), frontier models, the cloud platform and Workspace distribution ; plus genuine model choice including Claude.

Key limitation / criticism: It’s a developer/IT platform, not a turnkey agent ; most SMBs and agencies won’t build here. Separate compute billing makes budgeting harder, and enterprises remain cautious about giving agents access to sensitive data.

Devin (Cognition)

What it is: An autonomous AI software engineer that runs in the cloud for hours, handling codebase migrations, bug backlogs and feature work. Devin Review now runs an automatic security review on every pull request.

Primary use case: Delegating real software engineering ; long-horizon coding, migrations, debugging.

Ideal user: Startup-to-Fortune-500 engineering teams.

Pricing: Devin 2.0 cut the floor from $500 to about $20 (Core), with Ultra tiers using the most powerful models (Ultra costs ~40% more than the default agent). Kimi K2.7 and GLM 5.2 don’t consume quota for Pro/Max/Teams until early July; Sonnet 5 uses ~30% less quota than Sonnet 4.6 through August 31, 2026.

Model lineup: Integrates GPT-5.5, Claude Opus 4.8, Claude Sonnet 5, Kimi K2.7, GLM 5.2. Claude Fable 5 was added June 9, removed June 12 under the export directive, and restored July 1 ; it tops Cognition’s FrontierCode benchmark. Devin Ultra uses the most capable available models.

Standout differentiator: The first agent Cognition’s research lead (SVP Silas Alberti) trusts to run unattended overnight and deliver production-ready code (with Fable 5).

Key limitation / criticism: Coding-only; premium cost for heavy autonomous use; benchmark leadership on Cognition’s own FrontierCode drew “self-serving” criticism.

OpenClaw (open-source, self-hosted)

What it is: A free, MIT-licensed, self-hosted AI agent framework that runs as a persistent local “Gateway” on your own machine, connecting to messaging channels (WhatsApp, Telegram, Slack, Discord and many more). Launched November 2025 as “Clawdbot,” renamed twice, it crossed 100k GitHub stars in January and roughly 350k by April 2026 ; among the fastest-growing repos in history. Creator Peter Steinberger (founder of PSPDFKit) joined OpenAI in February; a non-profit foundation now stewards it with OpenAI sponsorship.

Primary use case: Private, self-hosted personal/developer automation ; overnight coding agents, daily digests, webhook-triggered debugging ; with full data ownership.

Ideal user: Developers and power users who want no vendor lock-in and are comfortable with terminal setup. Features added through 2026: ContextEngine/Active Memory, dual-engine model routing, Discord voice, xAI OAuth, a compliance plugin, OpenTelemetry observability, and OpenRouter smart provider switching.

Pricing: Free (MIT). You pay only for API calls (roughly $5-200/month depending on models and usage) plus optional VPS hosting ($5-15/month). Managed hosts exist (e.g., OneClaw from $9.99/month; Blink Claw ~$45/month bundling 200+ models).

Security/data model: Local-first ; data, memory and context stay on your device with bring-your-own-key. But the attack surface is significant: prompt injection is the top risk, and unvetted third-party skills can carry code and supply-chain risk. Audit skills before installing.

Standout differentiator: Full ownership, auditability and no per-task billing ; NVIDIA built its enterprise NemoClaw on the same codebase.

Key limitation / criticism: Setup requires terminal use; security is your responsibility; community skill registry is not fully vetted.

Also worth knowing: Manus and Genspark

Manus is a general-purpose agent that runs inside a virtual computer ; strong for individual research and web tasks, weaker on team features. Free tier (300 daily credits); Pro from $20/month (4,000 credits) up to $200 (40,000 credits); Team from $20/seat (2-member minimum). Credit burn is unpredictable and there’s no shared workspace or persistent memory. Test the monthly plan before making an annual commitment.

Genspark is an all-in-one “super agent” workspace ; Sparkpages (structured research), AI Slides, AI Sheets, and a “Call For Me” agent that places real phone calls. Its “Claw” AI employee (launched March 2026) takes tasks via WhatsApp/Telegram/Slack/Teams. Free (100 credits/day), Plus $24.99/month (10,000 credits), Pro $249.99/month. Reviewers praise the slides and research output but flag opaque credit consumption, weak support.

How to Choose: A Decision Framework

Work through these in order.

1. Where does your work live?

• In cloud SaaS apps (CRM, Slack, Drive) → ChatGPT Work or Copilot Cowork.
• In local files and messy folders on your machine → Claude Cowork or Perplexity Personal Computer.
• In code → Devin (autonomous) or Claude Code / Cursor (in-IDE).
• Everywhere, and you want to own the stack → OpenClaw.

1. What’s your existing platform? If you’re all-in on Microsoft 365, Copilot Cowork removes integration friction (at a price). If you’re on Google Workspace and want to build agents, Gemini Enterprise. Otherwise stay platform-neutral with ChatGPT or Claude.
2. Who’s the user? Solo/SMB → the $20 tiers of ChatGPT Work or Claude Cowork. Enterprise with governance needs → Copilot Cowork, Gemini Enterprise, or Perplexity Enterprise. Developers → Devin, OpenClaw, or Gemini’s platform.
3. How much do you trust it to act? If it will send emails, update systems or touch production, prioritise permission models and audit logs. Reviewers currently rate Claude Cowork’s permission model and Copilot’s M365 governance highest; ChatGPT Work’s cloud auto-review is newer.
4. Can you afford the burn? Budget for metered usage, not a flat licence. Start bounded (a weekly competitor briefing) before allowing consequential actions.

Benchmarks that should change your decision: If a single routine task consistently eats more than ~10-15% of your weekly allowance, move up a tier or change tools. If “Sites”-style public publishing matters and you’re UK-based, ChatGPT Work can’t do it yet. If more than one person needs shared context, avoid Manus.

Frequently Asked Questions

What is an AI worker agent? An AI worker agent is software that takes a goal, plans the steps, uses your tools and files, and completes a multi-step task with little supervision ; returning finished work (documents, spreadsheets, code, web apps) instead of just a chat answer. The distinction from a chatbot is action: a chatbot tells you how; an agent does it.

What is the best AI agent for a small business in 2026? For most SMBs and agencies, Claude Cowork or ChatGPT Work at about £16-20/month per user is the best starting point ; both now include agent features on their entry paid plans, both produce finished deliverables, and neither requires an enterprise contract. Choose Claude Cowork if your work is local files and writing; ChatGPT Work if it’s cloud apps and research.

ChatGPT Work vs Claude Cowork ; which is better? They differ by where the agent operates. ChatGPT Work is cloud-native and app-centric ; better for web research, multi-app orchestration and shipping polished output across 1,400+ connectors. Claude Cowork is desktop-first and filesystem-aware ; better for local file pipelines, writing quality, and controlled Microsoft 365 workflows. Many professionals run both; at roughly $20 each, that’s about $40/month.

How much do AI worker agents cost? Entry agent access is about $20/month (£16 + VAT) on ChatGPT Plus, Claude Pro and Manus Pro. Mid tiers run $100-200/month. Enterprise stacks cost more: Microsoft 365 Copilot is ~£24.70-30/user/month on top of M365; Gemini Enterprise is $21-50/user/month plus usage; Perplexity Computer’s Max tier is $200/month. On top of licences, all charge metered usage, and deployment/training is the larger real cost.

Are AI agents safe to give access to my files and email? They can be, with guardrails. Desktop agents (Claude Cowork, Perplexity Personal Computer) use scoped per-folder permissions and confirmation for destructive actions; enterprise tools (Copilot Cowork, Gemini Enterprise) add audit logs and identity controls. The unsolved risks across all of them are prompt injection (malicious instructions hidden in a file or web page) and destructive actions from ambiguous prompts. Start with read-only, bounded tasks and add write access gradually.

Which AI agent runs locally rather than in the cloud? Claude Cowork (desktop, direct local file access), Perplexity Personal Computer (always-on on a Mac, recommended Mac mini), and OpenClaw (fully self-hosted, MIT-licensed) are the local-first options. Cloud-only tools include ChatGPT Work, Copilot Cowork, Perplexity’s original Computer, and Gemini Enterprise.

What happened to Claude Fable 5? Fable 5 (a Mythos-class model above Opus) launched June 9, 2026, was suspended June 12 under a US government export-control directive after researchers documented a jailbreak, and was restored globally July 1 after Anthropic trained an improved safety classifier that blocks the jailbreak in more than 99% of cases. It’s now available on Pro, Max, Team and Enterprise plans (credits-based billing after July 7) and in Devin, where it tops the FrontierCode benchmark.

Do I need coding skills to use an AI worker agent? No, for the mainstream tools. ChatGPT Work, Claude Cowork, Copilot Cowork, Perplexity Computer, Manus and Genspark are built for non-developers ; you describe outcomes in plain language. You need technical skills only for developer platforms (Gemini Enterprise Agent Platform, Devin’s deeper features) and self-hosted OpenClaw, which requires terminal setup.

Can AI agents replace employees? Not wholesale in 2026. They compress tasks ; research, drafting, data cleanup, file organisation ; that used to take hours into minutes, and they can run recurring workflows unattended. But they hallucinate, need human review on anything consequential, and Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027. Treat them as leverage for your team, not a replacement for judgment.

Which AI agent is best for developers? For autonomous coding, Devin runs unattended for hours and now adds an automatic security review on pull requests. For in-IDE work, Claude Code and Cursor lead. For building and governing custom agents, Gemini Enterprise Agent Platform. For a private, self-hosted setup with no vendor lock-in, OpenClaw.

Is there a free AI worker agent? Yes. ChatGPT Work is included on the free ChatGPT tier (with hard usage caps) running GPT-5.6 Terra. Manus and Genspark offer free daily-credit tiers. OpenClaw is free and open-source (you pay only API and hosting costs). Claude Cowork, Copilot Cowork and Perplexity Computer require paid plans.

What’s the difference between Perplexity Computer and Personal Computer? Perplexity Computer is the cloud version ; it runs workflows in an isolated cloud sandbox and orchestrates ~19-20 models. Personal Computer is the local extension that runs on your Mac (ideally an always-on Mac mini), with access to local files and native apps like Mail, Messages and Calendar. Computer launched Max-only; Personal Computer opened to the $20 Pro plan on May 7, 2026.

Recommendations

If you’re a London agency or SMB deciding now:

1. Start with a two-tool trial. Put ChatGPT Plus (£16+VAT) and Claude Pro (~£16) side by side for a month, ~£35 total. Run the same real brief through both ; a competitor analysis into a deck. Let output quality and usage burn pick the winner.
2. Default to Claude Cowork for file- and writing-heavy work; ChatGPT Work for cloud-app research and mixed output. Note the UK caveat: ChatGPT Work’s “Sites” public publishing is blocked in the UK/EEA, so don’t build a client-facing publishing workflow around it yet.
3. Only move to Copilot Cowork if you already pay for Microsoft 365 and need Office-native execution with governance ; budget the full ~£24.70-30/user/month add-on plus the July 2026 base-price rises, and invest in training, or adoption will stall.
4. Skip Perplexity Max ($200/mo) unless you run genuinely multi-hour, recurring workflows. Test Personal Computer on Pro ($20) first.
5. For developer/automation work, trial OpenClaw on a spare machine before paying for a managed agent ; but audit every community skill and lock down permissions.

Staged rollout: (1) Pick one bounded, repeatable workflow. (2) Run it read-only for two weeks and measure time saved against usage burned. (3) Add write access and connectors only once the pattern is reliable. (4) Expand to a second workflow. Keep senior judgment in the approval loop throughout.

Thresholds that change the plan: If a routine task eats more than ~10-15% of your weekly allowance, upgrade a tier or switch tools. If you need shared team context, rule out Manus. If governance/audit is mandatory, shortlist Copilot Cowork, Gemini Enterprise or Perplexity Enterprise. If data must stay on-premises, choose OpenClaw or a desktop-first agent.

Caveats

• This market changes weekly. Model names, prices and features moved repeatedly through 2026 (Anthropic’s Fable 5 suspension/restoration, Perplexity opening Personal Computer to Pro, Microsoft’s July price rises). Verify current pricing on each vendor’s own page before committing ; all figures here are as of mid-July 2026.
• GBP figures are approximate. OpenAI and Anthropic bill in USD (with VAT and card FX on top); Microsoft and Google publish GBP list prices that drift with exchange rates. Expect £16-20 all-in for “$20” plans after VAT and FX.
• Usage metering makes real cost hard to predict. Credit and token billing means a heavy month can cost far more than the headline licence. Vendor productivity claims (time saved, ROI multiples) are largely self-reported and need testing against your own numbers.
• Some claims are vendor-supplied. Microsoft’s agent-growth metrics, OpenAI’s red-team blocking rates, and case studies like the “$225,000 of marketing tools replaced in a weekend” are company figures not independently audited. Treat them as directional.
• Security is an open question for agentic tools. Prompt injection remains unsolved industry-wide; the ZDNet comparison described a “trust gap” rather than a specific flaw. Independent audits of data routing (especially Perplexity’s “local” claims) have not been published.