FutureAbility · Market Research · 30 July 2026
In seven months the "agent that does the work" category went from one product to four. This paper compares them on the dimensions that actually decide whether they're useful - local file access, projects, instruction files, memory, Agent Skills, connectors, scheduling, models and cost - with Claude Cowork held as the reference implementation, because it defined the category and every rival has since copied part of it.
All four products promise the same thing: describe an outcome, the agent plans it, executes across multiple steps and tools, and hands back finished files. They differ enormously in where the work happens, what it can touch, and how you pay for it.
Claude Cowork
Anthropic
Reference
Defined the category
The only one of the four that is vendor-neutral about where your files and tools live, and the only one where agentic use is included in a flat seat price with no metered task billing. Weakest on enterprise audit logging and on memory carrying into agent sessions.
From A$24/mo (Pro) · Max 5x A$143 · Team A$29/seat
ChatGPT Work
OpenAI
Closest rival
Architecturally near-identical
The most faithful copy: real local filesystem access on desktop with an OS-level sandbox, AGENTS.md as the instruction file, the same Agent Skills standard, parallel subagents. Undermined for planning purposes by credit metering shared with Codex and a hard split between local and cloud modes that don't sync.
From A$29/mo (Plus) · Pro 5x A$143 · Business A$29/seat + credits
Gemini Spark + Antigravity
Split in two
Neither piece is complete
Google has the best cloud persistence story (runs 24/7 with your devices off) and native write access across Workspace. But Spark is US-only, English-only, personal-account-only and Ultra-gated; local file access is macOS and Apple Silicon only. The technically superior product, Antigravity 2.0, is free and cross-platform but positioned at developers.
A$149.99/mo for Spark (AI Ultra 5x) · Antigravity free
Copilot Cowork
Microsoft
Governance play
Runs on Claude models
Licensed Anthropic's technology, adopted its skills format, imports its plugin packages - and runs on Claude Opus 4.8 and Sonnet 5. In exchange you get Purview, eDiscovery, DSPM and Agent 365. You give up local file access entirely, cross-session memory entirely, and price predictability entirely.
A$44.90/seat licence + unbounded credits
The single most important finding
Three of the four vendors have adopted Anthropic's SKILL.md Agent Skills format as an open standard. OpenAI links to the agentskills.io specification directly. Microsoft accepts .skill files and automatically converts Claude plugin packages for use in Copilot Cowork. Google's Antigravity and Gemini Spark both read SKILL.md.
This means skills you build once are largely portable across all four platforms - a genuinely unusual outcome in enterprise software, and a strong argument for investing in skill authoring rather than betting on a single vendor. It also means "does it support skills?" is no longer a differentiator. The differentiators are now local file access, memory, and the billing model.
Every cell below was checked against vendor primary documentation on 30 July 2026. Where a vendor does not publish a figure, the cell says so rather than guessing. Items marked UNCONFIRMED could not be verified in a primary source and should not be relied on commercially.
| Dimension | AnthropicClaude Cowork | OpenAIChatGPT Work | GoogleGemini Spark / Antigravity | MicrosoftCopilot Cowork |
|---|---|---|---|---|
| 1 · Product and availability | ||||
| LaunchedFirst public availability | 12 Jan 2026 research preview, macOS, Max only. Windows, web and mobile followed. Web/mobile still labelled beta.Formal GA date never announced - UNCONFIRMED | 9 Jul 2026. Announced with GPT-5.6, globally on day one, all plans including Free.Still "gradually rolling out to eligible accounts" | 19 May 2026 (I/O) for Spark; 1 Jul 2026 for local Mac files. Antigravity 2.0 also 19 May 2026.Marketing page still says "Coming soon" | 16 Jun 2026 GA worldwide. Announced 9 Mar, preview 30 Mar. Agent Mode in Word/Excel/PowerPoint GA 22 Apr 2026. |
| Platforms | Widest macOS, Windows (incl. arm64), web, iOS, Android, ChromeOS. Linux via Claude Code. |
Broad macOS, Windows, web, iOS, Android. Codex only selectable on desktop. |
Fragmented Spark: web, mobile, macOS only (Apple Silicon only, macOS 15+). Antigravity: macOS, Windows, Linux. |
Broad Browser, Outlook, Teams, M365 Copilot desktop app (Windows + Mac), iOS, Android. |
| Geographic and account limits | Few. Paid plans only - not on Free. Available in Australia. | Some. Computer Use and Record and Replay region-limited. Enterprise/Edu requires Enterprise Key Management enabled. | Severe. Spark is US only, English only, 18+, personal Google accounts only - explicitly not work or school accounts. Personal Intelligence excluded in EEA, UK, Switzerland. | Some. Anthropic-supported regions only; Anthropic models excluded from the EU Data Boundary. Tier-1 languages. |
| 2 · Local file access - the decisive dimension | ||||
| Reads and writes files on your actual computer | Yes You connect folders; every call is permission-checked and brokered through the desktop app. Works even for remote sessions, provided the desktop app is open. |
Yes - desktop only "Open a local folder or project and grant access only to the files the task needs." Web and mobile cannot access local files at all. |
Partial Spark: macOS only, since 1 Jul 2026, folder-level consent, revocable. Web/mobile Spark has none. Antigravity: full local access, all three OSes. |
No Microsoft states it plainly: "Cowork can't access or edit files stored locally on your device. It works with files in OneDrive and SharePoint." Cannot delete OneDrive/SharePoint files either. |
| Where code actually runsExecution sandbox | Two modes. Local: dedicated Linux VM isolated by Apple Virtualization.framework or Hyper-V. Remote (now default): temporary sandbox on Anthropic servers, destroyed at session end, mandatory egress proxy, no access to private or cloud-metadata addresses. | Explicit composer choice: Work locally or Cloud. Local uses OS-native enforcement - macOS Seatbelt, native Windows sandbox, or bubblewrap on Linux/WSL2. Cloud is a managed isolated environment with a hostname allowlist. |
Spark is cloud-based by design - "keeps working even when you close your laptop." Remote browser plus remote code execution. Antigravity runs locally with terminal sandboxing in preview on macOS/Linux, "coming soon to Windows". | Temporary isolated environment inside the Microsoft 365 service boundary. Microsoft markets the absence of local access as the security feature: "Cloud hosting means files are not stored locally." |
| Permission model | Three modes: Manual / Auto / Skip. Auto runs a prompt-injection and exfiltration safety review on every action and consumes more usage. Auto will never approve new folder access, file deletion, or creating scheduled tasks. Deletion always needs explicit approval in every mode. | Three modes: Ask for approval (default), Approve for me (a reviewer agent adjudicates), Full access. Sandbox and approvals are separate controls. Underlying modes read-only, workspace-write, danger-full-access. Admins constrain centrally via requirements.toml. |
Antigravity has the most granular model: per-action read_file, write_file, command, mcp, read_url, execute_url, unsandboxed, each set to Deny / Ask / Allow with Deny > Ask > Allow precedence. Spark uses folder linking in the sidebar. |
Inherits the user's own M365 permissions; sensitivity labels flow through. Approval dialogs carry risk-level indicators, scopable per recipient or domain - but persist only for the current session. Admins can pre-authorise for zero-touch runs. |
| Local browser control | Computer use is a research preview on Pro and Max, with no sandbox between Claude and your screen. Claude in Chrome available; Anthropic "strongly advises against" it for sensitive data. | Built-in browser in the desktop app plus Computer Use on macOS/Windows, per-app approval and an always-allowed list. Cannot automate terminals, cannot approve OS security prompts. Region-limited. | Spark drives a remote browser server-side, retaining your session cookies between runs. Local browser control was roadmapped at I/O but not shipped. | Drives a hidden tab in your local Microsoft Edge using existing SSO - credentials never leave the device. Edge only, web only, off by default, admin-gated. Hands back for CAPTCHA and MFA. |
| 3 · Projects and workspaces | ||||
| Project construct | Cowork Projects (9 Apr 2026) - own instructions, scheduled tasks, context source and memory. Separate from Chat Projects, which hold knowledge files plus custom instructions. | Projects - "smart workspaces" holding chats, reference files and custom instructions. Unlimited projects. Chat or Work can be launched from within a project's context. | No single construct. Gems (persona + instructions + knowledge, all tiers incl. free), Gemini Notebooks (formerly NotebookLM), and Antigravity Projects - which can span multiple folders, e.g. frontend and backend repos. | Copilot Notebooks is closest - up to 300 files, but only the first 100 references are used for grounding. Cowork sessions are tasks, not projects. Copilot Pages is a persistent canvas. |
| File limits | Chat Projects fit the 200K context (~150k words); RAG mode expands capacity roughly 10× on paid plans. Cowork projects point at a local folder - no file count limit. | Per project: Free 5 · Go/Plus 25 · Pro/Business/Enterprise/Edu 40. Max 10 files uploaded at once. Workspace agents get 512 MB per file, 10 GB total. | Gems: 10 knowledge files, 100 MB each. Gemini chat: 10 files per prompt, video ≤2 GB, others ≤100 MB. Code folder up to 5,000 files / 100 MB. | Notebooks up to 300 files but 100-reference grounding ceiling. Attachments must be under 200 MB. "Download All" zip capped at 50 files / 500 MB. |
| Project-scoped instructions | Yes - Cowork Projects have their own instructions; folder instructions can be updated by Claude itself. | Yes - project instructions override global custom instructions. | Yes - Gem instructions. Antigravity uses workspace rules in .agents/rules/. |
Not documented for Notebooks. UNCONFIRMED |
| Sync and sharing | Weakest point. Cowork projects are desktop-only, stored locally, no cloud sync, no project sharing even on Team/Enterprise. Sessions cannot be shared. Live artifacts can be shared org-wide on Team/Enterprise. | Split. Cloud chats sync across web, mobile and desktop. Local chats stay on that computer and local outputs never appear on web or mobile. Shared projects support Edit / Chat roles, up to 100 people. | Good. Cloud-native, so syncs by default. Workspace Studio agents are shareable "similar to sharing a Google Doc". | Good. Cloud-native, stored in SharePoint Embedded containers. Skills shareable to specific users with re-share control. |
| 4 · Instruction files - the CLAUDE.md question | ||||
| Filesystem instruction file | CLAUDE.md - but documented for Claude Code, not Cowork. Load order: managed policy → ~/.claude/CLAUDE.md → ./CLAUDE.md → ./CLAUDE.local.md, all concatenated, not overridden. Guidance: keep under 200 lines. @path imports to 4 hops.In Cowork proper the equivalents are UI-based: Global instructions and Folder instructions |
AGENTS.md - a documented open spec at agents.md. "Codex reads AGENTS.md files before doing any work." Global ~/.codex/AGENTS.md → project root walked down to CWD → merged root-down, nearer files winning. Budget 32 KiB. Fallback filenames configurable. Desktop custom instructions are stored in the global AGENTS.md. |
GEMINI.md in Antigravity - global at ~/.gemini/GEMINI.md, workspace rules in .agents/rules/. 12,000 characters per rule file. Activation modes per rule: Manual, Always On, Model Decision, or Glob. Consumer Gemini has no file-based config - UI settings only. |
No equivalent. No always-loaded workspace-root instruction file. Per-user custom instructions under Settings › Personalization. Developer layer uses declarative agent manifests with an 8,000-character instruction ceiling. Microsoft's own docs name-check CLAUDE.md and AGENTS.md - but only for developer Copilot. |
| 5 · Memory | ||||
| Cross-session memory in the agent | Projects only Anthropic's own words: "What Claude remembers about you in chat doesn't carry into Cowork sessions yet. Within Cowork, memory is supported in projects only." A real gap in the benchmark product. |
Yes, two systems Cloud ChatGPT memory is used by Work on web. Separately, local Codex memories - off by default, stored as files in ~/.codex/memories/, generated in the background from idle chats with secret redaction. |
Yes Spark explicitly draws on "chats, Personal Intelligence, Connected Apps, skills, location". Three layers: personal context, saved info, and Personal Intelligence across Google apps. |
None Microsoft's own responsible-AI card: "Within a single conversation, Cowork maintains context… Memory does not persist across conversations." Persistence must be faked using custom skills. |
| User-viewable and editable | Fully. Entries organised by category at Settings › Memory. Edit, delete, pause, or reset. Import/export to and from other AI tools (experimental). Incognito chats excluded. | Yes. An editable memory summary you type into directly, plus a Sources control per response showing which memories, files and instructions shaped the answer. Caveat: the summary "will not include everything ChatGPT remembers." | Yes, but deletion is two-step - you must delete the chats containing a memory and disconnect the source app, or "Gemini might still use the info." | Preview only. Copilot memory is still preview and Frontier-gated; stored in a hidden folder in the user's Exchange mailbox. Tenant control "Enhanced personalization" is on by default. Copilot Studio agent memory is deleted after 28 days of inactivity. |
| 6 · Agent Skills and extensibility | ||||
| Skills supportReusable instruction + script packages | Originated the standard Folder with required SKILL.md (YAML name ≤64 chars, description ≤1024) plus optional scripts/, references/, assets/. Three-level progressive disclosure: metadata ~100 tokens always loaded, body under 5k tokens on trigger, resources on demand. agentskills.io confirms the format "was originally developed by Anthropic, released as an open standard." |
Adopted the standard OpenAI states outright: "OpenAI Skills follow the Agent Skills open standard, so you can download them from one product and install them in another." Identical folder layout, identical progressive disclosure, plus an optional agents/openai.yaml. Discovery across CWD, repo root, $HOME/.agents/skills and /etc/codex/skills. |
Adopted the standard Antigravity: "Skills are an open standard… a folder containing a SKILL.md file," and links agentskills.io. Spark accepts SKILL.md or a zip. But Spark skills are crippled: plain-text files only - no PDF, DOCX, XLSX or images - and "scripts can't perform actions or requests on external websites." |
Adopted the format Literal SKILL.md with YAML frontmatter at OneDrive /Documents/Cowork/skills/, discovered automatically each session. 13 built-in skills. Accepts .skill files and auto-converts Claude plugin packages. Limits: 50 custom skills per user, body budget under 5,000 tokens.Participation in the agentskills.io standard itself: UNCONFIRMED |
| Plugins / distribution | Plugins bundle skills, connectors, slash commands and sub-agents. Team and Enterprise owners can run private plugin marketplaces (24 Feb 2026); org-managed plugins can be auto-installed and not edited. | Plugins distributed through a "universal plugin directory shared by ChatGPT and Codex". Uploaded skills are security-scanned and may be flagged Needs Review or Blocked. Five separate admin RBAC toggles plus an invocations dashboard. | Workspace Studio agents are the reusable unit for business tenants - natural-language built, shareable, with webhooks and custom Apps Script steps. | Cowork plugins = skills + connectors. 20 skills and 10 connectors per package. Admin-deployed plugins cannot be removed by users. Microsoft's caveat: custom skills "aren't validated by Microsoft". |
| Skills sync across surfaces | No. Skills do not sync across surfaces; claude.ai skills are per-user zip uploads with no central admin management. | No. "Personal Skills must be added separately on desktop and web/mobile and do not sync automatically." Standalone skills don't work in Work on the web at all - only plugin-bundled ones. | Partial. Spark skills are web and mobile; the macOS app is not listed as supporting them. | Yes - stored in OneDrive, so available wherever Cowork runs. |
| MCP support | Full. MCP is Anthropic's own protocol. Connectors Directory spans claude.ai, desktop, mobile, Claude Code and Cowork. Custom remote MCP available on every plan including Free. Local MCP servers do not run in remote sessions. | Full. Available on every plan - Plus, Pro, Business, Enterprise and API. Custom MCP apps on Plus and above (not Free or Go). Warning: "every MCP server adds more context to your messages and uses more of your limit." | Yes, at every layer. Consumer Gemini supports custom MCP servers - but "only available to users with Gemini Spark access and within Spark tasks." Antigravity has per-project MCP config. Gemini Enterprise offers BYO-MCP via Apigee. | Constrained. Implements the MCP Apps Extension (SEP-1865) for sandboxed HTML widgets, but rejects initialize, tools/list, sampling/*, elicitation/*, notifications and server-pushed streaming. Widget calls capped at 60/min per conversation. |
| Connectors | Verified plus community directory - Google Drive, Gmail, Slack, DocuSign, CRM, calendar. Claude for Small Business (13 May 2026) added QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, M365 and Slack at no extra fee. | "60+ apps" - Slack, Google Drive, SharePoint, GitHub, Atlassian, Teams, Salesforce, Adobe Acrobat, Gmail, Outlook. Four-level permission model with "Important actions" as the default. Enabled by default on Business, disabled by default on Enterprise/Edu. | All tiers including free. Google Workspace, GitHub, Spotify, Instacart, OpenTable, Zillow, Canva, Dropbox. Spark can write across Calendar, Gmail, Drive, Docs, Sheets and Slides. Workspace Studio adds Asana, Jira, Mailchimp, Salesforce and arbitrary webhooks. | 9 partners at GA (Enosix/SAP, Harvey, LSEG, Miro, monday.com, Moody's, Morningstar, S&P Global Energy, TeamsMaestro), 8 coming. Dynamics 365 Sales and Customer Service on by default. Separately 100+ Graph connectors, read-only by default. |
| 7 · Automation, scheduling and parallelism | ||||
| Scheduled and recurring tasks | Yes. /schedule or the Scheduled sidebar. Hourly, daily, weekly, weekdays or manual. Each run is a full session with connector, skill and plugin access. Runs remotely - no device needs to be online. Tasks needing local files must run locally. |
Yes. Standalone or in-chat, with RFC 5545 RRULE scheduling and event triggers. Not on Free. Local scheduled tasks require the machine on and the app running. They run unattended with your default sandbox - OpenAI flags full-access unattended runs as "elevated risk". | Yes, confusingly twice. Gemini chat scheduled actions (max 10 active, needs AI Plus and above) and Spark Schedules (time and condition triggered, max 15 concurrent tasks). Google has to publish a note explaining they're different things. | Yes. Natural-language scheduled prompts, capped at 5 per user. Event-driven tasks added July 2026 - triggers on inbound email matching criteria, Teams channel or chat messages, optionally only on @mention. Default to draft-and-approve. |
| Runs with your machine off | Yes for remote sessions and scheduled tasks. Phone notification on completion. But local file access needs the desktop app open. | Yes in Cloud mode. No for local mode or local scheduled tasks. | Yes - the deepest of the four. "Spark keeps working in the background even when you close your laptop or lock your phone." Marketed as 24/7. | Yes. "Minutes to hours (autonomous execution)". Progress preserved across disconnects; soft and hard pause, resume, cancel. |
| Parallel subagents | Yes - "breaks complex work into smaller tasks and coordinates parallel workstreams". Sub-agents are bundleable inside plugins. | Yes - "spawning specialised agents in parallel". Built-in default, worker, explorer plus custom TOML agents. On the web the subagent panel is read-only - you cannot stop or steer one. Proactive delegation needs the Ultra intelligence level. OpenAI warns subagents "consume more tokens". |
Antigravity yes - dynamically defined subagents, isolated git worktrees to prevent conflict, JSON hooks. Spark no - custom sub-agents were roadmapped at I/O, not shipped. | Effectively no. Cowork plans sequentially. The only documented multi-agent pattern is Sonnet + Opus Advisor, where Opus reviews Sonnet's output. Real orchestration lives in Copilot Studio, a separate product.Parallel fan-out primitive: UNCONFIRMED |
| 8 · Models | ||||
| Models available | claude-fable-5 (1M context, GA 9 Jun 2026), claude-opus-5 (1M), claude-sonnet-5 (1M), claude-haiku-4-5 (200k). Cowork has a model selector.Exact selectable list in Cowork not published - UNCONFIRMED |
GPT-5.6 in three variants: Sol (deepest reasoning), Terra (everyday default), Luna (fast and cheap), plus Sol Pro on Pro. GPT-5.5 Instant, GPT-5 Thinking Mini, legacy models. GPT-Live powers desktop voice; GPT-Image-2 for images. Context: 256K reasoning on Plus, 400K on Pro. | Gemini 3 Flash-Lite, 3 Flash and 3 Pro on all tiers including free - tiers differ by usage volume, not model access. Gemini 3.5, 3.5 Flash, 3.6 Flash, Omni, 3.1 Pro all appear on live Google pages concurrently. Spark runs on Gemini 3.5 Flash. Deep Think is Ultra-only. Antigravity also ships Claude Sonnet and Opus 4.6. | Runs on Anthropic. Picker offers Auto, Claude Sonnet 5, Claude Opus 4.8, Claude Fable 5 (preview, requires data retention), Sonnet + Opus Advisor, and GPT-5.5 via Azure AI Foundry. Anthropic became a formal Microsoft subprocessor on 7 Jan 2026. Disabling Anthropic models disables Cowork. Microsoft's own "Cowork 1" model was announced but has not shipped. |
| Context window | 1M tokens on Fable 5, Opus 5 and Sonnet 5. | Reasoning total 256K (Plus), 400K (Pro). Instant 27K - 128K. | 32k free · 128k AI Plus · 1,000,000 on AI Pro and Ultra. Deep Think 192k. | Inherits the Claude models' windows. Prompt input capped at 250,000 characters. |
| 9 · Pricing | ||||
| Entry individual planAUD per month, ex GST | Free A$0 - no Cowork Pro A$24 annual / A$29 monthlyUS$17 / US$20 converted |
Free A$0 - limited Work on desktop Go A$11 (may include ads) Plus A$29US$8 / US$20 converted |
Free A$0 AI Plus A$7.99 AI Pro A$32.99Google publishes AU prices |
No individual path. Cowork requires a Microsoft 365 Copilot licence on a business tenant. |
| Higher individual tiersThe heavy-use plans, AUD per month | Max 5x - A$143 Max 20x - A$287US$100 / US$200 converted. 5x and 20x refer to usage headroom above Pro |
Pro 5x - from A$143 Pro 20x - A$287US$100 / US$200 converted. Pro adds GPT-5.6 Sol Pro and unlimited standard models |
AI Ultra 5x - A$149.99 AI Ultra 20x - A$329.99Published AU prices. Spark requires Ultra - and a US personal account |
No consumer tier. The step up is Microsoft 365 E7 at A$148.20, which bundles E5, Copilot, Entra Suite and Agent 365. |
| Business and teamAUD per seat per month, ex GST | Team standard A$29 annual / A$36 monthly Team premium A$143 annual / A$179 monthly Enterprise A$29 + usage at API rates Cowork included at every tier, no meteringTeam spans 2 to 150 seats |
Business A$29 annual / A$36 monthly Enterprise and Edu - contact sales Credits charged on topBusiness starts at 2 seats. Business per-seat limits equal Plus |
Workspace Starter A$9.90 Standard A$19.80 Plus A$30.90 Gemini Enterprise from A$30 to A$43Gemini bundled into Workspace at no extra cost since Jan 2025. AI Expanded Access add-on price not published |
Copilot Business (max 300 seats) A$31.40 list / A$26.91 promo M365 Copilot add-on A$44.90 Business Standard with Copilot A$35.20 Business Premium with Copilot A$47.90 Agent 365 A$22.40 Zero Cowork credits included at any tier |
| Basis of the AUD figure | Converted. Anthropic publishes USD only and adds 10% GST unless an ABN is registered. | Converted. OpenAI publishes no AUD list; prices render client-side by region. Australia is an Enterprise data-residency location. | Published by Google on the AU storefront. Workspace prices are annual-commitment; flexible runs about 19% higher. | Published by Microsoft on the AU storefront, ex GST. Suite list prices rose 1 July 2026. |
| Billing model for agentic workThe commercial question that matters most | Flat seat, no metering Cowork is included in the plan. No separate credit meter, no per-task charge. Limits are the plan's 5-hour and weekly caps. Extra usage can be bought as usage credits if you want it, but nothing forces you to. |
Metered credits "ChatGPT Work usage inside ChatGPT uses the same pricing, credits, and usage limits as Codex." Credits are token-denominated - GPT-5.6 averages 5-40 credits per message. Plus and Pro users can buy more credits without upgrading. |
Compute-based Google moved from prompt counts to a compute-used model on 17 May 2026. Limits refresh every 5 hours up to a weekly cap. Exceed it and you're downgraded to smaller models rather than blocked. Pay-as-you-go AI credits exist for Antigravity and Flow, "coming soon" to the Gemini app. |
Unbounded metering Critically: "No Cowork entitlements are included with the Microsoft 365 Copilot subscription." Copilot Credits at A$0.0143 each (US$0.01) cover model use, context retrieval, tool calls and runtime. Prepaid P3 plans discount 5-20% but unused credits expire at the end of the annual term. |
| Published agentic usage limits | No Cowork-specific quota published. Plan 5-hour session and weekly limits apply, with a separate weekly cap on Opus. Cowork burns usage faster than chat, and Auto mode faster than Manual. | Fully published, per model, per 5-hour rolling window. GPT-5.6 Sol: Plus 15-90 local messages, Pro 20× 300-1800. Business per-seat limits are identical to Plus. Enterprise on flexible pricing has no fixed limit - "usage scales with credits." | Spark: 15 concurrent tasks. Chat scheduled actions: 10 active. Gemini Agent has a monthly request cap whose number Google does not publish. Free users are throttled first when capacity is tight. | No fair-use quota at all - only a money cap. The sole per-user ceiling is the admin-configured credit limit; hit it and "they lose access to agents and services for the rest of the month." In-flight tasks complete and "aren't billed (at Microsoft's sole discretion)". Concurrency, timeout and per-user quotas: UNCONFIRMED. |
| Cost per taskVendor-published estimates | Not applicable - no per-task charge. | 5-40 credits per GPT-5.6 message. Rate card published per model, e.g. Sol 125 credits per 1M input tokens, 750 per 1M output. | Not published per task; compute-based. | Light 100-300, Medium 300-700, Heavy over 700 credits. Note these bands changed in six weeks - the June guide said 70-200 / 400-600 / over 1500. Microsoft's own estimator assumes 125 / 500 / 1200 credits per prompt on Opus 4.8. Implied: roughly A$203/user/month for a knowledge worker and A$327 for a technical user - entirely on top of the A$44.90 licence. |
| 10 · Governance and compliance | ||||
| Audit and compliance | The real weakness Anthropic documents that Cowork activity is not captured in audit logs, the Compliance API, or data exports - OpenTelemetry only. EDR tools cannot inspect the VM or remote sessions. |
Strong Compliance Logs Platform covers skills and app calls. Admin action control (all / read-only / custom), Parameter Constraints using regex and ranges, domain restrictions, RBAC. Data residency in 10 regions including Australia. |
Mixed Workspace admin controls exist, but Spark is personal-account-only so sits outside tenant governance entirely. Gemini Enterprise adds VPC-SC, CMEK, data residency, HIPAA and FedRAMP High. |
Deepest of the four Purview audit log, DSPM, eDiscovery, IRM, Data Lifecycle Management, Communication Compliance, sensitivity labels applied at every step, Copilot data residency, Agent 365 identities for agents. But DLP was not ready at GA - listed as "coming soon". |
| Training on your data | Commercial plans are not trained on. Incognito chats excluded from memory. | No training on Business, Enterprise, Edu or API data by default. But on Free, Go, Plus and Pro, OpenAI "may use information accessed from projects/apps to train our models" if "Improve the model for everyone" is on. | Personal Intelligence requires Web and App Activity, YouTube History and Gmail smart features to be on. Consumer terms apply. | Enterprise terms, DPA, EDP and the Customer Copyright Commitment. Claude Fable 5 preview requires data retention by the model provider, with a persistent banner. |
| 11 · Setup burden | ||||
| What it takes to get running | Lowest Subscribe, install the desktop app, connect a folder. No tenant, no Azure subscription, no admin. A sole trader can be productive in ten minutes. |
Low Subscribe and install. Business starts at 2 seats. Enterprise/Edu additionally requires Enterprise Key Management enabled. |
Low but gated Antigravity is free and instant. Spark requires a US personal Google account and an AI Ultra subscription - a blocker for an Australian business, not a setup problem. |
Highest Seven distinct admin steps: assign the Copilot USL, enable Anthropic models tenant-wide, link or create an Azure subscription (Global or Billing admin only), create a spending policy, set monthly and per-user limits, enable usage-based-billing discovery, optionally enable browsing. Trap: once a billing method is set on a policy you cannot change it - you must delete and recreate. |
| 12 · Fit for a 5-50 person business | ||||
| Verdict | Best fit Flat, predictable per-seat cost with agentic use included. No IT department required. Vendor-neutral about your file store and tools. Team plan spans 2 to 150 seats. Claude for Small Business (13 May 2026) added the SMB connectors an Australian consultancy actually needs - QuickBooks, HubSpot, Canva, DocuSign, M365, Google Workspace. |
Strong second Cheap entry, 2-seat minimum, no tenant required. The catch is that credit metering makes monthly cost genuinely hard to forecast, and Workspace Agents - the shareable team agent - are unavailable on every consumer plan including Pro at A$287. |
Not viable today Spark is US-only and personal-account-only, so an Australian business cannot buy it for staff. Antigravity is free, excellent and cross-platform - but it is a developer tool, not a knowledge-work agent. Revisit when Spark reaches Workspace accounts. |
Only if already all-in on M365 Licensing is fine - no seat minimum, and Copilot Business is capped at 300 seats. The problems are the setup burden and unbounded metered cost on top of the seat price.Important: whether Cowork actually runs on the cheaper Copilot Business SKU is NOT CONFIRMED by Microsoft - all Cowork docs say "Microsoft 365 Copilot licence". Verify in a live tenant before advising a client. |
Of all the dimensions above, this is the one that changes what the product is for. It splits the four cleanly into two camps.
Claude Cowork and ChatGPT Work both operate on the actual filesystem, brokered through a desktop app, with an OS-enforced sandbox and folder-level consent. That means they can work the way you already work: the client folder on your Mac, the messy spreadsheet in Downloads, the PDF someone emailed you. Nothing has to be uploaded first.
Copilot Cowork cannot do this at all, and Microsoft is explicit and unapologetic about it: files live in OneDrive and SharePoint, full stop. Microsoft frames the absence as a feature - "cloud hosting means files are not stored locally… your tasks keep running even when your laptop is off." For a large regulated enterprise that is a defensible trade. For a consultancy that receives client files by email and works in local folders, it means a permanent upload-and-download tax on every task.
Google sits awkwardly between the two. Gemini Spark gained local file access on 1 July 2026 - but macOS only, Apple Silicon only, US only, Ultra only. Antigravity 2.0 has the most sophisticated permission model of any of the four, with per-action Deny/Ask/Allow precedence, and runs on all three operating systems for free. It just isn't aimed at the person who wants their invoices sorted.
A caveat that applies to all four, and to Cowork most of all
Anthropic's own guidance asks users to "monitor Claude for suspicious actions that may indicate prompt injection." Simon Willison's response, on the day Cowork launched, remains the sharpest criticism in the category: "I do not think it is fair to tell regular non-programmer users to watch out for 'suspicious actions that may indicate prompt injection'!"
Google's documentation is even blunter about its own product, warning that Spark could "run a command that can be malicious or install compromised software, to achieve persistent data exfiltration through arbitrary code execution." OWASP's position at Infosecurity Europe 2026 was that prompt injection remains an "unresolved problem" within generative AI architecture. Every product in this table gives an agent your files and your credentials. None of them has solved this.
This is the one dimension where Claude Cowork does not lead - and it's worth being honest about, because it's the feature people assume works.
"What Claude remembers about you in chat doesn't carry into Cowork sessions yet. Within Cowork, memory is supported in projects only."
Anthropic · Claude support documentation · retrieved 30 July 2026
"Within a single conversation, Cowork maintains context… Memory does not persist across conversations."
Microsoft · Copilot Cowork responsible-AI application card · retrieved 30 July 2026
OpenAI has the most developed answer: cloud memory used by Work on the web, plus a separate local Codex memory store written as files under ~/.codex/memories/, generated in the background from idle chats with secret redaction, and off by default. It also ships the best transparency control of the four - a Sources button under any response that shows exactly which memories, files and instructions shaped that answer, with per-item correction.
Google's memory is the most ambitious in scope - Personal Intelligence reaches across Gmail, Calendar, Drive, Photos, Contacts, Search, Maps and YouTube history - and does carry into Spark sessions. But it excludes work and school accounts entirely, is unavailable in the EEA, UK and Switzerland, and doesn't apply to Gems, meaning Google's "project" construct and Google's "memory" construct don't compose with each other.
For Tony's purposes the practical answer is: put anything that must always apply into instruction files, not memory. Anthropic's own guidance for Claude Code says the same thing about CLAUDE.md, and OpenAI says it explicitly of AGENTS.md: "Treat memories as a helpful recall layer, not as the only source for rules that must always apply."
Anthropic published the Agent Skills specification on 18 December 2025. Seven months later it is the de facto standard across all four vendors - and Microsoft will import your Claude plugins directly.
| Vendor | Reads SKILL.md | Instruction file | Notable |
|---|---|---|---|
| Anthropic | Authored the standard | CLAUDE.md | Published as an open standard; reference implementation. Private plugin marketplaces on Team/Enterprise. |
| OpenAI | Yes - explicitly | AGENTS.md | Links agentskills.io in its own docs; skills "portable between products". Universal plugin directory shared with Codex. |
| Yes, both surfaces | GEMINI.md | Antigravity links agentskills.io. Spark skills restricted to plain text and cannot make external network requests. | |
| Microsoft | Yes - and converts Claude plugins | None | "If your file is a Claude or other compatible plugin package, Cowork automatically converts it into a publishable package." |
What this means commercially
Skill authoring is now the most transferable AI investment available. A well-built skill - a FutureAbility document generator, a client-onboarding workflow, a D365 filing routine - will run on all four platforms with modest adaptation. That materially reduces platform risk, and it argues for building the skill library aggressively rather than waiting to pick a winner.
It also means the reverse: "supports skills" has stopped being a reason to choose one product over another. Choose on local file access, memory, governance and billing model instead.
A caveat on the widely-repeated adoption statistics. Claims that Microsoft and OpenAI shipped support "within 48 hours" and that "32 tools" read the same SKILL.md by March 2026 trace to secondary blogs rather than press of record, and should not be quoted as fact. What is verifiable is that agentskills.io itself states the format "was originally developed by Anthropic, released as an open standard," and lists Cursor, GitHub Copilot, VS Code, OpenAI Codex, Gemini CLI, OpenCode and Goose among adopters.
All figures in Australian dollars, ex GST. Google and Microsoft publish AU prices directly; Anthropic and OpenAI publish USD only, converted here at 1 USD = A$1.4325 (30 July 2026). Anthropic adds 10% GST unless an ABN is registered. Modelled for a ten-person Australian consultancy. This is where the four products diverge most sharply, and where the headline seat price is most misleading.
10 seats, annual, ex GST
A$287/mo
A$29/seat · premium seats A$1,433/mo · converted from USD
Predictable. One line item.
10 seats, annual, ex GST
A$287/mo
A$29/seat, plus credits · converted from USD
Base predictable, heavy agentic use is not.
10 seats, annual, ex GST
A$198/mo
Workspace Standard A$19.80/seat · published AU price
Cheapest on paper, the agent is not available to you.
10 seats, ex GST
A$449/mo
A$44.90/seat - before any Cowork usage · published AU price
Realistic all-in: A$449 licence plus about A$2,030-3,270 usage.
The cost-blowout precedent is real, recent, and documented
GitHub moved Copilot to usage-based billing on 1 June 2026. Developers reported monthly bills going from A$42 to A$1,074, and from A$72 to A$4,298 (US$29 to US$750 and US$50 to US$3,000). GitHub's own research found agentic tasks consume roughly 1,000× more tokens than a single-turn query. Uber reportedly exhausted its entire 2026 AI coding tools budget by April and capped spend at about A$2,149 per user per tool (US$1,500) (widely reported; primary source not located).
Microsoft's published Cowork task-cost bands moved substantially within six weeks - the June guide said light 70-200 and heavy over 1,500 credits; the July guide says light 100-300 and heavy over 700. Date-stamp any Copilot Cowork cost figure you put in front of a client. Independent analysis suggests context retrieval, not the model, drives most of the cost - which means cost scales with how much tenant data the agent has to search, not with how hard the thinking was.
Attributed opinion and hard adoption data. Vendor marketing excluded.
"Microsoft is seeking to commoditize its complements, but Anthropic has a point of integration of their own; it's good enough that Microsoft is making a new bundle on top of it."
Ben Thompson, Stratechery · 10 March 2026 · on Microsoft building Copilot Cowork on Anthropic's technology
"Anthropic knows that it needs to get to the enterprise data - and is going for the integration of the productivity suites. That is a good first step, but the key one will be to get to the transactional data."
Holger Mueller, Constellation Research · 24 February 2026
"Claude Cowork is regular Claude Code wrapped in a less intimidating default interface and with a filesystem sandbox configured for you without you needing to know what a 'filesystem sandbox' is."
"I would be very surprised if Gemini and OpenAI don't follow suit with their own offerings in this category."
Simon Willison · 12 January 2026 · the second prediction was correct within seven months
"A long-running agent doesn't behave like a chatbot: It behaves like a distributed system, and distributed systems demand orchestration, identity, and context discipline that most companies have never built."
Forrester · The State Of Agentic AI In 2026: Companies Are Chasing, Few Are Catching · 3 June 2026
| Source | What it measures | Result |
|---|---|---|
| Menlo Ventures 2025 State of GenAI |
Enterprise LLM API market share | Anthropic 40% OpenAI 27% Google 21% Anthropic was 12% in 2023; OpenAI was 50% |
| Ramp AI Index April - May 2026 |
Share of US businesses with paid vendor spend | Anthropic 34.4% · OpenAI 32.3% - the first Anthropic lead in the dataset. Overall business AI adoption 50.6%. Anthropic reportedly wins ~70% of first-time business AI purchases. |
| Forrester Security Survey 2026 |
Security decision-makers naming agentic AI as a concern | 49% |
| Anthropic Economic Index June 2026 · vendor-published |
What 1.2M anonymised Cowork sessions were used for | Business process and operations 33.4% · content creation and copywriting 16.4% |
| CMU TheAgentCompany | Real office tasks completed by the best agents | 30.3% - a useful antidote to vendor demos |
| Gartner / MIT | Project failure projections | Gartner: over 40% of agentic AI projects cancelled by end-2027. MIT: 95% of GenAI pilots fail to scale. |
ZDNet ran both agents against real files under the headline "I let ChatGPT Work and Claude Cowork loose on my files - only one made me nervous." The verdict was that results and strengths were broadly similar, but "Claude currently feels considerably safer." Wording via syndication - zdnet.com could not be fetched directly, so treat the quote as approximate.
Beyond that, independent adversarial testing of these products is thin. I found no Gartner or Forrester research scoped specifically to SMB agentic suites, no Forrester Wave for agentic assistants, and no Australian analyst commentary of any kind. Much of the post-launch coverage of ChatGPT Work in particular simply restates OpenAI's own framing. Treat vendor reliability and speed claims as unproven.
The change-management point worth repeating to clients
Steve Hasker, CEO of Thomson Reuters, on why enterprises stall even with good tools: "A general counsel's office, a law firm, a tax accounting firm or an audit firm needs to rewire the processes… I think it's 18 months away before that sort of change management catches up with the standard of the tool."
Forrester's framing of the same problem: "Three-quarters of enterprise leaders tell us they're adopting agentic AI. Only a small minority have it running in meaningful production beyond 'agentish' chatbots." And on the hidden blocker: "Underneath all of it sits the trust tax. Every autonomous action has to be logged and defensible to an auditor, and right now that cost is too high." That last point is precisely where Claude Cowork is weakest and Copilot strongest.
Assessed for an Australian professional services firm of 5 to 50 people, with no dedicated IT function.
Claude Cowork remains the recommendation, and the reason is billing, not capability. It is the only product of the four where agentic work is included in a flat per-seat price with no metered task charge. For a firm that cannot absorb an unpredictable monthly variance, that single fact outweighs most feature differences. Add to it the lowest setup burden of the four - no tenant, no Azure subscription, no admin console - real local file access on Windows and macOS, the widest platform coverage, and a Team plan that starts at two seats and scales to 150.
ChatGPT Work is a credible alternative and closing fast. Architecturally it is the most faithful copy of Cowork: genuine local filesystem access with OS-level sandboxing, AGENTS.md as a documented open instruction-file spec, the same skills standard, parallel subagents, and better audit logging than Anthropic offers. Two things count against it for a small firm. Credit metering shared with Codex makes the monthly bill hard to forecast. And the local/cloud split is genuinely awkward - local chats stay on one machine and never sync, so a two-person team working on the same client can't see each other's local sessions.
Copilot Cowork is only rational if you are already fully committed to Microsoft 365 and value governance above cost certainty. What you get is real: Purview, DSPM, eDiscovery, IRM, Agent 365 identities for agents, sensitivity labels flowing end to end. What you give up is substantial: no local file access at all, no cross-session memory at all, sequential rather than parallel execution, and a metered cost with zero included allowance that Microsoft's own estimator puts at A$203-327 per user per month on top of the A$44.90 licence. The setup is seven admin steps including linking an Azure subscription - and one of them, the billing method on a spending policy, cannot be changed once set. There is a real irony here: Copilot Cowork runs on Claude Opus 4.8 and Sonnet 5, so you are paying Microsoft a premium and a metered markup to use Anthropic's models with less capability than Anthropic's own product.
Gemini is not currently purchasable for this use case. Gemini Spark is restricted to US personal Google accounts, explicitly excluding work and school accounts, and requires AI Ultra at A$149.99 per month. An Australian business cannot deploy it to staff. Google Antigravity 2.0 is free, cross-platform, and arguably the most technically sophisticated agent harness of the four - dynamic subagents, isolated git worktrees, the best permission model, scheduled tasks, and it will happily run Claude Sonnet and Opus 4.6 - but it is a developer tool. Worth watching closely: if Spark reaches Workspace accounts and Australia, the economics change quickly, because Gemini is bundled into Workspace at no extra cost.
Stay on Claude for the core agentic workflow, and keep investing in the skills library - that investment is now portable to all four platforms, which removes most of the lock-in risk that would normally argue for hedging. Maintain a Microsoft 365 Copilot capability for client-facing Power Platform work, because that is where the clients are, but price Cowork usage separately and explicitly in any proposal rather than treating it as a per-seat add-on. Independent analysis calls treating Cowork as just another per-seat add-on "the most common and most expensive mistake expected in 2026 Microsoft renewals" - a useful line in a client conversation, and a genuine advisory opportunity.
Two things to verify before quoting any of this to a client
1. Whether Copilot Cowork runs on the cheaper Copilot Business SKU. Every Microsoft Cowork document says "Microsoft 365 Copilot licence" or "USL". Only secondary sources claim Copilot Business qualifies, and the Copilot Business FAQ says "same capabilities" without ever naming Cowork. Test it in a live tenant.
2. Australian dollar pricing for Anthropic and OpenAI. Neither publishes AUD prices. Anthropic bills in USD and adds 10% GST unless an ABN is registered. OpenAI's pages render prices client-side by region and no AUD list could be retrieved from a primary source. Every A$ figure circulating for these two is a third-party conversion. Only Google and Microsoft publish real Australian prices.
Vendor primary documentation, checked individually on 30 July 2026. Secondary and analyst sources listed separately.