09 · Plans, pricing & cost model
What each plan unlocks, why Cowork burns through limits faster than Chat, and how to talk about cost with an executive audience without over-promising.
Pricing is the fastest-changing content in this pack. The figures below are as published on claude.com/pricing and claude.com/product/cowork on 2 August 2026. Re-check before every cohort, and say "as of today" out loud when you present them.
Published plan pricing
| Plan | Price | Includes |
|---|---|---|
| Free | $0 | Neither product |
| Pro | $17/month on annual ($200 billed up front), or $20/month billed monthly | Cowork; Claude Code |
| Max | From $100/month — choose 5× or 20× more usage than Pro (the Cowork product page lists Max 20× at $200/month) | Cowork; Claude Code |
| Team | Standard seat $20/seat/month annual ($25 monthly). Premium seat $100/seat/month annual ($125 monthly). For teams of 2–150. | Cowork and the Slack connector |
| Enterprise | Seat price + usage at API rates; $20/seat, usage cost scaling with model and task. Self-serve and sales-assisted. | Admin controls, usage analytics, OpenTelemetry observability |
claude.com/pricing for the figures; claude.com/product/cowork for what each tier includes. Anthropic's own footnote: "Usage limits apply. Prices shown don't include applicable tax. Price and plans are subject to change at Anthropic's discretion."
Team offers two seat types at very different prices — $20 standard vs $100 premium (annual). On Enterprise, Claude Code on the web specifically "requires a premium seat or a Chat + Claude Code seat." So "we have Team/Enterprise" does not by itself tell you whether a given person can use a given feature. Ask which seat type.
Both products are included in the subscription — there is no separate SKU for Cowork or for Claude Code. What varies is which features your plan tier unlocks and how much usage you get before hitting limits.
What each tier actually unlocks
The full matrix is in §05 section I. The decision-relevant summary:
| If you need… | Minimum plan | Why |
|---|---|---|
| Cowork at all | Pro | Paid plans only; not on Free |
| Claude Code at all | Pro (or a Console API key / Bedrock / Vertex / Foundry) | The CLI works with third-party providers; everything claude.ai-flavoured needs a subscription |
| Computer use | Pro or Max — and nothing above | Explicitly not on Team or Enterprise |
| Dispatch | Pro or Max — and nothing above | Same constraint |
| High-volume agentic work | Max 5× or 20× | Cowork is compute-intensive; see the limits discussion below |
| SSO, analytics dashboard, server-managed settings, automated Code Review | Team | Org-level features begin here |
| SCIM, Compliance API, Enterprise Analytics API, ZDR eligibility | Enterprise | ZDR additionally requires separate enablement for qualified accounts |
| Cowork admin monitoring (OTel) | Team | Team and Enterprise only |
| Cowork cloud sessions on Enterprise | Enterprise + admin action | Off by default; owner must enable and grant via a custom role |
Two of the most demo-able features — computer use and Dispatch — are unavailable on the plans that large enterprises actually buy. If your audience is Enterprise-seated, plan your demos accordingly, and be ready to explain that these are research-preview / consumer-tier capabilities that have not yet cleared the bar for managed deployment.
The usage-limit conversation
This is where executive expectations most often break. Three documented facts, in order of importance:
| # | Fact | Source framing |
|---|---|---|
| 1 | Cowork consumes limits faster than Chat. | "Complex, multi-step tasks are compute-intensive and require more tokens to execute." |
| 2 | Auto mode costs more than Manual or Skip. | "Auto mode consumes more of your usage limit than the other modes" — the extra safety checks are themselves model calls. |
| 3 | Parallelism multiplies cost. | Claude Code: "Running several sessions or subagents at once multiplies token usage." |
Why an agentic session costs more than a chat turn
Worth explaining once, properly, because it makes every downstream cost question self-answering:
- A chat turn is roughly one model call.
- An agentic task is a loop: gather context → act → verify → repeat. Each iteration is at least one model call, and each tool result is added to the context that gets re-sent.
- Sub-agents run their own loops with their own input and output tokens.
- Safety-screened modes add classification calls on top.
- So a single "organise this quarter's contracts and summarise the exposure" brief can be dozens of model calls, not one.
"Chat is asking a question. Cowork is assigning a project. You wouldn't expect those to cost the same in a person's time either." Then the practical follow-on: match the tool to the size of the job. Don't use an agentic session to reformat one paragraph.
Cost visibility, by product
| Mechanism | Cowork | Claude Code |
|---|---|---|
| Per-request cost telemetry | ✓ api_request events carry cost_usd, token counts and cache stats (Team/Ent. OTel). Values are approximations. | ✓ OTel metrics on every provider |
| In-product usage breakdown | ◐ Not documented | ✓ /usage shows which skills, subagents and MCP servers drive your limits |
| Analytics dashboard | ✗ | ✓ Team/Enterprise |
| Programmatic analytics | ◐ Via your OTel backend | ✓ Enterprise Analytics API |
| Hard spend caps | ✗ Not documented | ✓ Per-developer day/week/month limits via the Claude apps gateway, enforced live per request |
| Team spend limits | ✗ | ✓ Documented in the Costs guide |
Documented levers for reducing cost (Claude Code)
These mostly transfer conceptually to Cowork even where the exact control doesn't exist:
- Context management — keep CLAUDE.md / instructions under 200 lines; move reference material into skills that load on demand; use
/contextto see what's consuming space. - Model selection — Sonnet handles most tasks; reserve Opus for complex reasoning.
- Extended thinking settings — tune effort to the task.
- Subagents for noisy work — a subagent that reads fifty files returns one summary, keeping the main context small.
- Prompt caching — managed automatically; note that a model switch triggers a slow uncached turn, and
/compacthas a cost. - Preprocessing hooks — shrink inputs before they reach the model.
Total cost of ownership: the dimensions that aren't the subscription
| Cost dimension | Cowork | Claude Code |
|---|---|---|
| Setup effort | Low — install the desktop app, connect folders, pick connectors | Medium — install, authenticate, write CLAUDE.md, configure permissions, possibly MCP and hooks |
| Per-user onboarding | Low — no terminal literacy required | Medium — assumes CLI comfort; Anthropic ships a champion kit and communications kit for exactly this reason |
| Ongoing maintenance | Project instructions and connectors, per user (not shareable) | CLAUDE.md, skills, hooks — but version-controlled and shared, so effort amortises across the team |
| Infrastructure | None — runs on Anthropic servers | None for local; optional gateway, CI runners, dev containers if you want them |
| Governance effort | Currently higher, because fewer hard controls exist | Higher upfront (managed settings, deny rules), lower ongoing |
| Existing cloud commitments | Not applicable | Can draw down AWS / Google Cloud / Azure spend via Bedrock, Claude Platform on AWS, Google Agent Platform, Microsoft Foundry |
Claude Code can be billed through Amazon Bedrock, AWS Marketplace, Google Cloud or Microsoft Foundry — which for many enterprises means it draws down an existing committed-spend agreement rather than creating a new vendor line. Cowork has no documented equivalent. For a CFO audience this is often the single most decision-relevant fact in the whole comparison.
Documented limits and constraints, collected
| Limit | Value | Applies to |
|---|---|---|
| Team plan seats | 2–150 | Plan constraint, not a product limit |
| Cowork file read size | 50 MB per individual file | Cowork |
| Plugin package (uncompressed) | 200 MB | Both |
| Files per plugin package | 5,000 | Both |
| Marketplace repository archive | 512 MB | Both |
| Plugins per marketplace / marketplaces addable | 500 / 25 | Both |
| Auto-memory index loaded at session start | First 200 lines or 25 KB of MEMORY.md | Claude Code |
| Recommended CLAUDE.md size | Under 200 lines | Claude Code |
| Local transcript retention | 30 days default (cleanupPeriodDays) | Claude Code |
| Dispatch permission timeout | 10 minutes, then auto-denied | Both (via Dispatch) |
| OTel model-response truncation | 60 KB | Cowork |