Anthropic

Best for: Picking the right Claude subscription — the Pro/Max gap is the most-confused decision in the Anthropic lineup.

• Free ($0/mo) — Sonnet only, ~30 messages every few hours, no Projects, limited file uploads
• Pro ($20/mo) — Sonnet 4.6 default, Opus 4 access, ~5× Free usage, Projects, Memory, Research, Connectors, Claude Code on a shared quota
• Max ($100 or $200/mo) — 5× or 20× Pro caps, priority access, much higher Claude Code usage, expanded Research limits
• Team ($25/seat/mo annual billing, or $30/seat/mo monthly, ≥5 seats) — Pro features + shared workspace, central billing, admin console, Team Memory
• Enterprise (custom) — SSO/SAML, SCIM, audit logs, no training on your data by default, custom retention, expanded context
• Claude Code is included on every paid tier; heavy coding agent users almost always justify Max

Limitations: The $100 vs $200 Max split is genuinely confusing — most individuals are best served by Pro or Max $100; Max $200 is for full-time agentic coding. API usage is billed separately on every tier.

Best for: Choosing the right model for the job — the three tiers map cleanly to depth, balance, and speed.

• Opus 4.7 (flagship, released Apr 16 2026) — deepest reasoning, longest agentic runs, hardest coding and research; $5 input / $25 output per 1M; 200K context (1M on Opus 4.7)
• Sonnet 4.6 — the default workhorse: near-Opus quality on most tasks at a fraction of the cost; $3 input / $15 output per 1M, 200K context
• Haiku 4.5 — fast and cheap; $1 input / $5 output per 1M; ideal for classification, extraction, and high-volume agent loops
• Extended Thinking (opt-in budget of internal reasoning tokens) is available on Opus and Sonnet — billed as output
• Prompt caching gives up to 90% off cached input tokens (5-minute or 1-hour TTL); Batch API gives a flat 50% discount on async jobs
• All Claude 4.x models accept text, image, and PDF input; vision quality is strong on charts, screenshots, and handwriting

Limitations: Reasoning tokens make cost prediction harder — budget output volume to be much higher than the visible answer. Older Claude 3.x models still appear in some integrations; migrate to 4.x for new builds. Opus 4.7 ships with a new tokenizer producing up to ~35% more tokens for the same input, so real per-request cost can rise even at the unchanged headline rate.

Best for: Autonomous coding tasks — multi-file edits, test running, refactors, and long-horizon agent work in a real codebase.

• Local terminal agent with full read/write/run access to a chosen directory; reads CLAUDE.md automatically for project context
• IDE extensions for VS Code and JetBrains; web app at claude.ai/code; CLI works on macOS, Linux, and Windows (PowerShell)
• --dangerously-skip-permissions enables fully autonomous work (use only in sandboxed dirs or worktrees)
• Skills, slash commands, hooks, and MCP servers all extend the agent — install once, reuse everywhere
• Routines schedule repeating jobs on Anthropic-hosted cloud agents; they run even when your laptop is off
• Included on every paid plan; usage draws from your plan quota, with much higher limits on Max

Limitations: Quota visibility in the CLI is improving but still imprecise. Routines are powerful but easy to misuse — pair with verification before letting one push code or send messages.

Best for: Screen-aware persistent assistant for knowledge workers — file and task automation without writing code.

• Sees any open app natively — no screenshots, no manual paste-in needed
• Stable filesystem and terminal access via a built-in Desktop Commander connection that does not drop on tab switch
• Built for review and sustained desktop work: Word docs, Excel sheets, PDFs, email threads, file folders
• Same Claude account and Memory as claude.ai — context carries across surfaces
• Available on macOS and Windows; the closest Anthropic equivalent to OpenAI ChatGPT Agent for desktop tasks

Limitations: Not a coding agent — for repo-level work, use Claude Code. Cowork is in active development; some integrations are still rolling out.

Best for: Driving real web UIs — navigating, clicking, filling forms, copying credentials, and automating dashboards that have no API.

• Reads the live DOM, clicks, types, and scrolls inside any tab
• Excellent for SaaS setup tasks: Vercel, Supabase, Stripe, Cloudflare, GitHub configuration
• Connects to claude.ai as an MCP tool — your chat conversation can drive the browser
• Available on Pro and above; sandboxed-by-default with explicit permission prompts for destructive actions

Limitations: Heavily JS-driven SPAs and aggressive bot-detection (some banking, enterprise SSO) still trip it up. Treat as a copilot for browser tasks, not an unattended worker.

Best for: Spreadsheet work in place — formulas, cleanups, modeling, and natural-language transformations on real workbooks.

• Runs inside Excel as a sidebar; reads the active sheet and ranges, writes formulas and values back
• Strong on data wrangling: dedupe, normalize, pivot, fuzzy-match, fill-down, and explain-what-this-formula-does
• Pairs with Cowork when the task spans multiple files or apps; Excel add-in alone is sufficient for single-workbook work
• Anthropic's direct answer to Microsoft Copilot for Excel — does not require a Microsoft 365 Copilot license

Limitations: Power Query / VBA support is limited compared to native Excel automation. Available on Pro and above; Enterprise admins can deploy at the tenant level.

Best for: Connecting any AI client to any tool or data source through a single open protocol — the USB-C of agent integrations.

• Open spec (modelcontextprotocol.io) — any client can speak to any server, with no Anthropic lock-in
• Now supported by Claude (chat, Code, Cowork), OpenAI (Responses API, Codex), Google (Gemini), Microsoft (Copilot Studio), Cursor, Windsurf, Zed, and more
• Hundreds of community + vendor servers exist for GitHub, Slack, Postgres, Stripe, Linear, Notion, Filesystem, Puppeteer, Sentry, etc.
• Stdio (local) and HTTP-SSE (remote) transports; OAuth + scoped permissions for production deployments
• You can author a custom MCP server in TypeScript, Python, or any language — the protocol is small and stable

Limitations: Authentication patterns for remote MCP are still settling. Server quality varies wildly — pin versions and review code before connecting one to production data.

Best for: Building production AI features — embedding Claude into your own product, with separate billing from claude.ai subscriptions.

• Messages API — single endpoint for all Claude models with tool use, vision, PDF, prompt caching, and streaming
• Batch API — 50% discount on async jobs that complete within 24 hours
• Files API + PDF support — upload once, reference in many calls; PDFs are parsed natively without OCR plumbing
• Workbench in the Console — interactive prompt playground with model and parameter switching, prompt caching diagnostics, and copy-as-code
• Claude Agent SDK (TypeScript and Python) — wraps the Messages API with tool loops, MCP, memory, and compaction
• Bedrock and Vertex AI distributions for AWS and Google Cloud customers who need to keep traffic in their cloud

Limitations: Pricing is per-token, not per-subscription — a runaway loop can rack up real money fast. Capacity for Opus 4.x can tighten at peak — set up retries with backoff for production traffic.

Best for: Rolling Claude out across an organization with the controls security and legal teams require.

• SSO (Okta, Azure AD, Google) and SAML/SCIM provisioning on Team and Enterprise
• Admin console with usage analytics, member management, and per-seat policy controls
• Data not used for training by default on Team and Enterprise (and on the API)
• Custom data retention windows on Enterprise (zero-retention available for qualifying customers)
• Compliance: SOC 2 Type II, ISO 27001, ISO 42001, HIPAA BAA for qualifying Enterprise customers
• Bedrock and Vertex deployments inherit the underlying cloud's compliance posture (FedRAMP, IL5, etc.)

Limitations: Audit log depth and DLP integrations are evolving — large enterprises should request a security review before standardizing. Bedrock/Vertex pricing differs from direct Anthropic API pricing.