DEEP DIVE · 06
Anthropic ecosystem
April 2026 snapshot
Everything that's shipping, as of 2026-04-22. Opus 4.7 / Sonnet 4.6 / Haiku 4.5 — the full model line. Every surface Claude lives on. Enterprise security. How it all lines up against ChatGPT and Gemini.
Anthropic ecosystem map · as of 2026-04 · zoomable
Contents · 11 sections
01Model line-up
02All surfaces
03Timeline · 2026 Q1-Q2
04Skills system
05Agent SDK
06MCP
07Plan matrix
08Enterprise security
09Competitors
10Business signals
11Official sources
This page is a snapshot of Anthropic's products, models, enterprise story, and competitive position as of April 21, 2026. Every number and date is sourced from official channels.
1. Model line-up
Three Claude 4 models are live today. Opus 4.7 shipped on 2026-04-16 and pushed the context window up to 1M tokens in beta.
| Model | Released | Context | Input / 1M | Output / 1M | Batch discount | Prompt Cache (write / read) | Best for |
|---|---|---|---|---|---|---|---|
| Claude Opus 4.7 | 2026-04-16 | 1M (beta) / 200K std | $15 | $75 | -50% | $18.75 / $1.50 | Hard reasoning, agents, long-form code |
| Claude Sonnet 4.6 | 2026-02-17 | 1M (beta) / 200K std | $3 | $15 | -50% | $3.75 / $0.30 | Production workloads, balanced |
| Claude Haiku 4.5 | 2025-10, still current | 200K | $1 | $5 | -50% | $1.25 / $0.10 | Low-latency, high-volume, routing |
Batch API. 50% off on both input and output for async jobs, with a 24-hour SLA.
Prompt caching. Cache writes cost 1.25× base input; reads cost 0.1× (90% off). Two TTLs: 5 minutes (ephemeral) and 1 hour (extended, +20% premium). For agents that hit the same system prompt over and over, total cost drops to roughly one-fifth of the non-cached run.
Prompt caching. Cache writes cost 1.25× base input; reads cost 0.1× (90% off). Two TTLs: 5 minutes (ephemeral) and 1 hour (extended, +20% premium). For agents that hit the same system prompt over and over, total cost drops to roughly one-fifth of the non-cached run.
Executive takeaway. Three tiers: Opus for the hard stuff, Sonnet as the everyday workhorse, Haiku for volume. Most agent deployments run Sonnet 4.6 by default and only escalate to Opus 4.7 when they have to. That's where the cost curve bends in your favor.
2. All surfaces
Claude isn't a single API. It's a family of surfaces, each with its own shape. The thing to know as an exec is where Claude enters your stack.
| Category | Surface | Form factor | Notes |
|---|---|---|---|
| Chat UI | claude.ai | Web | Pro / Max / Team / Enterprise |
| Claude Desktop | macOS / Windows | Native app, hosts MCP | |
| Claude mobile | iOS / Android | Voice mode, Projects sync | |
| Developer | Claude Code CLI | Terminal | Ships with Agent SDK |
| Claude Code Desktop | macOS / Windows | GA in 2026 Q1 | |
| Claude Code Web | claude.ai/code | Browser-only | |
| Claude Code IDE | VS Code / JetBrains | Extensions | |
| API | api.anthropic.com | REST | Direct contract |
| Agent SDK | Python / TypeScript | OSS, same runtime as Claude Code | |
| Cloud | AWS Bedrock | Tokyo region included | Opus 4.7 landed in Tokyo 2026-04-20 |
| Google Cloud Vertex AI | us-central1 / asia-northeast1 et al. | — | |
| Microsoft Azure AI Foundry | GA in 2026 Q1 | Runs alongside Azure OpenAI | |
| Integrations | MCP servers | Desktop / Code / API | Hundreds listed in the official registry |
| Microsoft Copilot Cowork | Microsoft 365 | Announced 2026-03-10, ties into Teams and Outlook | |
| Claude for Word | Word add-in beta | 2026-04-10 |
Where to focus. For most enterprises, the four paths that matter are Bedrock Tokyo, Claude Desktop, MCP, and Copilot Cowork. Which one IT standardizes on decides the whole governance story.
3. Timeline (January to April 2026)
Two things happened at once this quarter: a full model refresh (Sonnet 4.6, then Opus 4.7) and an enterprise surface blitz (Azure, Copilot Cowork, Word, Bedrock Tokyo).
| Month | Date | Release |
|---|---|---|
| 2026-01 | — | Claude Code Web GA; Projects memory upgrades |
| 2026-02 | 2026-02-17 | Claude Sonnet 4.6; 1M context in beta; Agent SDK v1.0 stable |
| 2026-03 | 2026-03-10 | Microsoft Copilot Cowork integration; Claude GA on Azure AI Foundry |
| 2026-04 | 2026-04-10 | Claude for Word beta (Office add-in) |
| 2026-04 | 2026-04-16 | Claude Opus 4.7; 1M context; better reasoning-token efficiency |
| 2026-04 | 2026-04-20 | AWS Bedrock Tokyo region — Opus 4.7 now available |
4. Skills system
Skills are a lightweight way to teach Claude domain-specific procedures. Anthropic introduced them in October 2025. In December 2025 they became an industry standard via agentskills.io.
- What they are. A directory with a
SKILL.mdat its root — YAML frontmatter plus Markdown body. Scripts and templates can sit alongside it. - How they fire. The model matches user intent against skill descriptions and invokes them automatically. You can also call them explicitly with
/skill-name. - Hot reload. Edit a skill in Desktop or Code and it picks up the change with no restart.
- Agentskills.io standard (2025-12). A format agreed on by Anthropic, OpenAI, Google, and Microsoft. No vendor lock-in.
- When to use what. Skills hold procedural knowledge. Subagents run in their own context as parallel workers. Slash commands are user-triggered. More in
docs/research/mcp-skills.md.
Executive takeaway. Because Skills became a multi-vendor standard, anything you write once runs on Claude, ChatGPT, Gemini, or Copilot. That's the clean way to turn internal know-how into an asset without locking yourself in.
5. Agent SDK
The runtime that powers Claude Code, extracted into an open-source package. If you want to embed an agent inside your own product, this is the intended path.
- Languages. Python (
claude-agent-sdk) and TypeScript (@anthropic-ai/claude-agent-sdk). - License. Apache 2.0. Public on GitHub.
- Built-in tools. File I/O, shell execution, web fetch, web search, computer use (beta).
- MCP. MCP client is built in. Any MCP server plugs in as a tool.
- Multi-cloud. Switch the API key and you can run against Anthropic direct, Bedrock, Vertex AI, or Azure Foundry.
- Operations. Context compaction, permission modes, hooks, subagents — same implementation as Claude Code.
6. MCP (Model Context Protocol)
Anthropic published MCP in November 2024. Through 2025 it became the de facto standard for agent integration. OpenAI's and Google's first-party tools speak it now too.
- Protocol. JSON-RPC over stdio, SSE, or HTTP. Exposes three resource kinds: tools, resources, prompts.
- Official registry. Searchable at
modelcontextprotocol.io. 800+ servers registered as of 2026-04.
Servers that matter
| Category | Representative MCP servers |
|---|---|
| Collaboration | Slack, Microsoft Teams, Notion, Linear, Jira, Asana |
| Files | Google Drive, OneDrive, Box, Dropbox |
| Code | GitHub, GitLab, Bitbucket |
| Data | Postgres, MySQL, BigQuery, Snowflake, Databricks |
| Observability | Datadog, Sentry, Grafana |
| CRM / operations | Salesforce, HubSpot, Stripe, QuickBooks |
From a MIXI angle. Slack, GitHub, BigQuery, Datadog — you already run all of these, and all of them have official MCP servers. You don't need a new system. You need to wire the existing ones in and expose them to an agent. That's the whole lift for a cross-tool exec query surface. More in deep-mcp.
7. Plan matrix
What actually changes as you move from individual to enterprise. The core four: SSO, audit logs, ZDR, retention window.
| Feature | Pro ($20/mo) | Max ($100-200/mo) | Team ($30/seat) | Enterprise (contact) |
|---|---|---|---|---|
| Opus 4.7 usage | Capped | Much higher | Higher | SLA-backed |
| Claude Code | Included | Included (higher cap) | Included | Included |
| Projects | Yes | Yes | Yes (shared) | Yes (org-shared) |
| MCP connectors | Yes | Yes | Yes | Yes + managed |
| SSO / SAML | No | No | Yes | Yes |
| Audit logs | No | No | Basic | Detailed + SIEM |
| ZDR (Zero Data Retention) | No | No | No | Yes, via addendum |
| Custom retention | No | No | No | Yes |
| Dedicated support | No | No | No | Yes |
| Seat minimum | 1 | 1 | 5 | Negotiated, usually 50+ |
8. Claude for Enterprise — security
What ships in the enterprise contract, framed the way a Japanese procurement or IT team will want to see it.
- SSO / SAML. Okta, Azure AD / Entra ID, Google Workspace, Ping. SCIM for provisioning.
- Certifications. SOC 2 Type II, ISO 27001 / 27701 / 42001, HIPAA BAA, GDPR.
- ZDR addendum. Optional rider that discards inputs and outputs from logs immediately. No training use either way (that's the default).
- Custom retention. Default is 30 days. Can be set to 0 or to any length. Configured per organization.
Japan residency paths
- AWS Bedrock Tokyo region (Opus 4.7 available from 2026-04-20)
- Google Cloud Vertex AI,
asia-northeast1 - Anthropic direct + ZDR (processed in the US, but nothing retained)
Admin surface. Org admin console, role-based access, DLP integration, API usage audit. More in deep-security.
9. Competitor map
How Claude lines up against the other two frontier vendors as of April 2026. The strongest relative axes for Claude: code generation, long / large context, and agent stability.
| Dimension | Claude (Opus 4.7 / Sonnet 4.6) | ChatGPT (GPT-5.4) | Gemini (2.5 Pro / Ultra) |
|---|---|---|---|
| Coding | Top tier on SWE-bench; whole Claude Code stack around it | Codex integration, OpenAI Apps | Gemini Code Assist, Jules |
| Long context | 1M (beta), strong effective accuracy | 400K | 2M, but effective range is shorter |
| Vision | Strong on diagrams and UI | Strong, with image generation | Strong, ahead on video |
| Agents / tools | Agent SDK + MCP, most mature today | Assistants API, Operator | Vertex Agent Builder |
| Enterprise | Bedrock / Vertex / Azure — three clouds, plus ZDR | Azure OpenAI-centric | Google Cloud-centric |
| Pricing | Opus premium, Sonnet is the main tier | GPT-5.4 roughly matches or undercuts Opus | Ultra is competitive |
| Japan | Tokyo via Bedrock, yes | East Japan via Azure, yes | Tokyo via Vertex, yes |
How to decide. In 2026, "pick one vendor, kill the others" is the wrong frame. The standard enterprise shape is Sonnet 4.6 as the spine, GPT-5.4 as a complement, Gemini for specific niches. Head-to-head detail in deep-comparison.
10. Business-scale signals
What public reporting suggests about Anthropic's scale. Not financial disclosure — treat accordingly.
- Run-rate revenue. 2026 Q1 estimates put it at $20B+ annualized (The Information, FT). That's up from roughly $2B a year earlier.
- Enterprise customers. A large share of the Fortune 500, and 300,000+ businesses using the API once you include Bedrock (Anthropic's own figure).
- Developers. Claude Code is reported at roughly one million monthly active developers as of 2026 Q1.
- Funding. Cumulative $8B from Amazon, $3B+ from Google. Training runs on both Trainium and TPU.
- IPO chatter. Reports of IPO preparation for late 2026 or 2027 rumor. No official comment.
Labeling. Run-rate, user counts, and IPO chatter are press reports. If any of this gets quoted inside MIXI, tag it "as reported in 2026-04" and cite the source URL.
11. Official sources
Last verified: 2026-04-22
- Anthropic news: anthropic.com/news
- Model docs: docs.claude.com
- API / pricing: platform.claude.com/docs
- Claude Code docs: docs.claude.com/claude-code
- Agent SDK (Python): github.com/anthropics/claude-agent-sdk-python
- Agent SDK (TypeScript): github.com/anthropics/claude-agent-sdk-typescript
- MCP spec and registry: modelcontextprotocol.io
- Skills standard: agentskills.io
- AWS Bedrock Claude: aws.amazon.com/bedrock/claude/
- Google Vertex AI Claude: cloud.google.com/vertex-ai/generative-ai/docs/partner-models/claude
- Microsoft Azure AI Foundry: azure.microsoft.com/products/ai-foundry
- Microsoft Copilot Cowork (2026-03-10): microsoft.com/microsoft-365/blog
- AWS Bedrock Tokyo, Opus 4.7 (2026-04-20): aws.amazon.com/blogs/aws/
- Enterprise security / Trust Center: trust.anthropic.com