TL;DR
| Question | Answer |
|---|---|
| Can I use OpenClaw with Anthropic? | ⚠️ YES — With separate billing |
| Ban risk | 🟡 MEDIUM (Policy shifted from ban to pay-as-yougo) |
| Key evidence | April 2026: Subscription limits no longer cover third-party tools; usage bundles required |
| Effective date | April 4, 2026 at 3PM ET |
| Options | 1) Usage bundles (pay-as-you-go) 2) API keys 3) Migrate to Kimi k2.5 — 8x cheaper |
April 2026 Policy Update: From Ban to Pay-as-you-go
Anthropic has significantly shifted its policy on third-party tool usage. As reported by The Verge on April 4, 2026, the company moved from an outright ban to a pay-as-you-go billing model for OAuth token usage in third-party tools.
What Changed (Effective April 4, 2026)
| Aspect | Before (Jan-Feb 2026) | After (April 2026) |
|---|---|---|
| Policy | Explicit ban on OAuth tokens in third-party tools | Subscription limits don’t cover third-party usage |
| Enforcement | Technical blocks, account bans | Billing separation, usage bundles required |
| Options | API keys only | Usage bundles OR API keys |
| Cost | $200-1000+/month (API) | Usage bundles at “discount” vs API |
The New Options
Option 1: Usage Bundles (New)
- “Discounted usage bundles” available for third-party tool usage
- Billed separately from Claude subscription
- Pay-as-you-go model (not unlimited)
Option 2: One-Time Credit
- Anthropic providing one-time credit equal to monthly plan cost
- Automatic for affected subscribers
- ⚠️ Expires: April 17, 2026 — must redeem by this date
- Can request full refund via email link
Option 3: API Keys (Existing)
- Continue using commercial API keys as before
- Metered pricing, full compliance
- Boris Cherny (Head of Claude Code) personally submitted PRs to improve OpenClaw’s prompt cache efficiency — API users benefit from these optimizations
Background: The One-Week Delay
Peter Steinberger (OpenClaw creator, now at OpenAI) and board member Dave Morin attempted to negotiate with Anthropic. According to Steinberger: “tried to talk sense into Anthropic, best we managed was delaying this for a week.”
This explains why the effective date (April 4) is later than initially planned—negotiations secured a brief delay but no fundamental change to the policy direction.
Why This Still Matters
Even with the pay-as-you-go option, this represents:
- End of unlimited third-party usage — Subscription arbitrage is dead
- Additional cost for OpenClaw users — Bundles add expense on top of subscriptions
- Continued platform control — Anthropic maintains tight grip on how Claude is accessed
Primary Source Evidence
The Ban Quote
Source: Anthropic Claude Code Legal and Compliance
Quote:
“Using OAuth tokens obtained through Claude Free, Pro, or Max accounts in any other product, tool, or service — including the Agent SDK — is not permitted and constitutes a violation of the Consumer Terms of Service.”
— Anthropic Claude Code Legal and Compliance, February 2026
What This Means:
- Claude Free, Pro, and Max OAuth tokens are explicitly banned from third-party tools
- The Agent SDK is also covered by this ban
- Violation = Consumer Terms of Service breach
- Account termination possible “without prior notice”
Verification Date: 2026-02-24
Enforcement Timeline
Anthropic moved from implicit restrictions to explicit enforcement, then to billing separation:
| Date | Action | Evidence |
|---|---|---|
| 2025-06 | Windsurf capacity cuts | First third-party tool restrictions |
| 2025-08 | OpenAI API access revoked | Competitive benchmarking blocked |
| 2026-01-09 | OAuth harness block deployed | Technical enforcement begins |
| 2026-01 | Account bans (some false positives) | Acknowledged by Anthropic |
| 2026-02 | Policy formalization | Legal page published with explicit ban |
| 2026-04-04 | Pay-as-you-go model | Subscription limits no longer cover third-party tools; usage bundles introduced |
The January 2026 OAuth Block
What happened:
- Anthropic deployed technical blocks against OAuth “harness” patterns
- OpenCode, Roo Code, and similar tools broke
- Some users received account suspension notices
- Anthropic acknowledged false positives and reversed some bans
Detection Methods (documented):
- Client fingerprinting via HTTP headers
- Pattern analysis of request volume/timing
- Binary signature detection
- User-agent analysis
What’s Covered by the New Policy (April 2026)
Subscription Limits No Longer Cover Third-Party Tools
| Token Type | Usage | Status | Billing |
|---|---|---|---|
| Claude Free OAuth | Third-party tools | ⚠️ Usage bundles required | Pay-as-you-go |
| Claude Pro OAuth | Third-party tools | ⚠️ Usage bundles required | Pay-as-you-go |
| Claude Max OAuth | Third-party tools | ⚠️ Usage bundles required | Pay-as-you-go |
| Consumer OAuth | Agent SDK | ⚠️ Usage bundles required | Pay-as-you-go |
| Commercial API key | OpenClaw | ✅ Allowed | Metered API pricing |
| Bedrock/Vertex | OpenClaw | ⚠️ Check enterprise terms | Enterprise pricing |
Key Change (April 2026): OAuth tokens are no longer “banned” but subscription limits don’t cover third-party usage. You need usage bundles OR API keys.
Usage Bundles vs API Keys
| Factor | Usage Bundles | API Keys |
|---|---|---|
| Cost | “Discounted” vs standard API | Full metered pricing |
| Setup | OAuth login + bundle purchase | Generate key in Console |
| Limits | Bundle-sized quotas | Account limits |
| Best for | Occasional third-party use | Heavy/production usage |
Note: Commercial API keys remain the safest long-term option for third-party integrations. Consult your Anthropic sales representative for enterprise terms.
Detection & Enforcement
How Anthropic Detects Violations
HTTP Header Analysis
- User-agent strings
- Client identifiers
- Request patterns
Usage Pattern Analysis
- Request volume anomalies
- Timing patterns (automated vs human)
- Feature usage patterns
Binary Fingerprinting
- Client application signatures
- Known third-party tool patterns
OAuth Token Flow
- Token usage outside Claude Code app
- Non-browser client detection
Enforcement Actions
| Action | Likelihood | Trigger |
|---|---|---|
| Warning email | Medium | First detection |
| Temporary suspension | High | Confirmed violation |
| Permanent ban | High | Repeated violation |
| API key revocation | Medium | Commercial API abuse |
Key Clause:
“We may suspend or terminate your account without prior notice.”
Why Anthropic Banned This Pattern
Strategic Context
Anthropic is building a walled garden:
- Claude Code — Official agentic tool, requires subscription
- Agent SDK — For enterprise customers only
- Consumer plans — No third-party access allowed
Business Logic:
- Protect Claude Code revenue
- Control user experience
- Prevent “shadow AI” proliferation
- Maintain audit/compliance control
Contrast with OpenAI
| Aspect | Anthropic | OpenAI |
|---|---|---|
| Philosophy | Walled garden | Marketplace |
| Official CLI | Claude Code only | Codex CLI (uses OAuth) |
| Third-party OAuth | Explicitly banned | Explicitly supported |
| Enforcement | Technical blocks + bans | None documented |
If You’re Currently Using OpenClaw with Anthropic
Immediate Actions
- STOP using Anthropic OAuth tokens with OpenClaw — Don’t wait for a ban
- Export your data — Conversations, configurations, custom skills
- Rotate your API key — If keeping Anthropic for other uses
- Switch to an allowed provider — See migration paths below
If You Receive a Ban Notice
- Appeal if false positive — Anthropic has reversed some bans
- Document your usage — Be ready to prove compliance if disputed
- Don’t evade — Creating new accounts violates terms further
- Migrate completely — Use this as catalyst to move
⚠️ Data Loss Warning
Critical: Anthropic bans can result in permanent loss of your data.
What You Lose When Banned
| Asset | Status | Recovery Path |
|---|---|---|
| Conversation History | ❌ LOST | No export during suspension |
| Claude Memories | ❌ LOST | Personalized context gone |
| Projects/Artifacts | ❌ LOST | Any unsaved work inaccessible |
| API Keys | ⚠️ Revoked | Must generate new ones (if account restored) |
| OpenClaw Config | ✅ SAFE | Stored locally on your machine |
Can You Export After Being Banned?
Short answer: Probably not.
- ❌ Self-service data export is disabled during account suspension
- ⚠️ Some users report success requesting export during the appeal process
- ❌ No guarantee — Anthropic’s priority is compliance, not data recovery
Protect Yourself NOW
Before you get banned:
Export your Claude data immediately
Claude.ai → Settings → Data Export → Request Download- Takes 24-48 hours to generate
- Includes conversations, memories, account data
Screenshot critical conversations
- No bulk export for individual chats
- Copy/paste important code or insights to local files
Don’t rely on Claude’s “memory”
- Keep your own documentation of preferences
- Store reusable prompts locally
Use version control for generated code
1 2 3 4# Commit anything important immediately git add . git commit -m "Backup Claude-generated code" git push
The bottom line: If Claude knows your codebase, your style, your preferences — that’s all ephemeral. Treat it like cache, not storage.
Migration Paths
Recommended: Kimi k2.5
Why:
- 76.8% SWE-bench (80% of Opus 4.5’s 80.9%)
- 8x cheaper API ($3/1M vs $25/1M output tokens)
- Native vision capabilities
- Explicitly allowed with OpenClaw
Migration:
- Export OpenClaw config
- Swap endpoint:
api.anthropic.com→api.moonshot.cn - Update model:
claude-sonnet-4→kimi-k2.5 - Test with small tasks
- Decommission Anthropic key
Free Access: NVIDIA NIM — Free Kimi k2.5 API credits
Alternative: Google Gemini
Why:
- Free tier available
- No restrictions found
- Strong context window (1M tokens)
Guide: /value/free-stack/#google
Ultimate: Self-Hosted
Why:
- Zero provider risk
- You are the provider
- No terms to violate
Guide: /tools/self-hosting/
Verification Ledger
✅ VERIFIED: Strong Evidence
Explicit ban on consumer OAuth in third-party tools
- Source: Claude Code Legal and Compliance page
- Quote: “not permitted and constitutes a violation”
- Date verified: 2026-02-24
- URL: code.claude.com/docs/en/legal-and-compliance
OAuth harness block deployed January 2026
- Source: GitHub issues, user reports
- Impact: OpenCode, Roo Code broken
- Date: 2026-01-09
Account bans (some reversed)
- Source: Community reports, Anthropic acknowledgment
- Scope: Some false positives reversed
- Date: January 2026
“Without prior notice” termination clause
- Source: Consumer Terms of Service
- Quote: “suspend or terminate… without prior notice”
- Date verified: 2026-02-24
FAQ
Can I use OpenClaw with Anthropic?
Yes, but with separate billing. As of April 4, 2026, Claude subscriptions no longer cover third-party tool usage. You have three options:
- Usage bundles — “Discounted” pay-as-you-go for OAuth-based third-party access
- API keys — Commercial API (metered pricing, full compliance)
- Switch providers — OpenAI or Kimi k2.5 (see Migration Paths)
What changed in April 2026?
Anthropic shifted from an outright ban (Jan-Feb 2026) to a billing separation model (April 2026). OAuth tokens still work, but subscription limits don’t apply to third-party tools. You need usage bundles for continued OAuth-based access.
What happens if I keep using Anthropic OAuth with OpenClaw?
Before April 4, 2026: Account suspension risk.
After April 4, 2026: Usage will be billed via pay-as-you-go bundles (or blocked if no bundle purchased).
Anthropic may still suspend for terms violations, but the primary enforcement is now billing-based rather than access-based.
I already got banned in January/February 2026. Can I get my data back?
Probably not for past bans. Self-service export is disabled during suspension. Some users report success requesting export during the appeal process, but there is no guarantee.
For April 2026 changes: This is a billing change, not a ban. Your data remains accessible.
What’s the safest alternative to Anthropic?
OpenAI (explicitly allows third-party OAuth via Codex CLI) or Kimi k2.5 (8x cheaper, no restrictions). See Migration Paths.
What are the actual cost impacts?
Based on user reports post-April 4, 2026:
| Usage Pattern | Old Cost (Subscription) | New Cost (API/Bundles) |
|---|---|---|
| Light agent use | $20/month (Pro) | $20-50/month |
| Moderate use | $100/month (Max 5x) | $100-300/month |
| Heavy/Orchestration | $200/month (Max 20x) | $500-1000+/month |
| Per-task cost | Included in sub | $0.50-$2.00 per agent task |
Key insight: A single 200K context session with Opus can cost $5-10 in API fees alone. Heavy OpenClaw users report 5-10x cost increases.
Can I get a refund instead?
Yes. Anthropic is offering full refunds for affected subscribers who prefer to cancel rather than migrate. Contact Anthropic support to request.
Should I use usage bundles or API keys?
API keys are safer long-term. Usage bundles are new, pricing is unclear, and policy could shift again. API keys have stable commercial terms and clearer compliance boundaries.
Bundle discount: Up to 30% off when pre-purchasing — only worthwhile if you’re committed to staying with Anthropic.
What Would Change This Rating
To upgrade (less risky):
- Anthropic publishes explicit authorization for third-party tools
- Technical blocks removed
- Official API key exception documented
Unlikely — Anthropic’s walled garden strategy is deliberate.
Review cadence: Monthly during policy volatility; quarterly otherwise.
Government Pressure: The DOD Ultimatum & Anthropic’s Response (February 2026)
On February 25, 2026, Defense Secretary Pete Hegseth issued an ultimatum to Anthropic CEO Dario Amodei:
- Deadline: Friday, February 27, 2026 at 5:01pm
- Threats: Supply chain exclusion, Defense Production Act invocation, “supply chain risk” label
- Trigger: Anthropic’s refusal to allow “any lawful use” with no safety carve-outs
Anthropic’s Public Response (February 26, 2026):
One day before the deadline, Amodei published a statement pointedly titled “Statement from Dario Amodei on our discussions with the Department of War” — using the pre-1949 name for the DoD.
Two explicit redlines confirmed publicly:
| Redline | Status |
|---|---|
| Mass domestic surveillance | 🔴 Refused — “incompatible with democratic values” |
| Fully autonomous weapons | 🔴 Refused — AI “not reliable enough”; human oversight required |
Anthropic stated it does cooperate with foreign intelligence, counterintelligence, and partially autonomous systems.
“Regardless, these threats do not change our position: we cannot in good conscience accede to their request.” — Dario Amodei, February 26, 2026
Why this matters for OpenClaw users: Anthropic’s public refusal to cede policy control to government pressure reinforces the company’s walled-garden philosophy. If the company holds redlines on billion-dollar government contracts, it will hold redlines on third-party OAuth enforcement. These are the same organizational values.
Full Coverage:
- /posts/anthropic-dod-pressure-2026/ — Full timeline, DPA context, official statement breakdown
- /verify/ai-provider-military-use-policies/ — Military use policy comparison
Related Links
Provider Policy Hub:
- /verify/openclaw-provider-policies/ — All providers summary
- /posts/openclaw-provider-policy-check-2026/ — Breaking news, migration paths, enforcement timeline
Contrast With:
- /verify/openclaw-openai-policy/ — OpenAI’s allowed status
- /verify/openclaw-google-policy/ — Google’s split enforcement (similar ban pattern)
Anthropic Context:
- /verify/claude-code-terms/ — Claude Code terms deep-dive
- /verify/anthropic-api-terms/ — API terms analysis
- /posts/anthropic-tos-changes-2025/ — Timeline of policy changes
- /risks/anthropic/account-bans/ — Account ban risk factors
Migration:
- /tools/opencode/ — Free Kimi access
- /value/free-stack/ — Complete free tier guide
Sources
Primary Sources
- Claude Code Legal and Compliance — February 2026
- Anthropic Consumer Terms of Service
- Anthropic Commercial Terms
Enforcement Evidence
- GitHub Issues: OpenCode, Roo Code projects (January 2026)
- Community reports on X/Twitter
- Anthropic support acknowledgments
News Coverage
- The Verge: Using OpenClaw with Claude AI is about to get more expensive — April 4, 2026 — Primary reporting on April 2026 policy change (Archive)
- Business Insider: ClawdBot rename after Anthropic outreach
- TechCrunch: Anthropic OAuth enforcement
Last verified: 2026-04-06 Next review: 2026-05-06 Evidence level: High (official policy + The Verge primary reporting + technical enforcement + account bans + CEO statement)