On April 4, 2026 at 12pm PT, Anthropic cut Claude Pro and Max subscription access inside every third-party agentic tool. The company cited infrastructure strain. The timing, the exceptions, and the founder of the biggest affected project all tell a more complicated story.

We covered the initial block yesterday. This piece is about what doesn’t add up.

Quick Summary

Detail	Info
What Anthropic says	”Unusual traffic patterns without telemetry” caused infrastructure strain
What’s blocked	All third-party agentic tools using Claude Pro/Max subscriptions
What’s not blocked	Claude Code’s own looping and agent features, same token consumption pattern
Affected users	~135,000 active OpenClaw instances (revised from initial 40K reports)
Who’s calling foul	OpenClaw founder Peter Steinberger, now at OpenAI
The accusation	Anthropic copied popular OpenClaw features into Claude Code, then killed the open alternative
Compensation	One-time credit (expires April 17, 2026) + up to 30% off pre-purchased API bundles

Bottom line: The cost math justified action. The selective enforcement — blocking third-party loops while keeping first-party loops running — makes “infrastructure strain” sound like a convenient half-truth.

Why Anthropic’s Official Rationale Has a Gap

Boris Cherny, Head of Claude Code at Anthropic, framed the block around “unusual traffic patterns without telemetry.” Not just cost. Not just volume. The absence of telemetry, meaning Anthropic couldn’t see what these tools were doing with Claude’s outputs or how they were chaining calls.

That’s a more interesting justification than raw infrastructure strain. It’s a control argument. Anthropic wants visibility into how Claude gets used at scale, and third-party tools operating as black boxes made that impossible.

Fair enough. Except Claude Code itself has built-in looping features that chain sequential Claude calls, maintain large context windows across iterations, and run autonomous multi-step tasks for extended periods. The multi-agent harness Anthropic announced the same day runs three coordinated Claude instances for up to four hours straight.

Those features consume tokens in patterns identical to what OpenClaw was doing. The difference: Anthropic controls the telemetry.

So the real line isn’t “you’re using too many tokens.” It’s “you’re using tokens in ways we can’t monitor.” That distinction matters, because it shifts the conversation from capacity management to platform control.

The TOS Timeline Nobody Mentions

This block didn’t come out of nowhere, even if it felt that way to users.

How Anthropic’s Third-Party Restriction Escalated

1. February 2024: Anthropic’s Terms of Service first prohibited routing subscription access through third-party tools. The clause existed. Almost nobody noticed.
2. February 20, 2026: Anthropic formally re-clarified the restriction. OpenCode (a separate open-source project) removed Claude support after receiving legal requests from Anthropic.
3. April 4, 2026, 12pm PT: Hard enforcement. All third-party agentic tools blocked from Claude Pro/Max subscriptions. Approximately 135,000 active OpenClaw instances affected.

That’s a two-year runway from policy to enforcement. Anthropic can reasonably argue they gave warning. But a TOS clause nobody reads and a legal letter to one project aren’t the same as communicating directly with the 135,000 users who built workflows on this access pattern.

The phased escalation looks deliberate in hindsight. It also looks like Anthropic waited until Claude Code had absorbed enough of OpenClaw’s most popular features before pulling the trigger.

Peter Steinberger’s Accusation

Peter Steinberger created OpenClaw (originally Clawdbot, then Moltbot; Anthropic’s trademark team forced the rename, which should have been a signal). He left the project and joined OpenAI in February 2026.

His public response to the block: Anthropic’s decision was “a betrayal of open-source developers.”

Steinberger went further. He pointed out that Anthropic systematically adopted popular OpenClaw features (autonomous coding loops, multi-file analysis, persistent context across sessions, MCP tool integration) into Claude Code. Then, once the first-party product covered those use cases, they cut off the open-source alternative that pioneered them.

That’s a serious accusation. And given that Steinberger now works at OpenAI, it’s worth weighing his incentives alongside his claims. He has both the most direct knowledge of how this played out and the most obvious reason to frame it negatively.

But the timeline supports his version. Claude Code gained agent-team capabilities, background tasks, and extended autonomous sessions throughout late 2025 and early 2026. The block came after those features shipped, not before. Coincidence is possible. It’s not the most parsimonious explanation.

The Cost Math Was Genuinely Broken

I don’t want to dismiss the economics, because they’re real.

One autonomous OpenClaw agent running a full day of sustained agentic workloads can burn through $1,000 to $5,000 in API-equivalent costs. The Claude Max subscription costs $200/month. Even Pro at $20/month was being used for workloads worth hundreds in compute.

That’s not sustainable. It’s not even close. The business model of flat-rate subscriptions depends on average usage staying well below the cost ceiling. Tools like OpenClaw didn’t just push above the average. They optimized for extracting maximum compute from a fixed-price subscription. Every user was a power user. That breaks the model.

Our AI pricing comparison breaks down the per-token costs across providers. At standard Claude API rates, the workloads OpenClaw users were running would cost 5-25x their subscription fee.

So yes, Anthropic had a financial reason to act. The question isn’t whether the economics justified some response. It’s whether the specific response — a total block with no migration period, while keeping functionally identical first-party features unrestricted — was really about economics or about something else.

What This Looks Like From the Outside

Two things can be true at once:

Anthropic was losing money on power users routing subscriptions through third-party tools. The unit economics were structurally broken, and the longer they waited, the worse the bleeding got as OpenClaw’s user base grew.

Anthropic also benefits competitively from eliminating the open-source alternative to Claude Code. With OpenClaw effectively dead as a Claude-powered tool (the API pivot will take months and will cost users dramatically more), Claude Code becomes the only game in town for agentic Claude workflows. That’s not an accident. That’s a moat.

The infrastructure strain argument and the competitive moat argument aren’t mutually exclusive. Anthropic likely acted for both reasons. But by framing it purely as a cost and telemetry issue, they’re asking users not to notice the competitive benefit.

Steinberger noticed. So did the 135,000 developers who now have to either pay API rates, switch to Claude Code, or evaluate whether other models can handle their workloads at better unit economics.

What Are Your Options Now

If you were affected, the practical picture hasn’t changed much since our initial coverage:

• Claude API is the sanctioned path for programmatic access. The 30% discount on pre-purchased bundles expires eventually. Anthropic hasn’t specified when, but the credit redemption deadline of April 17, 2026 suggests the window is short. Lock in a bundle if you’re going API.
• Claude Code covers most development-focused agentic workflows that OpenClaw handled. It’s first-party, it has telemetry Anthropic approves of, and it isn’t going anywhere. The trade-off: you’re inside Anthropic’s ecosystem now, on their terms.
• Other providers are worth evaluating if you’re not locked into Claude’s reasoning quality. OpenAI and Google both offer API access without this kind of restriction on third-party tooling — for now. My prediction from yesterday stands: expect similar restrictions from every major provider within 12 months.
• Wait for OpenClaw’s API pivot. The team is working on it. The economics will be worse. The functionality should be similar. No timeline yet.

Our Take

Anthropic made the right call on the economics and the wrong call on the execution. The unit economics of subscription-powered agentic workloads were broken. Everybody building on that pattern knew it was too good to last, even if nobody wanted to admit it.

But the selective enforcement undermines the stated rationale. If the problem is “unusual traffic patterns without telemetry,” then the solution is to require telemetry, not to block third-party tools while running functionally identical patterns through your own product. Call it what it is: competitive positioning with infrastructure management as cover.

The TOS existed since 2024. Anthropic could have communicated directly to OpenClaw’s user base months ago, offered a migration timeline, and built API bridge tooling. Instead, they waited until Claude Code had feature parity, then enforced a two-year-old clause with zero transition period.

I think Steinberger is partially right and partially motivated. The “betrayal” framing is strong. Anthropic is a business managing its costs, not a nonprofit breaking a promise. But the pattern of adopting community-pioneered features into first-party products and then cutting off the community? That’s worth naming, even if the economics independently justified the block.

The AI tool ecosystem is watching. If Anthropic handles the post-block transition well (extended credit window, real API migration tooling, maybe an official third-party integration program with proper telemetry requirements), this becomes a rough but defensible business decision. If they let the 135,000 affected developers twist while promoting Claude Code as the replacement? The “walled garden” label sticks.

Frequently Asked Questions

Why were 135,000 users reported instead of 40,000?

Initial reports cited approximately 40,000 active OpenClaw users. Revised figures accounting for all third-party agentic tools (not just OpenClaw’s primary interface) and counting active instances rather than unique accounts put the affected population closer to 135,000. The discrepancy reflects how broadly third-party Claude routing had spread beyond OpenClaw itself.

Didn’t Anthropic’s Terms of Service already prohibit this?

Yes. The TOS clause prohibiting subscription routing through third-party tools has existed since February 2024. It was formally re-clarified on February 20, 2026, when Anthropic sent legal requests to projects like OpenCode. Enforcement on April 4 was the final step in a phased escalation, not a new policy, but most users were unaware of the clause until the block hit.

How is Claude Code different from OpenClaw if both loop Claude calls?

Functionally, both chain sequential Claude calls with large context windows for autonomous tasks. The difference is telemetry and control. Claude Code runs inside Anthropic’s ecosystem, giving them full visibility into usage patterns, model outputs, and error rates. OpenClaw operated as a black box from Anthropic’s perspective. The block is about who controls the pipeline, not about what the pipeline does.

Is Peter Steinberger a reliable source given he joined OpenAI?

Steinberger has the most direct knowledge of how Anthropic interacted with OpenClaw, having created the project and navigated Anthropic’s trademark and legal pushback firsthand. He also has an obvious competitive incentive as an OpenAI employee. Both things are true. His factual claims about the timeline (feature adoption into Claude Code followed by enforcement) are verifiable independently of his framing.

Will the 30% API bundle discount last?

Anthropic hasn’t announced an end date for the discount. The one-time credit must be redeemed by April 17, 2026. The bundle discount appears to be a standing transition offer, but given that it’s positioned as a migration incentive, purchasing sooner rather than later is the safer bet. Check the Anthropic API guide for current pricing details.

---

Last updated: April 5, 2026. Sources: Anthropic announcement via Boris Cherny (Head of Claude Code), Peter Steinberger’s public statements, revised user impact figures from community tracking. See our initial coverage for the breaking report.

Related reading: OpenClaw Review | Cursor vs Claude Code vs Copilot | Anthropic API Guide | AI Pricing Comparison