On June 15, Anthropic splits every Claude paid subscription into two separate credit pools — interactive and programmatic. If you run any third-party agent harness against Claude, including OpenClaw, your flat-rate model access ends on that date. The replacement is a fixed, non-rollover monthly credit billed at API rates. Pro subscribers get $20. Max 20x subscribers get $200. No surplus. No carryover. No exceptions.

The five weeks between now and June 15 are a pricing window. What you decide about your tooling stack before that deadline will determine whether your monthly AI spend doubles, triples, or holds. Most of the coverage around the OpenClaw ban has been about the drama — the creator getting locked out, the legal threats, the community outrage. This piece is about the dollars. Specifically, the structural cost mechanics that made the ban inevitable and the math that makes the so-called reinstatement functionally different from what it replaces.

The Flat-Rate Arbitrage

Every time a platform ships a flat-rate subscription over variable-cost infrastructure, a subset of users extracts multiples of the stated price. Anthropic's Claude subscriptions were no different.

A Max 20x subscription costs $200 per month. OpenClaw users running autonomous agent loops against Claude models were consuming, by Anthropic's own internal estimates, between $1,000 and $5,000 in equivalent API value monthly. The subscription absorbed that cost. The user saw a flat number on the credit card statement. The delta between $200 paid and $5,000 consumed sat on Anthropic's balance sheet.

This is arithmetic, not a moral argument. Anthropic priced its Max tier for interactive use — conversations, code reviews, document analysis with a human in the loop and natural pauses between requests. Agentic workloads, where a harness fires hundreds of sequential API calls with no human checkpoints, shatter the usage assumptions baked into that price point. The subsidy was large. It was real.

The concession is worth stating without hedging: Anthropic had a legitimate cost problem. No company absorbs a 25:1 ratio of compute cost to subscription revenue on a growing user segment without eventually closing the gap. On the pure question of whether the flat-rate model was sustainable under agentic load, the answer is no. The ban, in the narrowest possible framing, was defensible on cost grounds alone.

Everything else about how Anthropic handled it — the timing, the legal instruments, the individual account suspensions — erodes that defense.

The Cache Miss Tax

The cost gap between subscription revenue and actual compute was not purely a volume problem. It was an efficiency problem. And Anthropic's own caching architecture made it structurally worse for third-party tools than it needed to be.

Anthropic's prompt caching stores previously processed token sequences so that subsequent requests matching the same prefix skip reprocessing. Cache reads cost 0.1x the standard input rate — a 90% discount on input tokens. When Anthropic's first-party tools like Claude Code hit the API, they are engineered to maximize cache hit rates by keeping system prompts stable and conversation prefixes intact across requests.

OpenClaw was not engineered this way. GitHub Issue #49700 in the openclaw/openclaw repository documents the architectural root cause. The function `buildAgentSystemPrompt()` concatenates static and dynamic content into a single monolithic string. Workspace file listings, project context, tool configurations — all of it mutates between requests. Anthropic's caching requires an exact prefix match. Any change in the system prompt invalidates the entire cache for that session.

The measured result: approximately 10% cache hit rate for OpenClaw sessions. Every request processed 100,000 to 300,000 tokens from scratch with no cache benefit. Responses slowed progressively as conversation history accumulated. A request that would cost Anthropic roughly $3 in cached compute was costing $30 without it. Multiply by hundreds of requests per day per heavy user. Multiply again by the user segment that adopted OpenClaw specifically because it unlocked autonomous workflows.

Anthropic described these users as "really hard for us to do sustainably." The statement was vague on purpose. The cache data is precise. The platform was absorbing ten times the budgeted compute cost per request because the third-party harness was architecturally incompatible with the optimization layer that made flat-rate subscriptions viable in the first place.

A $200 subscription absorbing $5,000 in compute is a pricing failure; a 10% cache hit rate turning $3 requests into $30 ones is an engineering failure — and Anthropic chose to solve only the first.

Two documents govern what happened next, and they point in different directions.

In February 2026, Anthropic updated its terms of service with a specific new clause: "The use of OAuth tokens obtained via Claude Free, Pro, or Max accounts in any other product, tool, or service is not permitted." This was a narrow commercial rule — a preemptive legal foundation for blocking third-party authentication through subscription accounts. Clean. Specific. Targeted at the billing mechanism, not at individual users.

On April 4, enforcement landed. Claude Pro and Max subscribers could no longer use their subscriptions with third-party harnesses. OpenClaw usage required a separate API key at standard per-token rates.

Six days later, the situation escalated into something the February terms update does not explain. Peter Steinberger — OpenClaw's creator, who had left the project in February to join OpenAI — posted that Anthropic had suspended his personal Claude account. The email he received cited "suspicious signals associated with your account" and "a violation of our Usage Policy." Not a violation of the OAuth token restriction. A violation of the broader Usage Policy — a different instrument entirely.

Steinberger stated publicly he was already on API billing. He was complying with the new rules. If that is accurate, the OAuth token restriction does not apply to him. The Usage Policy invocation implies a different rationale, one Anthropic has never publicly clarified. His account was reinstated within hours. When asked to compare his treatment by the two major AI labs, Steinberger offered five words: "One welcomed me, one sent legal threats."

The February OAuth prohibition is a clear commercial boundary. The April Usage Policy suspension is a vague behavioral instrument. Anthropic deployed both within the same week against the same ecosystem. The first restructures billing. The second flags conduct. Conflating the two — which Anthropic effectively did by enforcing them simultaneously — converts a legitimate pricing correction into something that reads, from the outside, like punitive action against a departing community figure.

The Reinstatement That Changes Nothing

On May 13, Anthropic announced a reversal. Third-party agent usage would be permitted again on paid subscriptions through a new "Agent SDK" credit pool. The framing was conciliatory. The structure is not.

Under the new system taking effect June 15, every paid subscriber receives a monthly programmatic credit equal to their subscription price. Pro at $20 gets $20 in credits. Max 5x at $100 gets $100. Max 20x at $200 gets $200. Credits are billed at API rates. They expire at month's end.

Run the numbers on what $200 in API credits actually buys. Sonnet 4.6 — the default model for most agentic workloads — costs $3 per million input tokens and $15 per million output tokens. Without prompt caching, a heavy agentic session consuming 300,000 tokens per request at the output rate exhausts $200 in roughly 13 million output tokens. A full working day of aggressive autonomous agent usage — the workflow that made OpenClaw worth adopting — can drain that credit pool inside a week.

The Batch API's 50% discount applies to asynchronous jobs. Prompt caching at 0.1x input rates would cut costs dramatically. But both optimizations require architectural cooperation from the harness. The same cooperation that OpenClaw's current system prompt design does not provide at meaningful rates.

Calling this a reinstatement is technically correct and practically misleading. It is a migration from flat-rate to metered billing with a monthly coupon. Light users will not notice. The power users who triggered the original ban will hit the credit ceiling by mid-month and face standard API pricing for the remainder.

So What Do You Actually Do

If you are running OpenClaw or any third-party harness against Claude, audit your actual token consumption before June 15. OpenClaw's session logs track input and output tokens per request. Multiply by your daily request volume. Compare the total to your subscription tier's credit allotment at API rates — not subscription rates. The gap between those two numbers is your new monthly cost exposure.

If the gap is under $50 per month, the Agent SDK credit absorbs most of it. Your subscription remains the cheapest path and you can stop reading here. If the gap is large, you have two levers worth pulling before switching providers.

First: fix your cache hit rate. The architectural change documented in GitHub Issue #49700 — separating static and dynamic system prompt components so the cache prefix stays stable — can move hit rates from 10% to above 90%. At Anthropic's 0.1x cache read pricing, that turns $30 uncached requests into $3 cached ones. Second: route batch-eligible workloads through the Batch API for the additional 50% discount. Combined, these two optimizations can reduce effective costs by up to 95% relative to OpenClaw's default uncached, synchronous configuration.

$200 in Agent SDK credits at a 10% cache hit rate buys roughly one week of heavy agentic use. The same $200 at a 90% hit rate, routed through the Batch API, stretches to approximately two months. An 8:1 difference. That ratio is the number that should drive every decision you make before June 15 — not your subscription tier, not your model preference, not which lab's creator ban offended you more. Your cache hit rate. Everything downstream of it follows.

FAQ

What exactly did Anthropic block in April 2026?

Anthropic prevented Claude Pro and Max subscribers from using their flat-rate subscription access with third-party AI agent frameworks, effective April 4, 2026. The policy change meant OpenClaw and similar harnesses could no longer authenticate through subscription OAuth tokens. Users who wanted to continue running third-party agents against Claude models needed to switch to direct API billing at standard per-token rates, removing the flat-rate cost advantage that had made agentic workloads economically attractive on subscription plans.

Was OpenClaw's creator permanently banned from Claude?

No. Peter Steinberger's personal account was suspended temporarily around April 10, 2026. Anthropic cited "suspicious signals" and a Usage Policy violation in the suspension email. Steinberger stated publicly he was already complying with the new API-only rules. His account was reinstated within hours. He had left OpenClaw in February 2026 to join OpenAI, where Sam Altman announced he would work on next-generation personal agents.

What are Agent SDK credits and when do they take effect?

Agent SDK credits are a new programmatic usage allowance Anthropic announced on May 13, 2026, taking effect June 15. Each paid subscriber receives monthly credits matching their subscription price — $20 for Pro, $100 for Max 5x, $200 for Max 20x. Credits are billed at API token rates, not absorbed by the flat-rate subscription, and they do not roll over to the following month. They apply to third-party agents including OpenClaw.

Why was OpenClaw so expensive for Anthropic to support on subscriptions?

OpenClaw's system prompt architecture concatenates static and dynamic content into a single string that changes between requests. Anthropic's prompt caching requires an exact prefix match, so any mutation invalidates the cache. This produced roughly a 10% cache hit rate, meaning each request reprocessed 100,000 to 300,000 tokens from scratch. A request costing $3 in cached compute was costing Anthropic approximately $30 uncached, and heavy users were generating hundreds of these daily against a $200 flat subscription.

How much does Claude API access cost per model as of May 2026?

Opus 4.7 is priced at $5 input and $25 output per million tokens. Sonnet 4.6 costs $3 input and $15 output. Haiku 4.5 runs $1 input and $5 output. The Batch API halves all of these rates for asynchronous processing. Prompt cache reads are billed at 0.1x the base input rate, delivering up to 90% savings on cached input tokens. Combining both discounts can reduce effective costs by up to 95%.

Will the Agent SDK credits be enough for heavy OpenClaw usage?

For most power users, no. A $200 Max 20x credit pool at Sonnet 4.6 output pricing covers roughly 13 million output tokens — enough for light to moderate agentic use but insufficient for full-day autonomous agent loops. Without prompt caching optimization, heavy users will exhaust credits within the first week of each billing cycle and face standard API rates for the remainder of the month.

What is the single most effective way to cut OpenClaw costs on Claude after June 15?

Improve your prompt cache hit rate. The fix documented in OpenClaw GitHub Issue #49700 — restructuring the system prompt to separate static from dynamic components — can lift cache hit rates from approximately 10% to above 90%. At Anthropic's 0.1x cache read rate, this converts a $30 uncached request into roughly $3. Combined with the Batch API's 50% discount on asynchronous jobs, the same $200 in monthly credits can stretch from one week of coverage to approximately two months.

Yes. In February 2026, approximately six weeks before the April enforcement, Anthropic added explicit language to its terms of service: "The use of OAuth tokens obtained via Claude Free, Pro, or Max accounts in any other product, tool, or service is not permitted." This clause established the contractual basis for blocking third-party harness authentication through subscription accounts and preceded the broader policy enforcement by over a month.