Anthropic Banned Claude Pro/Max from Third-Party Routers on April 4, 2026. Here's What Actually Happened.

On April 4, 2026, Anthropic updated its terms and sent notices to users: Claude Pro and Max subscriptions could no longer be used through third-party routers, proxies, or “smart model pickers.” The language was direct. The enforcement was not.

This was not the first signal. It was the loudest one yet. And it exposed the structural weakness in an entire category of products that had been built on the assumption that consumer subscriptions could be arbitraged at scale.

The announcement and the ToS language

Anthropic’s update (and the accompanying email to Pro/Max subscribers) included language that had been tightening for months:

“You may not access, use, or resell Claude Pro or Claude Max through any third-party service, proxy, aggregator, or routing layer that is not explicitly authorized by Anthropic. This includes services that pool, share, or resell access to consumer subscriptions.”

Similar clauses had appeared in xAI’s June 2026 updates (anti-resale language targeting “unofficial clients and routers”) and in OpenAI’s signals around Codex CLI fingerprinting on non-native platforms. Anthropic simply made the ban explicit and tied it to the consumer subscription tier.

The date matters. April 4, 2026, is now the reference point for when the gray-market subscription router era entered its endgame.

Timeline of warnings (not surprises)

Mid-2025 to early 2026: Multiple providers added or strengthened “no resale / no proxy” language in ToS. Most router operators treated it as unenforced boilerplate.
June 2026 (xAI): Public statements and ToS updates targeting “unofficial clients” and services that “aggregate or resell” Grok consumer access.
Late 2025 – 2026: Reports of OpenAI Codex on Linux using rustls producing detectable fingerprints, with enforcement actions against certain wrapper tools.
April 4, 2026 (Anthropic): Explicit ban + notice for Claude Pro/Max. Gray-market routers that relied on OAuth or shared consumer sessions began losing access or going dark.

The pattern was visible to anyone reading the ToS diffs and the public statements. The routers that died or pivoted after April 4 were not victims of sudden policy; they were the ones that had bet the policy would never be enforced.

Impact on the gray-market layer

Several classes of products were directly affected:

Routers that logged in to consumer accounts and exposed a unified API or chat interface.
“One key for all your subs” services that used OAuth flows or session sharing.
Tools that presented themselves as “smart pickers” while actually fronting Pro/Max access.

Some shut down the Claude path. Others went fully dark or rebranded as “enterprise only” (with first-party keys required). A few attempted workarounds (residential proxies, browser automation) and were subsequently detected and banned.

In RoutePlane’s ToS-Risk classification, these paths are Tier 3 (🔴 High):

Gray-market OAuth / reverse-engineered / actively detected.
Continued use requires the user to acknowledge they bear the ban risk.

We do not block Tier 3 by default. We surface it at session start and in routeplane status. The user decides. But we do not pretend the risk does not exist.

Why this was structural, not tactical

The deeper issue is economic and contractual.

Consumer subscriptions (Claude Pro/Max, ChatGPT Plus, Grok consumer, etc.) are priced for individual use with rate limits and fair-use assumptions. When a router pools hundreds or thousands of those subscriptions behind a single interface, two things happen:

The provider sees anomalous usage patterns (volume, distribution of prompts, geographic signals, TLS fingerprints).
The economic model the provider sold to the subscriber is violated at scale.

Providers had been clear in their contracts that this was not allowed. The routers that treated the clause as marketing language rather than a hard boundary were always one enforcement wave away from breakage.

After April 4, the question for any router operator became: “What happens when the next provider follows?”

The honest path forward

There are two durable ways to route across providers without building on contested ground:

First-party API keys (Tier 1 🟢). OpenAI, Anthropic, Google, Mistral, OpenRouter, etc. These are the accounts you pay the provider directly. No resale, no proxy, no shared consumer session. This is the path RoutePlane is designed around.
Documented, provider-supported harnesses (Tier 2 🟡). Examples include GitHub Copilot via the official extension, or ChatGPT Codex CLI on macOS using the platform’s native TLS stack. These are still subscription-based, but they follow the path the provider has explicitly blessed.

Everything else — shared consumer logins, reverse-engineered OAuth, residential proxy pools, “undetectable” wrappers — is Tier 3. It may work today. It may stop working tomorrow. The user owns the risk.

What this means if you route your own traffic

We build RoutePlane as a local routing daemon. The graveyard shaped the design:

Routing policy lives in a YAML file on your machine. It cannot be taken away by a provider changing a ToS or a hosted service pivoting.
We prefer explicit policy (tiers, failover, cost ceilings) over learned classifiers that rot with every model release.
We do not put ourselves in the token path. You pay the providers. We sell the router, once.

The April 4 ban did not kill local-first, BYOK routing. It killed the illusion that you could reliably arbitrage consumer subscriptions through someone else’s hosted layer without eventually being noticed.

If you are still routing Claude Pro or Max through a third-party service that is not you and is not Anthropic, you are on Tier 3. RoutePlane will tell you that instead of hiding it.

CTA: If you route your own traffic and want a local daemon that surfaces ToS risk instead of pretending it doesn’t exist, see pricing or the activation guide.

Sources and references (as of 2026-07): Anthropic April 4, 2026 announcement and ToS update; xAI June 2026 anti-resale statements; public reports on Codex fingerprinting enforcement; RoutePlane ToS-Risk tier definitions in the daemon and documentation. Specific router names that went dark are omitted or generalized where public confirmation is incomplete.