Detailed Analysis
Anthropic's Claude Pro and Max subscription tiers operate under Consumer Terms of Service that carry significant implications for users handling proprietary or business-sensitive code, a reality that diverges sharply from the company's earlier reputation for stringent data privacy. Since September 28, 2025, Free, Pro, and Max accounts have been governed by Consumer Terms under which training on user data is opt-in by default — meaning users must actively disable it — and data retention extends to five years when training is enabled, versus a 30-day window when it is not. Critically, Consumer-tier accounts receive no Data Processing Addendum (DPA), a legally binding document that governs how personal and sensitive business data is handled, which is a standard requirement for GDPR compliance and enterprise data governance. Notably, "extra usage" billing bolted onto Pro and Max plans does not elevate those sessions to Commercial Terms; they remain subject to Consumer protections regardless of the additional payment.
The article's headline claim that code is "not protected like you probably think" requires important disambiguation in light of available research. The framing concerns legal and contractual data protection — not technical security vulnerabilities. Claude Code's OAuth-based authentication system does include robust technical safeguards: tokens are scoped exclusively to verified Claude Code clients, telemetry enforcement rejects unauthorized third-party proxies, and a permission system defaults to read-only access with explicit approval required for writes or system-level commands. These measures effectively ended subscription-sharing exploits like OpenClaw by mid-2025. The article's concern is therefore not that data is technically intercepted or exposed to unauthorized parties, but that the contractual and legal framework governing what Anthropic itself may do with that data is materially weaker on consumer plans than on API-based Commercial accounts.
The legal dimensions are particularly significant for businesses operating under EU trade secret law — such as Germany's GeschGehG — or similar frameworks in other jurisdictions. Those regimes typically require that organizations take "reasonable protection measures" to maintain the legal status of trade secrets. Routing proprietary source code through a Consumer-tier AI service that lacks a DPA and permits potential training on that data could be construed as failing to meet that threshold, with downstream consequences during IP disputes, acquisitions, or regulatory due diligence. By contrast, Anthropic's API accessed with a user's own key falls under Commercial Terms, which include no training on customer data, a DPA available through the Console, and cleaner GDPR standing — making it the legally defensible path for sensitive work.
The cost differential between these two tiers, however, is substantial enough to represent a genuine business decision rather than a trivial upgrade. Anthropic's own reported figures suggest the average Claude Code developer on API billing incurs roughly $6 per day in costs, with intensive agentic workloads reaching $20–50 per day — compared to $20 per month for Pro or $100–200 per month for Max. Depending on usage intensity, the API route can represent a 2–5x cost increase, meaning the Commercial Terms and DPA are effectively paid features accessed through pricing rather than subscription tier. This creates a two-tiered privacy landscape in which stronger legal protections are structurally inaccessible to individual developers or small teams whose economics make the flat-rate subscription the only viable option.
This situation reflects a broader tension emerging across the AI industry as subscription-based consumer products scale into professional and enterprise workflows without corresponding evolution in their legal frameworks. The policy change that introduced opt-in training for Pro and Max users was implemented quietly in late September 2025 amid a period of competing major announcements, limiting its visibility among the developer community. As AI coding assistants become deeply embedded in commercial software development, the divergence between consumer-facing marketing narratives — emphasizing privacy and safety — and the granular legal realities of Consumer versus Commercial Terms is increasingly consequential. Developers and businesses integrating these tools into production pipelines will need to treat data governance as a first-class technical requirement rather than an afterthought, auditing not just the tools they use but the specific contractual tier under which those tools operate.
Read original article →