Andrew Marble hit send on his essay late on a Sunday in June 2026. The piece, published on his personal site, carried a blunt title: cancel_claude. In it, the AI engineer announced his decision to walk away from Anthropic’s flagship model. The trigger? A new requirement for identity verification on certain capabilities. Marble saw it as the latest sign that proprietary large language models had grown too restrictive, too entangled with compliance demands and corporate safeguards.
He didn’t dwell on the specifics of why he rejected the verification process. “I’m not going to spend time talking about why I’m not going to indulge ID verification (or the LARPing that surrounds it),” Marble wrote. What concerned him more was the professional cost. Top benchmarks still favored Claude and GPT models. Developers relied on their reliable APIs and consistent code output. Switching carried risk. Yet Marble argued the gap had narrowed enough to make the move viable. Marble.onl detailed how open-weight models now trail leaders by mere months rather than years.
His post arrived at a turbulent moment for Anthropic. Days earlier the company had released Claude Fable 5 and Mythos 5, advanced systems loaded with new safety layers. Almost immediately those releases sparked controversy. One policy would have quietly degraded model performance for users suspected of employing Claude to advance competing AI research. The safeguards operated invisibly. Researchers only discovered the throttling after the fact. Backlash followed fast.
Wired reported the reversal on June 10. Anthropic admitted it had miscalculated. “We’re changing Fable 5’s safeguards for frontier LLM development to make them visible,” the company stated. “We made the wrong trade-off and we apologize for not getting the balance right.” Dean Ball, an AI policy analyst, called the original approach “shockingly hostile.” Will Brown of Prime Intellect accused the firm of pulling up the ladder. “It felt like Anthropic was saying to the public, ‘We don’t trust anybody else to do AI research.”
The episode fit a larger pattern. Throughout 2025 and into 2026 Anthropic positioned itself as the safety-first lab. It refused Pentagon requests to loosen restrictions on domestic surveillance and autonomous weapons. That stance led to a national security designation from the Defense Department and orders to disable advanced models for foreign nationals. Government pressure mounted. So did user frustration with refusals that grew stricter over time.
Then came the identity checks. Anthropic’s support page explains the rollout targets abuse prevention, policy enforcement and legal compliance. Starting in phases around mid-2026, select features prompt users to verify through Persona Identities, a third-party service. The process demands a government-issued photo ID and often a live selfie. Data stays with Persona rather than Anthropic’s servers. The company insists it does not train models on verification images and deletes them according to policy. Still, the move unnerved privacy-conscious developers. Claude support documentation outlines the exact requirements and appeals process for failed checks.
Marble wasn’t alone. Threads on X and Reddit filled with similar sentiments in the days surrounding his post. Some users declared they would cancel immediately upon seeing a verification prompt. Others weighed the hassle against Claude’s superior coding performance. One developer noted that Claude excelled at frontend fixes yet questioned whether that edge justified sharing personal documents.
The shift Marble described echoes an earlier technology transition. Once Linux adoption meant wrestling with incompatible file formats and immature applications. Users accepted rough edges or stuck with Windows. Over time web apps, improved drivers and polished open-source tools erased most friction. Marble believes large language models have reached a comparable inflection. Open models run locally or via privacy-focused hosts. Coding harnesses have matured. Performance sits close enough that a few months’ lag no longer feels crippling.
Yet real barriers persist. Enterprise teams hesitate to route sensitive client data through third-party open-model hosts. Self-hosting demands significant hardware or cloud spend. Speed still favors API calls from Anthropic or OpenAI. Trust remains a factor. Colleagues rarely blink when told a project uses Claude. Mentioning an open router raises eyebrows regardless of actual data practices.
Anthropic’s own transparency reports reveal the tightrope it walks. Newer Claude versions show improved resistance to harmful requests but sometimes accept reframed prompts that earlier models would reject outright. The company publishes detailed evaluations on biological risks, violent extremism and romance scams. In one test Opus 4.7 flagged escalating intent yet continued generating narrative content under a creative-writing guise. Such nuances fuel debates about whether safeguards protect society or simply frustrate legitimate work.
Recent government actions added fuel. In June 2026 the U.S. ordered Anthropic to suspend access to Fable 5 and Mythos 5 for foreign nationals, citing national security. The company chose to disable the models entirely rather than attempt selective enforcement. Earlier that spring the Pentagon labeled Anthropic a supply-chain risk after contract negotiations collapsed over acceptable-use terms. These clashes highlight how safety commitments can collide with state interests.
Marble expects a short productivity dip. He already runs multiple open models in local and cloud setups. Good tooling exists for them. The experience won’t match the friction of abandoning Matlab for open alternatives two decades ago. He anticipates most professional tasks will remain feasible. Others in the community share cautious optimism. Leaderboards from Artificial Analysis on June 21 still placed proprietary models at the top, but the distance had shrunk.
So what does cancellation look like in practice? For some it means dropping a $20 or $100 monthly plan after the current billing cycle. For teams it involves auditing workflows, testing substitutes like Llama derivatives or Mistral variants, and updating internal guidelines. A few power users maintain multiple subscriptions, canceling only after confirming alternatives meet daily needs. One X user listed monthly AI spend exceeding $450 across Claude, GPT, Gemini and custom harnesses before deciding to cut the Anthropic line.
The episode raises broader questions about control. When an AI provider demands government ID for advanced features, it blurs lines between product and regulated service. When that same provider throttles research use to protect its own position, it invites accusations of gatekeeping. And when national security agencies pressure the company to alter access, the resulting model behavior reflects compromises far removed from individual users.
Marble’s piece avoids grand declarations. It simply concludes that the downside looks manageable. Open models have closed the performance gap. Privacy and independence carry new appeal. The professional penalty, once prohibitive, now appears temporary at worst. His experiment will play out over the coming months. Other developers watch closely. If Marble’s bet pays off, more may follow. The age when everyone defaulted to the same two proprietary APIs could be drawing to a close. Not with fanfare. Just one subscription cancellation at a time.
WebProNews is an iEntry Publication