The Great AI Exodus: Why Anthropic Just Made Open Models the Pragmatic Choice

The commentary around Anthropic's mandatory identity verification for Claude has been predictable. Privacy advocates are alarmed about handing a government ID and a live selfie to Persona Identities , a company backed by Peter Thiel's Founders Fund. Users outside the US feel locked out of a tool they depend on. Lawyers are writing blog posts about fourth amendment implications.

All of that is valid. Nearly all of it misses the point.

This isn't a story about privacy. It's a story about lock-in — and the moment it finally broke.

The ID That Broke the Camel's Back

Anthropic's policy, which requires a government-issued photo ID plus a live selfie to continue using Claude, went from quietly published in April 2026 to prominently enforced around June 21-22. The company states it's "to prevent abuse, enforce our usage policies, and comply with legal obligations." The data is held by Persona, not Anthropic.

Individual users should find this invasive. For teams building production workflows on Claude, it's a supply chain event. If your entire code review pipeline, your agent orchestration layer, or your customer-facing AI product depends on a model that can cut you off because your passport expired — you have a problem.

What sets this moment apart from every previous "users hate this closed model" cycle: the alternatives are actually ready.

"What sets this moment apart from every previous 'users hate this closed model' cycle: the alternatives are actually ready."

The Open-Model Readiness Gap Has Closed

Six months ago, switching off Claude meant swallowing a capability hit. Open models were catching up, sure, but the professional penalty was real. Andrew Marble made this exact argument in his June 21 analysis , concluding that the gap has narrowed to the point where "there is minimal downside to switching to open models."

The numbers back him up.

GLM-5.2, released June 16 under an MIT license , is the strongest open-weights model ever tested. Artificial Analysis ranks it first among open models on their Intelligence Index with a score of 51, ahead of MiniMax-M3 (44) and DeepSeek V4 Pro (44). Simon Willison called it "probably the most powerful text-only open weights LLM" .

The real shock is the pricing:

The Tech Stackups comparison tells the same story: both models built a 3D WebGL platformer from scratch. Opus was faster and cleaner. GLM-5.2 cost $5.39 vs Opus's ~$21.92 — roughly a quarter of the price. And GLM-5.2's weights can be downloaded and run anywhere, forever, under a license that can't be revoked.

The Supply Chain Risk

One part of this story keeps getting overlooked.

On June 12 — ten days before the ID verification enforcement — the US government forced Anthropic to suspend access to its most powerful models, Fable 5 and Mythos 5, for all foreign nationals, including Anthropic's own employees overseas. A government directive citing national security shut down production access to the best models on the market, effective immediately.

Anthropic complied. Customers who had built workflows on Fable 5 lost access overnight.

This isn't a hypothetical. It happened last week. If you're an EU-based startup building an agentic workflow on Anthropic's top model, you now know that a US government directive can terminate that access with zero notice. The ID verification requirement, combined with the Fable suspension, creates a one-two punch: your access can be restricted by policy, and it can be cut off by government order.

Open weights can't be revoked. A downloaded model can't be remotely disabled. This isn't an abstract freedom argument — it's a concrete risk calculation.

"Open weights can't be revoked. A downloaded model can't be remotely disabled. This isn't an abstract freedom argument — it's a concrete risk calculation."

The Economic Argument

The old math was simple: proprietary models were better. You paid more, but you got more. The professional cost of switching was real because the capability gap was real.

That math has inverted for a growing set of use cases.

GLM-5.2 costs roughly 1/5 the price of Opus on output tokens. You can run it yourself on your own hardware, or call it through any of the nine providers on OpenRouter . There are no rate limits. No identity checks. No government directives.

And the gap is narrowing fast. GLM-5.2 beats Opus on AIME 2026 (99.2 vs 95.7). It trails on SWE-bench Pro (62.1 vs 69.2) but leads the open-weights pack. With every release cycle, the margin shrinks a bit more.

The question is no longer "is the open model good enough." It's "how much are you willing to pay for that last 5-10% of capability, and are you willing to accept the access risk that comes with it?"

What This Means

Three shifts are converging at once:

1. Anthropic is tightening access — mandatory ID, government-mandated model suspensions, a growing compliance burden
2. Open models crossed a threshold — GLM-5.2, Qwen 3 fine-tuning, and the Apertus sovereign AI infrastructure platform mean production-ready alternatives now exist
3. The switching cost collapsed — the tools to self-host, the middleware to bridge APIs, and the community knowledge to make it work have matured

The developers leaving Claude aren't just the privacy-conscious fringe. They're EU engineering teams who just realized their entire infrastructure rests on a foreign government's export control policy. They're startups that can't afford the 5x price premium. They're anyone who's ever had a CI pipeline break because an API key got revoked.

This is the great AI exodus — not a dramatic overnight migration, but a structural shift in the industry's center of gravity. The arguments for open models used to be ideological. Now they're pragmatic.

And pragmatic wins every time.

Further Reading

1. GLM-5.2 vs Claude Opus — Tech Stackups — A head-to-head comparison building a 3D WebGL platformer, with cost, speed, and quality data for both models.
2. GLM-5.2 on Artificial Analysis — Independent benchmarks showing GLM-5.2 as the leading open-weights model across agentic and reasoning tasks.
3. Anthropic Statement on Fable 5 and Mythos 5 Suspension — The June 12, 2026 government directive that forced Anthropic to cut access to its most capable models, demonstrating the supply chain risk inherent in closed models.
4. Simon Willison — GLM-5.2 Review — Hands-on testing and commentary from one of the most respected voices in the AI community, confirming the model's capability and limitations.
5. Claude Opus 4.8 Announcement — Anthropic's own announcement of their current best available model, including benchmark data and pricing used in this analysis.