There is minimal downside to switching to open models
TLDR Dev
Claude's new ID verification is pushing one dev toward open-weight models instead. He says the switch barely hurts anymore.
Andrew Marble, writing on June 21, has a confession: he used to treat open-weight models the way he once treated Linux back in the mid-2000s — a hobbyist's plaything, fine for tinkering, risky for real work. That's changing fast, and the trigger isn't better open models alone. It's Claude's rollout of identity verification, a move Marble says he saw coming given the recent pile-up of "safeguards" and the general vibe shift around how frontier labs treat their users.
His argument leans on a specific analogy. Choosing Linux over Windows in 2006 meant real professional risk: garbled Word docs, missing PowerPoint fidelity, half-finished open-source tools that almost did the job. He stuck with Windows through grad school, he admits, until Matlab finally pushed him off. Today, he says, that gap has nearly closed — most productivity software runs in a browser now, and open-source tooling has matured into something you don't have to apologize for using.
LLMs, he argues, aren't quite there, but they're closer than people assume. Every major leaderboard, including Artificial Analysis on the day he's writing, still gets topped by Claude and GPT. Compatibility still favors the proprietary side too — Claude Code just works, the APIs are polished, and crucially, nobody blinks if you tell a client you're routing data through OpenAI or Anthropic. Route the same query through DeepSeek or OpenRouter and suddenly people get twitchy about privacy, deserved or not. Self-hosting solves the privacy problem but trades it for cost, complexity, and latency — pick at least two.
What's shifted, in Marble's telling, is the size of the penalty for going open. He's already running a stable of open-weight models locally and in the cloud, coding harnesses for them have gotten genuinely usable, and the performance gap to the frontier has shrunk to a matter of months rather than years. He's careful not to frame this as some ideological stand against ID verification — he just doesn't want to do it, and he's more interested in whether skipping the top proprietary models actually costs him anything at work. His bet is that it won't cost much: some short-term friction, sure, but nothing like the Matlab-to-Octave wall he hit years ago.
My take
I'll take the open-weight tax over handing Anthropic my ID any day — this is exactly the kind of forced choice that pushes serious users toward open models faster than any benchmark chart ever could. Labs keep treating 'trust us' as a business model, and eventually enough professionals decide the compatibility gap is a smaller price than the privacy one. MIT-licensed frontier-adjacent models showing up within months of the leaders is the real story here, buried under the ID-verification headline.
Read more about this at: TLDR Dev
