The "Goodbye Claude" Videos Are a Leading Indicator
A wave of developers is posting that open source has caught up and they're canceling their frontier subscriptions. The vibe is early and a little performative. The data underneath it is real.
We build this newsletter on two signals we watch obsessively: what the experts we track share with each other, and what is actually climbing across the app charts and community feeds. (It is the same method behind our Q2 recap.) This week, in the very trend engine we use to assemble the issue, one of the fastest-rising AI videos was titled, with zero subtlety, "Open Source Is Back. Goodbye Claude." Tens of thousands of views, and it is not alone.
It is easy to wave that off as content. Developers love a declaration, and "I canceled my subscription" is a reliable way to get clicks. So we did what we always do: check whether the folk signal matches the hard data. This time it does.
Open source now moves more tokens than anyone
On Vercel's infrastructure, DeepSeek just passed everyone to handle just over a third of all tokens. On OpenRouter, DeepSeek's V4 Flash is processing about 5.3 trillion tokens a week while Anthropic's flagship Opus 4.8 handles just over 2 trillion. A year ago that ratio was unthinkable.
And the models keep landing. Tencent shipped Hy3 under a permissive Apache 2.0 license, beating a larger rival on nearly everything except code. Z.ai's GLM 5.2 now tops the open-weights intelligence rankings. Moonshot's Kimi K2.7 is good enough that Cognition built its new coding model on top of it. Alibaba's Qwen is everywhere. Even Mistral in Europe is committing to open weights. Most of the frontier-adjacent releases of the past month are things you can just download.
Read the fine print: they won the tokens, not the dollars
Here is where the "goodbye Claude" crowd gets ahead of itself. Token volume is not revenue, and the same data makes that plain. On that same Vercel dashboard, Anthropic still accounts for more than half of all AI spend. Opus costs roughly 23 times more per token than DeepSeek, so even as it moves fewer tokens it keeps capturing the lion's share of the money.
The cleanest framing I have seen comes from Decagon's CEO in that piece: frontier labs own discovery, open source owns production. The newest and hardest problems still need the most expensive model in the room. Once a workflow is understood and stable, it drifts down to whatever is cheap and good enough. So the labs keep the frontier and open source eats the back catalog. Both are true at the same time.
The thing actually collapsing is margin
If the labs aren't losing revenue, what are the "goodbye Claude" videos a leading indicator of? Pricing power.
A developer who put GLM 5.2 through Claude Code could not reliably tell it apart from Opus, at roughly $4.40 per million tokens, under a fifth of Opus's price. Both Z.ai and Fireworks expose Anthropic-compatible and OpenAI-compatible endpoints, so switching is a config change, not a rewrite. When "good enough" is one line of code away and five times cheaper, the ~90% gross margin the frontier labs earn on inference is the exposed number, well before their revenue is. Bezos said it first: your margin is my opportunity.
Watch the margin, not the mic-drop
None of this means Anthropic or OpenAI is in trouble this quarter. It means the moat was never the model weights. It is distribution, the tools built around the model, the trust to run it on your most sensitive work, and the shrinking set of problems only the frontier can solve today.
The developers filming "goodbye Claude" are early, and early is exactly what a leading indicator is. The token share moved first. The dollars follow the tokens, with a lag. We will keep watching both.
— Alexis