This was the week the AI industry stopped pretending subscriptions were sustainable. Every major story circled the same truth: the era of subsidized inference is ending, and everyone from individual developers to trillion-dollar platforms is adjusting at the same time.
The billing shift
GitHub Copilot announced usage-based billing effective June 1. Opus multipliers jumped from 3x to 27x. The $10/month plan still costs $10; it buys substantially less. The announcement reached 680 points on Hacker News.
Anthropic's HERMES.md billing bug hit #1 on HN at 1,126 points. Commit messages containing certain strings were silently routing requests to extra-usage billing. A separate bug injected a malware-scanning system prompt into every file read, wasting 20,000-40,000 tokens per session. Users paid for both.
ChatGPT launched a full advertising platform with a four-token attribution system, a merchant-side tracking SDK, and 720 hours of cookie persistence. Sam Altman called ads "a last resort" less than two years ago. The attribution loop reached 453 points on HN.
The pattern: every vendor that launched with flat-rate pricing has walked it back. Anthropic moved to pay-per-overage. Windsurf replaced credits with quotas. GitHub moved to usage-based billing. OpenAI added ads. The subsidy era lasted about 18 months.
The GitHub exodus
Mitchell Hashimoto, co-founder of HashiCorp, announced Ghostty is leaving GitHub after documenting daily outages that made development "impossible." The post reached 3,342 points and 993 comments, the highest-voted developer story of the month. The Register ran it under the headline: "GitHub 'no longer a place for serious work.'"
Within 48 hours, the Netherlands announced an open-source code platform for government (508 points). A developer proposed federated code hosting (529 points). Zed, an AI-native editor, hit 1.0 (1,831 points). HardenedBSD moved to Radicle. Wiz published a GitHub RCE vulnerability breakdown (411 points). A security researcher called GitHub Actions "the weakest link" (213 points).
Eight stories about one platform, on the front page of HN, in one week. The migration conversation moved from infrastructure teams to mainstream developers.
The model landscape
Microsoft and OpenAI ended their exclusive revenue-sharing deal (943 points). The same week, OpenAI announced its models would be available on Amazon Bedrock. The divorce and the diversification happened simultaneously.
OpenAI published "Where the goblins came from" (680 points), explaining how reinforcement learning rewards for a "Nerdy" personality mode accidentally trained the model to insert creature metaphors into unrelated responses. The fix was a line in the system prompt banning goblins. The transparency was welcome; the implication that reward signals leak unpredictably across training conditions was less reassuring.
Mistral Medium 3.5 launched. Claude.ai had two separate outages. The 4TB Mercor voice data breach exposed 40,000 AI contractors.
The ownership question
"Who owns the code Claude Code wrote?" reached 484 points and 447 comments, the most controversial thread of the week. Zig banned AI-generated contributions entirely, arguing that the project invests in contributors, not contributions. Vera, a new language designed for machines to write, launched the same week. A research paper showed that finetuning can activate verbatim recall of copyrighted books in LLMs.
Each of these is a different surface of the same question: when AI writes code, who is responsible for what it produces?
The vibe
Last month's vibe was "rationing." This week's vibe is "reckoning." The subsidies are ending. The platforms are fragmenting. The ownership is contested. The models do things their creators did not predict and cannot fully explain. Developer sentiment in HN threads split into two camps: people angry about the changes and people already trading tips on how to adapt. The anger is about the price catching up to the value. The adaptation is already happening.