Anthropic files its S-1, coding agents cost up to 75% less to run, and Microsoft offers 60,000 buyouts

On Monday's HN front page, Anthropic filed its S-1, the paperwork to sell its shares to the public. Anthropic brought in about 16 billion dollars in the first half of 2026, more than three times its revenue for all of 2025. For anyone who builds on Claude or Claude Code, the business behind the tools you pay for is now out in the open: what it earns, what it loses, and which risks it worries about.

The loss for those six months is almost 81 billion dollars. About 72 billion of it is on paper: convertible notes and SAFEs that were marked up as Anthropic's valuation climbed. That leaves a loss of roughly 8 billion dollars. The filing names open-weight models as a source of pressure on its prices, and that pressure is already showing up in what you pay per token.

Inside Anthropic, Claude wrote more than 90 percent of the contributions to the company's own codebase in September. Each engineer now adds more than 45 times as much code a day as the average in 2024. Anthropic has about 5,300 employees, and it builds Claude Code mostly with Claude Code, so the tool you use is the one its own engineers depend on every day.

Running a coding agent costs a lot less than it did in June. In an agent session, cached tokens are 96 to 99 percent of what you pay for, because every turn sends the whole conversation again. So the cut that matters is the cache price. Opus 5.5 charges 20 cents per million cached tokens, down from 50 on Opus 5, and GPT-6.1 Sol charges 10 cents. Together that makes Opus 5.5 about half the cost of Opus 5, and the Sol line about 75 percent cheaper. If you're choosing a model for agent work, compare the cache price first.

IBM's chief executive, Arvind Krishna, told investors that open-weight models are now about 90 percent as good as the frontier models, for a tenth to a quarter of the token cost. He puts most enterprise work in that cheaper tier, such as summarizing an email, examining a contract or answering a customer call. It fits in two or three racks on-prem, and the hard part is staff who can cluster the GPUs and cache the results. In the thread, jacquesm runs an eight-GPU rig and two DGX Sparks, and hasn't paid for a model in months.

Microsoft is offering voluntary buyouts to about 60,000 employees, roughly a quarter of the company, with a standard severance package. Gaming, Windows and Copilot take part fully, the AI and cloud teams partly, and security stays out. Vice presidents and below in the United States, Canada, India and the UK have until early December to decide. In the thread, ecshafer lays out the bind: take the money and job-hunt against thousands of former colleagues, or stay and hope a layoff doesn't follow.