Skip to content
S-compass SapientTech.dev
Go back

Anthropic's IPO filing puts numbers on Claude Code's economics

·
· Markdown

Coverage: the past 7 days. Searched “Claude Code”, “Anthropic”, “Claude Code Anthropic departures”, “Claude usage limits”, “Claude Agent SDK” on News, Hacker News, Reddit, GitHub, and X.

Anthropic filed its S-1 on October 26. The same week two Claude Code builders left, users reported smaller limits, Claude refused reverse engineering, and Duet opened a waitlist.

Anthropic filed its S-1 on October 26 to list on Nasdaq as ANT. First-half revenue was $16.3 billion, and the $80 billion net loss includes a $72 billion non-cash charge. The S-1 is Anthropic's first public account of its revenue, costs, and customer concentration, including how its heaviest Claude Code users spend.

Evidence and sources
Net revenue retention is about 900%. 11,900 customers pay over $100,000, 1,900 over $1 million, and about 250 over $10 million.
Opus 5.5 cut customer cost per task about 40% against Opus 5. Compute fell from 201% of revenue to 55%.
Claude Code's top enterprise users outspend other Claude users 20 to 1, per The New Stack's reading of the filing.

Anthropic opens a Claude Duet waitlist, ships MCP Tasks, donates Agent Skills

Claude Duet opened a Pro and Max waitlist October 21. Anthropic donated Agent Skills to the Agentic AI Foundation October 22. Claude Code gained MCP Tasks and Elicitations October 23. MCP Tasks need Claude Code 2.1.304 or later in the CLI, Agent SDK, or GitHub Actions, not VS Code, desktop, or Bedrock.

Evidence and sources
Duet is an agent on Opus 5.5 that reads email and calendar and works while you sleep, on the same usage limits.
Engineers from Anthropic, Earendil, Google, Microsoft, OpenAI, and OpenCode govern Agent Skills as a steering committee. Licenses do not change.
A server can now hand Claude a long job through MCP Tasks and pause mid-task to ask the user something through Elicitations.

Three Claude Code security findings in one week

Noma disclosed SneakPeek, a CVSS 8.7 Claude Code Desktop preview flaw patched in July. The Python Agent SDK fixed argument injection CVE-2026-96620. Claude Code put git emails into HTTP headers. The email leak reported in April continues, and one commenter blocks it with a PreToolUse hook that greps each tool call.

Evidence and sources
Files over 256 KB loaded via file: URLs, so a script could read and send SSH keys. Anthropic patched it July 7.
GitHub issue 53231 reported the email leak on April 25. The top Reddit comment says Claude sent a git email to Wikipedia.
One commenter calls a CLAUDE.md line a request the model weighs and blocks the call with a PreToolUse hook instead.

SemiAnalysis finds a Max limits A/B test as users report smaller limits

SemiAnalysis reported an A/B test October 22: some Max accounts get 20% less weekly budget. Theo Browne confirmed an Agent SDK bug October 25 that drains the 5-hour limit faster. Anthropic has not confirmed a test. On Reddit, Max 20x users report weekly budgets draining faster on unchanged workflows.

Evidence and sources
SemiAnalysis measured two Max 20x accounts the same day: weekly budgets of 5,266 million and 6,579 million tokens, 5-hour windows 19% apart.
Anthropic's Boris Cherny and Lydia Hallie had both told users that limits had not changed before SemiAnalysis published.
The bug leaves weekly limits alone. Agent SDK work draws on monthly API credit: $100 on Max 5x, $200 on Max 20x.

Previous Post
Today's top AI news, October 27, 2026
Next Post
A Fable 5.5 leak turned out to be a model mix-up