Good morning. It's Tuesday, October 27, and today we're breaking down the numbers behind Anthropic’s massive IPO, why agent harnesses can drive up costs, and OpenAI’s push to make computer-using AI faster and cheaper.
Plus: a new poll on Anthropic’s IPO and a guest essay exploring China’s progress in the AI race.
OpenAI released GPT-6.1 Terra, a model built to understand screens and work inside software. At medium reasoning effort, it scored 15 percentage points higher than GPT-5.6 Terra on professional computer-use tasks at about one-third the cost; API pricing starts at $1 per million input tokens and $5 per million output tokens.
Anthropic filed to go public on Nasdaq under the ticker ANT, potentially becoming one of the largest IPOs in history. Its filing highlights risks beyond spending, including heavy reliance on a handful of major customers and cloud partners. Google could provide up to $20 billion in additional financing, while Amazon holds $6.2 billion in convertible notes.
More than half of Anthropic’s customers spending at least $100,000 a year now use both the Claude developer platform and an app such as Claude Code or Cowork, up from about 30% at the end of 2025. Anthropic says this cross-selling helped existing customers increase their spending sharply as Claude spread across more teams and tasks.
Microsoft offered voluntary exit packages to about 60,000 employees—roughly 27% of its global workforce—in the United States, Canada, India and the United Kingdom. The company is trying to contain costs while spending heavily on AI; the offer makes those employees eligible to leave but does not mean all 60,000 will accept.
SPONSORED BY FACTORY
Clear Your Engineering Backlog
Factory is a self-improving system for your software development lifecycle. Droids write, review, validate, deploy, document, and monitor software across the full SDLC.
FORWARD FUTURE ORIGINAL
Is China Catching Up to the U.S. in AI?
The United States still owns the top of the AI leaderboard, with more money, computing power and several of the world’s best proprietary models. China has built a different advantage: more labs close to the frontier, the strongest open-weight models and many of the lowest-cost options for real work.
Anastasios Angelopoulos, cofounder and CEO of Ar breaks the race into three questions—who leads at the frontier, who has the stronger model ecosystem and who can deliver useful intelligence at the lowest cost. → Read the full article here.
IPOs
The Numbers Behind Anthropic’s IPO
Anthropic revealed $16.3 billion in revenue for the first half of 2026, a 14-fold increase from a year earlier, ahead of a potential $2 trillion IPO. The company reported an $80.7 billion net loss, largely driven by a one-time $72.4 billion accounting charge, while its operating loss was $693 million.
The filing also revealed up to $807 billion in infrastructure commitments over the next decade, including contracts for data centers, chips, and computing capacity. Anthropic spent nearly $9 billion on compute in the first half alone and warned it may need additional financing even after its IPO. The central question for investors is whether its explosive revenue growth can justify its enormous spending commitments. → Read the full article here. (Paywall)
POLL
Are You Buying Anthropic?
Anthropic is reportedly preparing to go public, giving investors a chance to bet directly on one of the leading AI companies.
Would you invest? |
AI HARNESS
An Agent’s Harness Can Triple Its Cost
A study from Shandong University found that more complex AI coding agent frameworks can cost up to three times as much without improving accuracy. Researchers tested Claude Code, OpenCode, and mini-SWE-agent across nearly 500 coding tasks and found that Claude Code and the much simpler mini-SWE-agent achieved statistically equivalent success rates.
The difference came down to overhead. Claude Code started with more than 16,000 tokens of context, compared with just 829 for mini-SWE-agent. The findings suggest that adding sophisticated tools, prompts, and orchestration doesn't necessarily make coding agents more effective, raising questions about whether the added complexity justifies the cost. Read the full article here.
HEADLINES
In Other News
AI-Generated 'Odyssey': Grok debuted Episode 1 of its AI-generated series of Homer’s ‘Odyssey’.
Claude Accesses Vote Counts: Anthropic says Claude skirted a paywall to get public Georgia vote totals, not voter records.
$40 Billion for AI Chips: Musk says SpaceX closed $40B in debt for GPUs expected to arrive in the first half of 2027.
Traders Build AI Infrastructure: Jane Street runs 4,000+ NVIDIA GPUs; Hudson River spends $1B a year on AI.
$95 Billion in Stock Sales: Meta will sell Class A, preferred and Class C shares to fund AI buildout.
AWS Infrastructure Deal: Verizon and AWS signed a $2B+ deal expanding fiber and compute for AI workloads.
