This brief covers AI news from 2026-10-27 UTC.
Anthropic opens its books, OpenAI ships a cheaper computer-use model, and Stripe buys the router that sits in front of everyone’s models.
Anthropic files S-1, disclosing $11.6B quarterly revenue
What happened: Anthropic filed a Form S-1 with the SEC on Oct 26, becoming the first major AI lab to file for public markets. The filing discloses $11.6 billion in quarterly revenue, a first operating profit, and roughly $800 billion in long-term compute commitments.
Why it matters: The filing puts Anthropic’s model economics and compute obligations on the public record, giving builders a rare look at the costs behind the API they call.
OpenAI ships GPT-6.1 Terra at $1 input, $5 output
What happened: OpenAI released GPT-6.1 Terra on Oct 26, a vision and computer-use model priced at $1 per 1M input tokens and $5 per 1M output tokens with a 1.05M context window, per its changelog.
Why it matters: A long-context model with computer use at that input price changes what a screen-driving agent costs to run at volume.
Google, Amazon, Broadcom line up $77B for Anthropic
What happened: Anthropic has lined up $77 billion in additional financing from Google, Amazon and Broadcom, each with its own conditions. Google’s $20 billion commitment is contingent partly on Anthropic meeting certain revenue targets, and Amazon’s is tied to computing capacity milestones.
Why it matters: Anthropic’s compute supply now runs through three vendors with performance strings attached, which shapes how much capacity its models can draw on.
Stripe to buy OpenRouter for more than $8B
What happened: Stripe has agreed to acquire AI model-routing platform OpenRouter in a deal valued by sources at more than eight billion dollars, according to a report citing Reuters.
Why it matters: OpenRouter is the switchboard many apps use to swap between models, so a payments company owning it puts billing and routing under one roof.
IBM makes Spyre a first-class PyTorch backend
What happened: IBM said torch-spyre, Spyre’s vLLM plugin, compiler improvements and CI work now make its inference accelerator a first-class PyTorch backend, delivering on a plan IBM and the PyTorch Foundation outlined in October 2025.
Why it matters: If ordinary PyTorch code compiles to Spyre, teams running inference on IBM Z, LinuxONE and Power can use their existing models and workflows instead of porting.
Cadence and Nvidia launch Maestro models for chip design
What happened: Cadence announced a partnership with Nvidia and launched its Maestro AI model family, fine-tuned on Nvidia’s Nemotron models to interpret the results of Cadence’s own EDA tools. Cadence senior VP Paul Cunningham said 20 to 30 percent of chips under development today are designed with the help of AI agents.
Why it matters: Domain-tuned models wired into existing verification tools are the pattern for agents in any workflow that already has a pass or fail test.
WebCoder-27B beats Claude Opus 4.8 on a frontend benchmark
What happened: Researchers from Tsinghua, Zhipu AI, Beihang and the University of Oslo trained WebCoder-27B from Qwen3.8-27B with GRPO in a persistent React/Vite environment. It scored 72.8 on requirement satisfaction on the new 40-task Webdev-Bench, against 64.8 for its base model, 65.1 for Claude Opus 4.8 and 59.6 for GPT-5.4.
Why it matters: Scoring the running app rather than the diff is a training recipe any team building a frontend agent can copy.