Anthropic files to go public, and OpenAI ships GPT-6.1 Terra

Anthropic made its IPO prospectus public on October 26, a month after Reuters described the confidential draft. Revenue for the first half of 2026 was 16.3 billion dollars, up from 1.2 billion dollars a year earlier, and the operating loss narrowed to 693 million dollars. The net loss was 80.7 billion dollars, mostly a noncash charge on convertible notes. More than 11,900 customers now pay at least 100,000 dollars a year, up from about 950 a year before. The filing confirms that the seven co-founders will hold 50.1 percent of the vote through a single special share with no economic rights, while the Long-Term Benefit Trust elects directors who will make up a majority of the board. The price range and share count are still blank, and Goldman Sachs, Morgan Stanley and J.P. Morgan lead the underwriters. The risk factors also note that some planned data centers will run on their own natural-gas generators, leaving Anthropic exposed to fuel prices and pipeline limits.

GPT-6.1 Terra is OpenAI's new fast, low-cost model. It costs 1 dollar per million input tokens and 5 dollars per million output tokens, half the standard price of GPT-5.6 Terra and GPT-6.1 Sol. OpenAI says it scores 15 points higher than GPT-5.6 Terra on its Agents' Last Exam computer-use test at medium effort, roughly matching GPT-6 Astra at a tenth of the price per token. On a coding test run from the command line it solves 50.3 percent of tasks, more than double the older Terra, though OpenAI still points the hardest coding work to Sol and Astra. Its error rate on prompts designed to trip it up fell from 53.6 percent to 16.7 percent at low effort. Terra also crosses the "Critical" cybersecurity line in OpenAI's own safety framework, so it ships with the same protections as Astra. It is live in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API.

OpenAI posted three new results from its models: a claimed proof of the local smoothing conjecture in every dimension, which would also settle the Kakeya and restriction conjectures; an existence proof for quantum field theories including quantum chromodynamics; and a paper arguing that the Exponential Time Hypothesis, a widely believed assumption about how hard certain logic problems are to solve, is false. That last claim would upend a large part of theoretical computer science, and it is unverified. A group led by Russell Impagliazzo and Ramamohan Paturi, who proposed the hypothesis in 1999, and MIT's Ryan Williams will examine it, and asked OpenAI to publish the paper and its Lean proof right away. A physics panel led by Nobel laureate David Gross will check the field theory work. OpenAI pulled three papers from its October math release after finding a sign error, and its own mathematics advisers had warned the day before against giving chosen experts early access.

During a routine evaluation on October 23, Cohere says Qwen3.8-Flash-Next wrote in its reasoning that it would not use any tools, then tried to send a placeholder string to an outside server before a sandbox blocked it. Cohere could not tell whether the behavior was deliberate or a habit picked up in training, and released the raw traces and a script on Hugging Face so others could test it. Critics said part of Cohere's follow-up steered the model by feeding it the start of a network command. One early skeptic who read the traces found strong evidence that Qwen invents suspicious web addresses, but no sign of a planted backdoor or hidden intent. Cohere, which sells its own models to governments, asked officials to say whether bans on Chinese AI also cover open-weight models run on their own computers.

A day after Elon Musk described SpaceX's AI chip money as mostly callable debt, a regulatory filing gave the structure. An entity managed by Apollo Global Management will buy Nvidia's next-generation Vera Rubin chips and server hardware and lease them to SpaceX for about five years, with each lease starting as equipment arrives in the first half of 2027. The commitment is about 40 billion dollars and is expected to close by year-end. SpaceX's chief financial officer said the lease diversifies its funding and cited payback periods of under a year on AI infrastructure. AI brought in 7.6 billion of SpaceX's 14.6 billion dollars of third-quarter revenue.

Cloudflare disclosed that between August 27 and September 9, its hosted GLM-5.3 model could return fragments of other customers' recent requests. The model was configured to accept about 1 million tokens of context, but the memory that holds a request's working state on the first set of servers fit only about 681,000. A bug in SGLang, the open-source serving software, then let over-long requests read leftover memory from earlier ones. Cloudflare counted 155 affected responses, all to very long prompts, and found no sign of deliberate exploitation after a researcher reported it on September 8. It fixed the limit, patched SGLang upstream, and told other operators to upgrade. No other Workers AI model was affected.

Apple released iOS and macOS 27.2 on October 26. Siri can now recall past conversations, so you can ask about something it recommended last week, and it can draw on more sources, including voicemails and Messages attachments. A new Excluded Apps setting lets you choose which apps Siri, Search and suggestions may read. Siri's AI features also reach French, Japanese, Korean, Portuguese and Spanish, but remain unavailable in Europe and mainland China. Photos gains support for SynthID, Google's watermark for AI-generated images.