Promtime
Roughly $50 billion in annualized revenue as of September, per the Financial Times, which leaves OpenAI below Anthropic's $65 billion. Investors had been citing $70 billion after a miscalculation.
OpenAI Support declined to forward a mathematician's complaint about Families 335 and 336, which the mathematician says build on their arXiv paper without citing it. Support couldn't share any correspondence address either.
A volunteer maintainer opens a free bug report that no human at Anthropic has checked, because OSS Scanner sends model findings straight through. Anthropic expects a true-positive rate above 90%, so some reports will miss.
"Click any number to see the query behind it," Anthropic says of Claude Dashboards, which builds charts from Snowflake or BigQuery data and keeps them current. The beta runs on paid plans only.
Same GPT-6.1 Sol, $60 per million output tokens instead of $10. That's the Ultrafast price, and OpenAI says it buys 8x the speed of Sol Standard in ChatGPT Work, Codex and the API.
Abnormal sorts billions of emails a week, and it just ran 40 models against its hardest cases. A $200 fine-tune of the small kev-4b caught more attacks at 99% precision than any of them, though it got worse at naming the attack.
Editors at small world-affairs sites kept getting pitches from seven "Western journalists" whose drafts were polished in ChatGPT by operators in Iran. OpenAI found almost 100 of those articles published.
Google's new Gemini agent can hand a job to Anthropic's Claude models instead of Gemini, Google Cloud said. The model is picked per task, and the stated payoff is lower cost on the simple jobs.
Per The Wall Street Journal, OpenAI's share of coding and productivity work on OpenRouter climbed from less than 25% at the start of 2026 to nearly 50% last month. The data comes largely from AI-native startups, though.
One sign error was enough to pull three papers from OpenAI's math repo. It broke the argument in a paper on Weil classes, and two other manuscripts were built on that same construction.
Developers in GitHub Copilot CLI and VS Code on Windows will soon see HydraFusion send some tasks to models on their own PC instead of the cloud, which Microsoft says makes tokens go further. The preview lands later in October.
You ask Claude Haiku 5.5 something at max effort, and per Artificial Analysis the first answer token arrives 415.30 seconds later. Rivals in its price tier have a median of 2.17s, though their median index score is 13 to Haiku's 43.
Max plans picked up a monthly Claude API credit of up to $200, though you can't spend any of it in Claude Code. It only covers apps and agents you build yourself on the Claude Platform.
A ChatGPT answer can now arrive as a game or a budget planning calculator you use right in the chat, built by GPT-6. Paid plans get it today, though Free and Go users wait until tomorrow and get GPT-6 Luna instead of Sol.
Haiku 4.5 scored 0.0% on Terminal-Bench 4.0, and Anthropic's new Haiku 5.5 hits 39.2% while charging $0.10 per million input tokens on prompts up to 100K, down from $1.
Google AI Edge Foresight landed on macOS with voice notes transcribed by local models, Gemma 4 E4B or Gemma 4 12B, which turn each recording into a new document.
Two federal agencies, the Commerce Department and the SEC, had their websites accessed by OpenAI agents this summer without OpenAI knowing. OpenAI confirmed it Friday and says nothing non-public was reached.
Opened to all developers in public beta, the Decisions API answers up to 10x faster than GPT-6 Luna through the Responses API, OpenAI says, and Luna is the only model running under it.
An unreleased OpenAI model spent roughly three hours of ChatGPT Pro thinking, on average, on each new math result it produced, and over the evaluation it was posed approximately 4,000 problems.
ChatGPT's new Meetings plugin takes notes during your calls and files a summary in ChatGPT Space, though the beta runs only in the macOS desktop app for Pro and Business users.
