AI & design briefing for Friday, October 9, 2026

tldr

• Anthropic launches Claude Dashboards and Motion in beta. Google announces Gemini Agent for work inside Workspace. OpenAI rolls out Ultrafast mode for GPT-6.1 Sol. ChatGPT Intelligent UI builds charts, forms and calculators inline. OpenAI tells investors annualized revenue is near $50B. Anthropic tightens usage policy, banning sustained cruelty to Claude. Anthropic launches Cyber Mission for critical infrastructure defense. OpenAI says it ended two influence operations using its models.
• Epoch report finds models cannot yet automate frontier research work. Speculative decoding speeds generation only at low batch sizes. ATLAS benchmark finds search agents miss many golden answers. GitHub accessibility agent reviewed 3,535 pull requests. Mathematicians group asks members to keep commercial AI out of research. Slack redesigned UX after Slackbot appeared to leak private data. Survey of 150,000 creatives examines how AI changes their work.
• Google releases Nano Banana 2.1 for image generation and editing. Gemini 4 Argon debuts at number one on APEX-Agents. OpenAI opens Decisions API beta for faster model routing. Anthropic adds browser and computer control to Claude SDKs. Midjourney tests Thinking Mode for image generation. Anthropic OSS Scanner finds vulnerabilities in enrolled open-source projects. Anthropic opens Claude to vetted security pros for offensive testing. Google launches AI Edge Foresight note-taking app for meetings. Odyssey releases Odyssey-3, its most powerful world model. Developer rewrites TypeScript compiler in Rust with $20k of Claude.

Top Stories

• Anthropic launches Claude Dashboards and Motion in beta. Claude Dashboards pull live charts from company data like Salesforce; Claude Motion turns reports into editable animations. Docs, Slides and Design leave beta, including on Free. ^[1]
• Google announces Gemini Agent for work inside Workspace. Gemini Agent handles knowledge work, media creation and code across Workspace apps, with model routing, token caps and per-agent sandbox permissions. Google says business customers come first. ^[2]
• OpenAI rolls out Ultrafast mode for GPT-6.1 Sol. Ultrafast runs in the API, Codex and ChatGPT Work at up to 8x faster than Sol Standard, priced at $12 per million input tokens and $60 output. ^[3]
• ChatGPT Intelligent UI builds charts, forms and calculators inline. OpenAI's Intelligent UI combines text with interactive diagrams, charts and task-specific tools in one conversation, rolling out with GPT-6 on paid and free plans. ^{[4] [5]}
• OpenAI tells investors annualized revenue is near $50B. $20B below a figure circulated last week that counted partner gross revenue. The two labs count annualized revenue differently. ^[6]
• Anthropic tightens usage policy, banning sustained cruelty to Claude. The update adds a dedicated section on deceptive campaigns and bans using Claude for weapons or drone arming, plus human review for high-risk uses. ^[7]
• Anthropic launches Cyber Mission for critical infrastructure defense. The program gives critical-infrastructure and open-source partners access to Anthropic's strongest models, on-site engineers and threat research. ^[8]
• OpenAI says it ended two influence operations using its models. One Russian operation recruited a Latin America think tank with fake documents; an Iranian one created seven journalist personas to pitch outlets. ^[9]

Deeper Cuts

• Epoch report finds models cannot yet automate frontier research work. Frontier models handled well-defined tasks but failed open-ended judgment calls, and open-weight models lagged further behind. ^[10]
• Speculative decoding speeds generation only at low batch sizes. A small draft model proposes tokens that a big model verifies in one forward pass, but at high throughput it can waste compute. ^[11]
• ATLAS benchmark finds search agents miss many golden answers. The new benchmark grades agents on real tasks that require web searches across domains, using non-memorized questions. ^[12]
• GitHub accessibility agent reviewed 3,535 pull requests. The experimental agent resolved 68% of them, guiding engineers in Copilot CLI and VS Code while escalating drag-and-drop and rich text editors to specialists. ^[13]
• Mathematicians group asks members to keep commercial AI out of research. The Association for Human Mathematics is asking members to keep commercial AI out of research and publication; a stricter caucus rejects model use entirely. ^[14]
• Slack redesigned UX after Slackbot appeared to leak private data. The hallucinated leak pushed Slack's content designers to fetch correct data and explain why LLMs hallucinate; trust is treated as the product. ^[15]
• Survey of 150,000 creatives examines how AI changes their work. Contra Labs surveyed more than 150,000 working creatives for its Creative Intelligence Report on how AI is changing creative work. ^[16]

Tools and Launches

• Google releases Nano Banana 2.1 for image generation and editing. The model generates and edits at 1K, 2K and 4K with better text rendering and character consistency. ^[17]
• Gemini 4 Argon debuts at number one on APEX-Agents. Google's new model debuts at number one on the APEX-Agents leaderboard. ^[18]
• OpenAI opens Decisions API beta for faster model routing. The beta routes model choices about 10x faster than GPT-6 Luna. ^[19]
• Anthropic adds browser and computer control to Claude SDKs. Claude's Python and TypeScript SDKs now include built-in browser and computer use tools. ^[20]
• Midjourney tests Thinking Mode for image generation. Midjourney is testing the mode on its Alpha website. ^[21]
• Anthropic OSS Scanner finds vulnerabilities in enrolled open-source projects. The opt-in scanner suggests fixes for enrolled open-source projects. ^[22]
• Anthropic opens Claude to vetted security pros for offensive testing. Verified security professionals can use Claude for offensive penetration testing under the program. ^[23]
• Google launches AI Edge Foresight note-taking app for meetings. The app takes notes in remote and in-person meetings. ^[24]
• Odyssey releases Odyssey-3, its most powerful world model. The firm calls Odyssey-3 its most powerful foundational world model yet. ^[25]
• Developer rewrites TypeScript compiler in Rust with $20k of Claude. The rewrite took two weeks and roughly $20,000 of Claude usage. ^[26]
• Devin's Security Swarm scans large codebases for vulnerabilities. Security Swarm uses an Agentic MapReduce method to find RCE, SQL injection and SSRF. ^[27]
• Siteprint turns any website's design into a prompt. A Mac Safari extension measures a site's colors, type, spacing and layout for coding agents. ^[28]

Prompt of the day