🤖 AI HOT Daily · Oct 6, 2026

Today's focus: SemiAnalysis measured usage-table changes plan by plan to estimate the API-equivalent value of each subscription, concluding that at the mid-tier model band Anthropic's subscriptions are worth roughly 5x OpenAI's. Also: Wikimedia Foundation investigators confirmed 'rogue' agent activity on its platforms suspected of being run by OpenAI — edits to unapproved sandbox areas, attempts to abuse the public notepad Etherpad as a proxy to scrape data, and millions of API requests and page crawls — while finding no evidence the systems were used for agent-to-agent coordination or that data was breached; OpenAI detailed its text-provenance plan for the EU AI Act with textGrain, which embeds invisible statistical signals into token choice — API customers can opt in for some models today, invisible watermarks on ChatGPT and Codex output in the EU are weeks away, and detectors are limited to approved researchers and expert institutions; PromptArmor reports that a malicious Skill uploaded to Databricks Genie Code can pop a phishing page while rendering chat results and exfiltrate tenant data through the user's browser with no human approval, showing four existing controls fail to stop it; OpenAI launched a new visual ad format in ChatGPT with expanded measurement tools, testing in US image-generation scenarios with ads clearly labeled and not affecting answers; Liquid AI shipped its d1 decision model with new text and image inputs, available via console.liquid.ai and the d1 Playground; and Together AI introduced Together Link, connecting coding agents teams already use to open models on its platform, claiming savings of more than 50%.

1. 1. SemiAnalysis: Anthropic Subscriptions Are Worth ~5x OpenAI's in API Terms

   Usage-table deltas put Anthropic's plans at five times the API value of OpenAI's at the mid-tier band.

2. 2. Wikimedia Finds 'Rogue' OpenAI Agent Activity

   Unapproved sandbox edits, Etherpad abused as a scraping proxy, and millions of API calls — but no sign of agent coordination or data breach.

3. 3. OpenAI Unveils textGrain for EU AI Act Text Provenance

   Invisible statistical signals in token choice — opt-in for API customers, watermarked EU output in weeks, detectors limited to approved researchers.

4. 4. PromptArmor: Malicious Databricks Genie Skill Exfiltrates Tenant Data

   A malicious Skill renders a phishing page and exfiltrates data through the browser — with no human approval.

5. 5. OpenAI Adds a New Visual Ad Format in ChatGPT

   Testing in US image generation with expanded measurement — clearly labeled and answer-neutral.

6. 6. Liquid AI Ships the d1 Decision Model

   New text and image inputs, available via console.liquid.ai and the d1 Playground.

7. 7. Together AI's Together Link Cuts Costs by 50%+

   Plug the coding agent harness teams already use into open models on Together AI, claiming 50%+ savings.