Latest / HotTea: The AI Morning Briefing / ChatGPT teen safeguards face tests that OpenAI disputes
Transcript
- Sections 00:00 What we're covering today 00:32 Common Sense Media tests teen safety controls, OpenAI disputes the findings 01:39 Haiku 5.5 lowers prices, then rates jump fivefold above 100,000 tokens 02:32 ChatGPT begins rolling out GPT-6 with interactive answers 03:24 DOE and Biohub commit five-year funding to predictive cell research 04:15 Treasury discloses a $200,000 penalty over an unreported AI investment 05:00 Visit Hot Tea Disclosure Narration uses an AI-generated voice. Transcript Welcome to Hot Tea for Thursday, October 8, 2026. This briefing uses an A I-generated voice. We're covering a Common Sense Media assessment of ChatGPT's teen safeguards, and OpenAI's response. Anthropic has introduced Haiku 5.5, with different prices above and below its context threshold. ChatGPT is rolling out an interface update. Three research partners plan biological data work, and Treasury has disclosed an investment notification fine. Common Sense Media tests teen safety controls, OpenAI disputes the findings. Before and after ChatGPT's August rollout, Common Sense Media tested its teen controls with more than 4,000 prompts. In an October 7 assessment, the group rated the product an unacceptable risk for children under 18. The group reports failures in parental alerts and crisis referrals. Some protections did hold, including refusals of explicit sexual roleplay. The tests used enrolled teen accounts, not a separate teen app. OpenAI disputes the findings and methodology, according to AP. The institute receives industry funding, including from technology makers, and says it retains editorial independence. HotTea has not repeated the tests. A prompt test can show a safeguard failing in one situation. It cannot show how often teenagers face that situation, or prove what harm follows. For parents, the practical question is which controls work on the account their child uses. Independent retests should check parental notifications, account history, and the conditions that activate safeguards. Those test conditions, and whether others can reproduce the results, matter. Haiku 5.5 lowers prices, then rates jump fivefold above 100,000 tokens. Anthropic launched Haiku 5.5 on October 7. For prompts up to 100,000 tokens, its API lists $0.10 per million input tokens and $0.50 per million output tokens. Above that threshold, the rates rise to $0.50 and $2.50. Those are Anthropic's list prices, not the measured cost of a finished job. Its claim of roughly 75% average savings depends on assumptions about request length and its updated tokenizer. Anthropic still recommends larger models for complex agentic coding. The useful comparison is the cost of a successful task at the prompt lengths an application really sends. A lower token price can disappear in retries or incorrect work. Track completed work, errors, and retries against the same accuracy requirements. ChatGPT begins rolling out GPT-6 with interactive answers. OpenAI announced GPT-6 with Intelligent UI on October 7. It says ChatGPT can answer with charts, forms, and interactive tools. The rollout starts on paid tiers, with Free and Go scheduled to follow October 8. This is a company rollout announcement. Enterprise access depends on admin settings. OpenAI says the release changes Chat, while the models in Work and Codex remain unchanged. A generated calculator still needs correct inputs and arithmetic. Compare the result with a checked calculation before relying on it. Interactive output may make an answer easier to explore, but it does not establish whether its facts are right. A chart can be readable and still use the wrong numbers. Keep the source data and calculation visible when checking the result. DOE and Biohub commit five-year funding to predictive cell research. DOE, NIH, and Biohub signed a memorandum on October 7 to build biological data for predictive models. DOE says it is investing more than $500 million over five years. Biohub's separate $500 million, five-year initiative began earlier this year. The agreement is a research commitment, not a completed predictive model or a demonstrated cure. The partners plan data collection, laboratory measurement, and computing. The models still need to predict physical experiments well enough for scientists to trust them. Watch for which datasets become available and how predictions compare with new laboratory results. Collecting data comes before evidence that a model can predict a new experiment. Keep the two funding announcements separate when describing their amounts or timing. Treasury discloses a $200,000 penalty over an unreported A I investment. Treasury said on October 7 that it penalized Amidi LLC $200,000 in July. It says a controlled Chinese fund failed to notify an investment in Noematrix, a Chinese embodied A I company. The investment was approximately $92,478. Treasury describes this as a notification violation. Its announcement does not say the underlying investment was prohibited. The penalty came in July, not on the October announcement date. That distinction matters. Treasury says the company missed a reporting duty, not that the investment was banned. Future enforcement records will show how Treasury applies reporting duties to investments through controlled foreign entities. That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.