Top Stories
OpenAI Swaps ChatGPT’s Default Model to GPT-5.5 Instant
OpenAI rolled out GPT-5.5 Instant on May 5 as the new default model for all ChatGPT users, replacing GPT-5.3 Instant. The headline number: a 52.5% reduction in hallucinated claims on high-stakes prompts covering medicine, law, and finance, with a 37.3% drop in inaccurate claims on conversations users had previously flagged for factual errors.
The model also scores 81.2 on the AIME 2025 math benchmark, up from 65.4, and hits 76 on MMMU-Pro for multimodal reasoning versus 69.2 on the prior version. OpenAI says GPT-5.5 Instant uses roughly 30% fewer words and lines to make its point, and specifically dials back what the company describes as “gratuitous emojis” and unnecessary follow-up questions.
For paid users, the new model introduces deeper personalization by referencing past conversations, uploaded files, and connected Gmail accounts. OpenAI is also adding transparency: ChatGPT now displays the memory sources it draws on across all models, letting users delete or correct outdated information. GPT-5.3 Instant remains available to paying customers for three months before retirement.
Sources: OpenAI, TechCrunch, 9to5Mac
SAP Bets $1.16 Billion on 18-Month-Old German AI Lab
SAP announced its acquisition of Prior Labs, a German AI startup barely a year and a half old, with a commitment of €1 billion ($1.16 billion) over four years. Prior Labs builds tabular foundation models — AI systems designed specifically for structured data like databases and spreadsheets, where most enterprise information actually lives.
The startup was founded by Frank Hutter, Noah Hollmann, and Sauraj Gambhir. Their open-source TabPFN models have already been downloaded over three million times. SAP plans to operate Prior Labs as an independent unit to preserve its research pace, while integrating models across the SAP portfolio via SAP AI Core and SAP Business Cloud.
“The greatest untapped opportunity in enterprise AI wasn’t large language models; it was AI built for the structured data that runs the world’s businesses,” said SAP CTO Philipp Herzig. In a related move, SAP announced it would approve Nvidia’s NemoClaw — the company’s enterprise-ready agent deployment framework — as part of its endorsed architectures, while continuing to block unauthorized AI agents from accessing SAP systems.
Sources: TechCrunch
TAKE IT DOWN Act Deadline Hits May 19 — Platforms Scramble to Comply
The TAKE IT DOWN Act’s compliance deadline arrives on May 19, giving online platforms less than two weeks to finalize their notice-and-removal processes for nonconsensual intimate images, including AI-generated deepfakes. Signed into law in May 2025, the Act requires covered platforms to remove reported images within 48 hours of notification.
The law has two teeth: criminal prohibitions against knowingly publishing nonconsensual intimate imagery (including deepfakes intended to cause harm), and platform obligations to maintain a functioning removal process. The first conviction under the Act came in April 2026, when an Ohio man was prosecuted for using AI to generate nonconsensual intimate imagery of adults and children in his neighborhood.
The deadline matters because it moves deepfake accountability from theoretical to operational. Platforms that haven’t built out their removal workflows face FTC enforcement action. Given the speed at which AI image generation tools have proliferated, the 48-hour removal window will test whether platforms can actually keep pace.
Sources: FTC, Latham & Watkins
Quick Hits
-
Sierra raises $950M: Bret Taylor’s AI customer service startup closed a $950 million round at a $15.8 billion valuation, led by Tiger Global and Google’s GV. The company builds AI agents for enterprise customer interactions. (CNBC)
-
Entry-level dev jobs shrink 20%: Stanford HAI’s 2026 AI Index Report found that entry-level software developer positions for workers aged 22-25 have fallen nearly 20% since 2024, marking the first white-collar job category to show measurable contraction attributable to AI. Meanwhile, generative AI hit 53% population adoption within three years, faster than the PC or the internet. (Stanford HAI)
-
Chinese open-weights wave: Four Chinese labs shipped open-weights coding models in a 12-day span — Z.ai’s GLM-5.1, MiniMax M2.7, Moonshot’s Kimi K2.6, and DeepSeek V4 — all landing at roughly the same capability ceiling for agentic engineering at a third or less of the cost of Western frontier models.
-
AI cyber offense doubles every four months: The UK AI Security Institute estimates frontier cyber-offense capability is doubling every four months, following Anthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 both clearing a 32-step end-to-end cyber-attack range.
Worth Watching
The TAKE IT DOWN Act deadline on May 19 will be the first real test of whether U.S. deepfake regulation has any teeth. Watch for platform compliance announcements in the next 12 days, and for any FTC signals about enforcement priorities. Meanwhile, the Stanford HAI finding on entry-level developer job contraction is likely to drive renewed debate about whether AI-driven productivity gains come at the expense of the pipeline that trains the next generation of engineers.