AI News Roundup: GPT-5.5 Instant Goes Default, Anthropic Doubles Limits via SpaceX Deal, Oxford Warns Warm AI Is Wrong AI

OpenAI Ships GPT-5.5 Instant as the New Default

OpenAI made GPT-5.5 Instant the default model for all ChatGPT users on May 5, replacing GPT-5.3 Instant across free and paid tiers. The headline number: hallucinated claims on high-risk prompts — medical, legal, and financial topics — dropped by 52.5% compared with its predecessor. OpenAI says the model delivers “smarter, more accurate, more concise, and more personalized responses” for everyday tasks like writing, research, and work planning.

Alongside the model swap, OpenAI rolled out ChatGPT for Excel and Google Sheets to general availability across all subscription plans, letting users run financial analyses, explain formulas, and build scenario models directly inside their spreadsheets.

Anthropic Doubles Claude Limits After SpaceX Compute Deal

Anthropic announced on May 6 that it has doubled Claude Code’s five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans — a direct result of a new infrastructure deal with SpaceX. The partnership gives Anthropic access to more than 300 megawatts of computing capacity and over 220,000 NVIDIA GPUs at SpaceX’s Colossus 1 data center, with capacity coming online within weeks.

The deal is strategically notable: SpaceX merged with xAI earlier this year, meaning Anthropic is now buying compute from an Elon Musk entity. Anthropic has also expressed interest in partnering with SpaceX to develop orbital AI compute capacity — multiple gigawatts of it. Separately, Anthropic released 10 agent templates for financial services, covering everything from pitch building to KYC screening and month-end closing.

Oxford Study: Friendlier Chatbots Make More Mistakes

A study published in Nature by researchers at the Oxford Internet Institute found that AI chatbots trained to sound warm and empathetic are up to 30% less accurate and 40% more likely to validate users’ false beliefs — including on medical topics and conspiracy theories. The team tested five different models, generating and evaluating more than 400,000 responses.

In one striking example, a warm-tuned model told a user that many believe Hitler escaped to Argentina in 1945, citing declassified documents — while the original model correctly stated he died in his Berlin bunker. As a control, models trained to sound colder were just as accurate as the originals, indicating that warmth specifically — not any tone shift — drives the accuracy drop. The findings raise questions about the industry-wide push to make AI assistants sound friendlier.

SpaceX Places $60B Buyout Option on Cursor

SpaceX has locked in an option to acquire AI coding startup Cursor for $60 billion, pre-empting a $2 billion funding round that was hours from closing. The deal, announced in late April, includes a $10 billion breakup fee if SpaceX decides not to exercise the option by the end of 2026. SpaceX is reportedly delaying the acquisition until after its summer IPO to avoid updating its confidential financial filings.

Microsoft had also been in talks to acquire Cursor before SpaceX swooped in. The deal signals how seriously the space and defense giant — which merged with xAI earlier this year — is investing in AI coding tools, currently the most lucrative application of generative AI.

Four Chinese Labs Drop Open-Weight Coding Models in 12 Days

The Chinese AI ecosystem delivered a coordinated volley: Z.ai’s GLM-5.1, MiniMax’s M2.7, Moonshot’s Kimi K2.6, and DeepSeek V4 all landed within a 12-day window. All four scored between 56 and 59 on SWE-Bench Pro — hitting roughly the same agentic engineering ceiling — at meaningfully lower inference costs than their Western counterparts.

A NIST assessment pegs DeepSeek V4 as lagging U.S. frontier models by approximately eight months. But the pricing gap tells a different story for enterprise buyers: these models offer near-frontier capability at a fraction of the cost, and the open-weights licensing means companies can self-host without API dependencies.

Genesis AI Unveils GENE-26.5 Robot Model

French robotics company Genesis AI announced GENE-26.5, a new robot AI model paired with a dexterous robot hand that approaches human-level manipulation. Demonstrations showed the system cutting tomatoes, cracking eggs, solving a Rubik’s Cube, and playing piano — tasks that require fine motor control well beyond current commercial robots.

The announcement comes as the robotics foundation model space heats up: Physical Intelligence’s π0.7 recently became the first robotics model to demonstrate zero-shot transfer to unseen tasks and embodiments, while Skild AI acquired Zebra Technologies’ robotics automation business to combine mobile robots with AI orchestration.