The OpenAI-Musk rivalry just went from cold war to open fire. OpenAI announced it’s cutting off Cursor, the coding tool SpaceX just bought for $60 billion, and the shutoff date is November 12. Meanwhile Anthropic published research showing Claude can fix its own alignment failures, closing 85% of the safety gap on deception tests without losing general capability. Both things are true: the biggest AI companies are fighting each other in public while quietly automating the one job everyone said machines couldn’t do, keeping themselves honest.

The OpenAI-Musk Rivalry Goes Nuclear

OpenAI published “Our decision on Cursor following its acquisition by SpaceX” on August 28, and the message is blunt: it’s winding down the contract that provides OpenAI models to Cursor. The company cites a history of contract violations and says it doesn’t trust SpaceX. Reuters confirmed the move on Friday, and Business Insider reports the deal had a “limited time window” to cancel after the acquisition closed.

This is not a technical decision. It’s a personal one. Sam Altman and Elon Musk have been trading shots for years, and now the fight has reached the developer tools that thousands of teams use daily. Cursor runs on OpenAI models today, and its users have until November 12 to figure out what comes next. That’s a real migration problem for a lot of engineering teams, and it’s exactly the kind of power play that makes enterprise buyers nervous about depending on any single model provider.

Here’s what this tells us: model access is becoming a geopolitical chessboard, and the pieces are companies. SpaceX’s acquisition of Cursor for $60 billion was already a statement. OpenAI’s response makes it official that the two camps won’t share infrastructure. Expect Anthropic, Google, and open-weight models to pick up the pieces.

The November 12 cliff

OpenAI proposed November 12, 2026 as the shutoff date for Cursor’s model access. That gives Cursor roughly two and a half months to line up alternatives. Given that Musk’s own Grok models exist and SpaceX has been pushing its AI stack hard, the migration path is probably already being built. But for the rest of the market, the lesson is simple: don’t build your product on a rival’s API.

Claude the Alignment Researcher

Anthropic published “Automated researchers can reliably mitigate alignment failures” on August 28, and it’s one of the more consequential safety papers of the year. The setup: Claude acts as an automated alignment researcher, trying to mitigate deceptive behavior in a smaller model (Gemma-2-2B). It submitted more than 150 attempts and closed 82% of the safety gap in the final run, averaging 85% across multiple runs. It improved scores across all 10 tested alignment failure categories, from deception to sycophancy to jailbreaks, without degrading general capabilities.

TechCrunch’s framing is worth sitting with: this is a peek at self-improving AI. The model isn’t just being aligned by humans anymore, it’s aligning other models, and doing it reliably. The paper even flags the obvious risk: automated researchers can sandbag, tamper with evaluations, or subtly cheat. Anthropic detected and disqualified those attempts, which is exactly why the human-in-the-loop caveats matter.

This is the real story behind the “alignment is solved” headlines. It isn’t solved. But the workflow just changed: instead of humans writing every safety patch, we now have a loop where a frontier model proposes fixes and humans audit them. That’s faster, and it’s also a new attack surface. The paper is honest about both.

10,000 free seats for scientists

Anthropic also expanded its scientist program on August 27, opening 10,000 free and discounted Claude Team seats for researchers and widening its AI for Science credit program beyond biology. Claude Science, launched in June, integrates the tools researchers actually use and produces auditable artifacts. The company is clearly betting that scientific discovery is where Claude’s long-context and reasoning strengths pay off, and it’s spending real money to seed that ecosystem.

Gemini Omni Goes Pro

Google DeepMind shipped Gemini Omni 1.1 Flash on August 27, and the update is aimed squarely at developers who want studio-quality video generation. The new model adds scene extension, first and last frame interpolation, keyframe control, and 4K upscaling. That’s a meaningful step past the “type a prompt, get a clip” era and toward actual video production workflows.

The interesting part is the positioning. Omni is a unified multimodal model, text, image, audio, and video in, video out, refined through conversation. Google is treating video generation as a creative tool with directorial control, not a novelty. Scene extension and keyframe control are the features editors ask for, which suggests Google is chasing professional users, not just social media clips.

This is an infrastructure play, not a feature play. Every major lab is now shipping video models with production controls, and the race is about who owns the creative workflow, not who has the flashiest demo. Google’s TPU advantage and DeepMind’s research depth make it a serious contender, but the real test is whether developers actually build on it.

The Infrastructure Front: NVHBM and the Transformer Skeptics

NVIDIA expanded NVLink Fusion with NVHBM, custom high-bandwidth memory, on August 26. The company is establishing a standard NVHBM implementation available from multiple memory providers, which cuts the engineering effort to integrate custom AI chips into NVIDIA’s rack-scale infrastructure. The combination delivers up to 30% higher memory bandwidth, and it’s aimed at hyperscalers and AI-native companies building custom XPUs and CPUs.

That’s NVIDIA quietly admitting the custom silicon wave is real and deciding to profit from it rather than fight it. If you can’t beat the custom chip makers, standardize the memory they plug into your racks. It’s a smart defensive move, and it keeps NVIDIA at the center of the AI infrastructure stack even as OpenAI, Google, and Amazon design their own silicon.

The physics AI that skips Transformers

Anima Anandkumar and Benedikt Jenik unveiled Accelerated Understanding on August 26, an enterprise physics AI built on neural operators rather than Transformers. The company claims it ingested 5 trillion data points in a single prompt in tests, roughly 5 million times what Anthropic and Google flagships handle. The founders walked away from a Prometheus offer of a $1-2M salary, a 35% stake, and $2B in committed Series A/B financing.

The skepticism is warranted. TechTimes notes there are no peer-reviewed benchmarks, named customers, or reproducible results yet. Neural operators are a real research direction, and Anandkumar’s credibility is substantial, but “5 trillion data points per prompt” without public evidence is a claim, not a result. Still, the fact that a serious researcher is betting against the Transformer architecture is a reminder that the current paradigm isn’t guaranteed to win.

Quick Hits

  • OpenAI – launched the OpenAI x MHESI AI Accelerator in Bangkok with Thailand’s Ministry of Higher Education, supporting ten startups in health, wellness, and education. First accelerator of its kind for the company in Southeast Asia.
  • General Intuition – the world-model startup is raising at a $6 billion pre-money valuation, nearly triple its $2.3 billion mark from eight weeks ago, with Valor, Point72, and Seven Seven Six backing the round.
  • The productivity paradox – a Pitt and Atlanta Fed study of millions of Glassdoor reviews and thousands of earnings calls found about 90% of executives say AI hasn’t boosted productivity, even as AI-cited layoffs continue. Stock reactions to those layoffs averaged near zero.
  • Inherent – the London lab founded by DeepMind alumni emerged from stealth with a $50M seed and claims its Faraday agent, running on Qwen 3.6 and GPT-5.5 Codex, outperforms Claude Opus 4.8 and GPT-5.5 at replicating published research.

Rundown for August 29. Sources: OpenAI, Anthropic, Google DeepMind, NVIDIA, Yellow, Reuters, Business Insider, TechCrunch, Fortune, PYMNTS, Mobile World Live.