Today’s stories trace a familiar split in AI: products keep shipping at a blistering pace while the harder questions about benchmarks, money, and trust get answered more slowly, or not at all.
The OpenAI Reshuffle: Products Launch as the Legal Bill Comes Due
OpenAI rearranged its own product line and leadership bench today, while a New York court motion raised uncomfortable questions about what the company already knew about its own chat logs.
- OpenAI Kills Atlas Browser, Folds Agentic Browsing Into ChatGPT. Nine months after it shipped as a standalone bet, OpenAI is shutting the browser down and rebuilding its agentic features as a ChatGPT desktop app and a Chrome extension instead.
- OpenAI Launches ChatGPT Work to Compete With Claude Cowork. The new GPT-5.6 powered workspace turns scattered team files into finished documents, planting OpenAI’s flag on collaborative work territory Anthropic staked out first.
- OpenAI’s Applications Chief Fidji Simo Steps Back After Health Leave. Simo is moving to a part time advisory role as a worsening neuroimmune condition forces OpenAI to split her responsibilities among Greg Brockman, Sarah Friar, and Jason Kwon.
- News plaintiffs accuse OpenAI of hiding chat log searches for two years. A new sanctions motion in the New York Times copyright case claims OpenAI could search ChatGPT logs long before it admitted as much, a discovery dispute that could weaken its fair use argument.
Winning the Test, Missing the Point: What GPT-5.6 Actually Proves
OpenAI’s newest model family shows how uneven progress looks up close, strong on aggregate scores and cost, still shaky on the benchmarks built to catch genuine reasoning failures.
- OpenAI ships GPT-5.6 in three tiers, Sol, Terra, Luna, after a gate. The flagship Sol tier lands close to Claude Fable 5 on blended benchmark scores at roughly a third of the price, but a wide gap opens on SWE-Bench Pro, the test built to catch real coding failures.
- GPT-5.6 Sol Wins an ARC-AGI-3 Game, Yet Still Fails Most of the Test. Sol becomes the first model to beat a public ARC-AGI-3 game, a genuine milestone that still leaves it clearing only 7.78 percent of the overall suite, with every rival scoring lower.
Money, Chips, and Trust: The Infrastructure Behind the Headlines
Away from model launches, the deals and governance moves that will decide who can afford to compete kept stacking up, from a training data startup doubling its price tag to Anthropic recruiting outside credibility ahead of a possible IPO.
- Mercor Reportedly in Talks to Double Valuation to $20 Billion. The AI training data startup already has a term sheet at twice its October valuation as revenue climbs and it folds in an acquisition of Deeptune.
- Meta’s Custom AI Chips Head Into Production This September. Manufacturing on Meta’s newest MTIA chip generation starts in September with Broadcom and TSMC, a direct hedge against the company’s own Nvidia spending.
- Anthropic adds ex-Fed Chair Bernanke to its governance trust. The Nobel laureate joins as the fourth member of a trust that holds no equity but appoints Anthropic’s board, a credibility move that lands as the company edges toward a possible IPO.
- Musk’s Anthropic Praise Reads Like Gratitude, Not a Truce. Elon Musk called Anthropic the AI leader and promised never to cut its SpaceX compute, a compliment that lines up neatly with a $40 billion contract rather than any real thaw between the two companies.
Under the Hood: The Research Fixing What Models Can’t Yet Do
Two infrastructure stories show the unglamorous work behind flashier launches, patching the operating system billions of people use and rethinking how video models trade speed for quality.
- Microsoft Turns AI Loose on Windows Bug Hunting, Widens Patch Pipeline. A multi model scanner called MDASH is finding Windows flaws faster than human reviewers could, though the resulting surge in confirmed bugs means IT teams now face more patches to triage each month.
- Nvidia’s Flex-Forcing Merges Two Video Generation Modes Into One Model. A new chunking method lets a single video diffusion model trade speed for quality at inference time, replacing the separate bidirectional and streaming checkpoints developers used to choose between.
Today’s Quick Hits
- Meta Opens Its Model API as Muse Spark Gets a 1.1 Upgrade. Developers can now build on the upgraded Muse Spark 1.1 through a public preview of Meta’s Model API, closing a gap the company had left open with no announced launch date.
- Prisma’s Type Safety Becomes the Guardrail for AI-Written Database Code. As coding agents generate more database queries and migrations, Prisma’s compile time checks are catching mistakes before they ever reach production.
- Z.ai Ditches GRPO’s Grouped Sampling to Fix Async RL Instability. Z.ai’s research team replaced grouped sampling with one rollout per prompt, giving up some throughput to gain training stability in agentic reinforcement learning.
- Anthropic Asks the Public for AI’s Hardest Questions. Anthropic will publish how it responds to public concerns about AI safety, though it has not said who will verify that the company actually follows through.