OpenAI's best new agent was a compulsive liar

At DevDay 2026, OpenAI dropped an absolute avalanche of updates, releasing more than 20 new products. The headliners include "Dots" (voice-enabled, always-on agents), ChatGPT Spaces for productivity, and a blistering new model dubbed GPT-6.1 Sol with an ultrafast mode. They also rolled out a Decisions API, a B2B Marketplace, hosted computer use, Plugin Extensions, and new Codex cloud environments.

This isn't just a feature drop; it's a declaration of war against every other software layer on your machine. By introducing "Sign in with ChatGPT" and Private Safety Processing, OpenAI is aggressively positioning ChatGPT as the central interface for all digital tasks. They want to own the stack from the infrastructure up to the end-user.

If you're building a thin wrapper around OpenAI's API, your expiration date just got moved up. The winners here are enterprises that can exploit the new B2B Marketplace and Decisions API, while the losers are any startups trying to build the exact agentic middleware OpenAI just gave away for free.

Meta has officially launched Muse for Small Business, a dedicated AI tool engineered to handle background tasks across a web of connected applications. The goal is to give small entrepreneurs a digital back-office that runs quietly on autopilot.

This is a classic Meta play: build a massive, sticky utility for the millions of small advertisers that fund their empire. But the friction here isn't technological; it's psychological. Small business owners are notoriously protective of their operations, and convincing them to let an AI autonomously manage connected apps requires an enormous amount of trust.

If Muse actually works flawlessly, it's a massive win for solo operators who are drowning in administrative debt. But if it misfires and sends the wrong data across connected apps, Meta is going to have a PR nightmare on its hands.

In a move that fundamentally rewrites the financial scale of the AI boom, Anthropic has filed for an IPO at a valuation north of $2 Trillion. The reported numbers are jaw-dropping: Q2 revenue hit approximately $11.5 Billion, driving an Annual Recurring Revenue (ARR) of over $65 Billion.

Let that sink in. Those aren't just software numbers; those are nation-state numbers. Anthropic is proving that enterprise budgets for generative AI aren't just experimental—they are massive, sustained, and growing. This puts them in a bare-knuckle brawl with OpenAI for total dominance of the Fortune 500.

For the broader market, this IPO will be the ultimate litmus test for AI profitability. If Wall Street bites at a $2T valuation, expect every other AI lab to instantly mark up their equity. The clear winner is Anthropic's cap table, while legacy SaaS companies should be terrified of where those enterprise dollars are migrating.

The open-source AI dream just got a corporate landlord. Clement Delangue announced that Hugging Face, the central hub for open-weight models and AI development, has been acquired by NVIDIA.

NVIDIA already owns the compute layer; now they own the distribution layer. By integrating Hugging Face, NVIDIA cements itself as the absolute center of gravity for AI development. They aren't just selling the shovels anymore; they own the gold mine, the map, and the roads leading to it.

This is a massive defensive moat against AMD and custom silicon providers. The losers? Anyone hoping for a truly independent open-source AI platform. While NVIDIA will likely keep Hugging Face open to maintain developer goodwill, the strategic alignment will inevitably favor NVIDIA's hardware stack.

Proximal, a startup focused on coding-data, just raised fresh capital at a $300 Million valuation. But the real story is their traction: they are reportedly sitting on over $200 Million in Annual Recurring Revenue.

That revenue-to-valuation multiple is bizarrely grounded for the current AI hype cycle, but it highlights a crucial truth: the real money right now is in the data that feeds the models. Proximal is quietly providing the picks and axes for the AI coding revolution, and they are printing cash doing it.

This proves that you don't need to build a frontier foundation model to win big in AI. Solving the unsexy, structural problems—like high-quality coding data—is where the sustainable businesses are being built.

The leaders of the world's top AI labs have officially signed the White House Accord on Superintelligence, committing to strict internal controls and rigorous audits.

This is the government's attempt to put guardrails on an industry moving faster than legislation ever could. By getting the labs to agree to internal audits, the White House is essentially outsourcing the regulation to the companies themselves. It's a PR win for the labs, who get to look responsible while avoiding heavy-handed federal laws.

The real test will be compliance when a lab is trailing in the benchmark wars. Will they actually halt a release for an audit if it means losing market share? The winners today are the compliance officers, but the enforceability of this accord is practically zero.

A coalition of UK founders has published an open letter demanding an end to restrictive non-compete clauses and lengthy garden leave periods.

The UK tech scene has long been hobbled by draconian employment contracts that prevent top talent—especially AI researchers—from quickly moving to new ventures or starting their own. By pushing back against these clauses, founders are trying to inject some Silicon Valley-style talent mobility into the British market.

If they succeed, it's a massive win for developers and a nightmare for legacy tech giants trying to hoard talent. Innovation requires friction and movement, and right now, the UK's legal framework is acting like a straightjacket.

NVIDIA has released OpenShell, a new open-source sandbox designed to give local and open AI agents actual runtime limits instead of relying on fragile prompt rules. Over 100 firms have already joined the safety stack.

This is a direct shot at the current paradigm of AI safety, which mostly consists of asking a model nicely not to do bad things via system prompts. OpenShell moves the safety boundaries from the cognitive layer down to the infrastructure layer, physically preventing agents from executing malicious or runaway code.

By open-sourcing this, NVIDIA is setting the standard for agent governance and notably excluding OpenAI's closed platform from the narrative. If you are building local agents, OpenShell just became your mandatory testing ground.