AI companies shipped more this week than most industries manage in a year — new models, a reworked ChatGPT, a record chip IPO — while courts and legislators scrambled to keep pace with what was already deployed months ago. The question worth carrying through today's briefing: when capability moves this fast, who actually gets to check the work?
OpenAI turns ChatGPT into a workplace platform with GPT-5.6
OpenAI broadly released its GPT-5.6 family on Thursday, two weeks after the Trump administration asked it to limit the rollout to government-approved users over cybersecurity concerns, a restriction the company says was lifted without objection.
The family comes in three tiers:
- Sol, the flagship, which CEO Sam Altman told CNBC is 54% more token efficient on agentic coding tasks than prior models
- Terra, a balanced everyday option
- Luna, the budget-friendly variant, priced at $1 input / $6 output per million tokens
Alongside the models, OpenAI launched ChatGPT Work, an agent that folds its Codex coding tool into the ChatGPT desktop app and can "stay with a project for hours," drafting documents, spreadsheets and presentations across connected tools like Slack and Google Drive. GPT-5.6 has also become the preferred model inside Microsoft 365 Copilot, touching Word, Excel, PowerPoint and Cowork. In the same wave of announcements, OpenAI confirmed it is shutting down its Atlas browser less than a year after launch, folding its features into ChatGPT instead.
The New York Times accuses OpenAI of hiding evidence
The copyright fight between news publishers and OpenAI escalated sharply this week. The New York Times and Daily News filed a motion asking a federal judge to sanction OpenAI for allegedly concealing its ability to search ChatGPT logs for copyrighted content.
The claims are specific:
- OpenAI privacy engineer Vincent Monaco reportedly testified in an April deposition that the company had already searched a de-identified sample of 78 million ChatGPT logs, and had built a "Bloom filter" under a project called "Project Giraffe" to detect regurgitation of copyrighted text
- Plaintiffs originally sought 120 million logs but were negotiated down to 20 million, which OpenAI then reportedly redacted 19 billion times, rendering the sample "unusable" in the court's own words
- Plaintiffs' counsel Ian Crosby said OpenAI "lied to The Times, The Daily News Plaintiffs, the public, and the court" for two years
OpenAI's spokesperson rejected the allegations, calling the motion an attempt to invade the privacy of ChatGPT users unrelated to the case.
Meta undercuts rivals on price with Muse Spark 1.1
Meta launched Muse Spark 1.1, its first AI model it will charge for, with CEO Mark Zuckerberg calling the pricing "very aggressive" and telling Bloomberg it beats Google's Gemini across several categories. Meta will charge developers $1.25 per million input tokens and $4.25 per million output tokens, undercutting Anthropic's Opus and coming in just above OpenAI's new Luna model. The launch was significant enough that Zuckerberg posted on X for the first time in three years to promote it, alongside a promise of "more to come soon."
Illinois signs AI safety law as Washington's own process stays opaque
Illinois governor JB Pritzker signed the Artificial Intelligence Safety Measures Act, modelled on California's SB-53 and New York's Responsible AI Safety and Education Act. The law requires developers of the largest models — those generating over $500 million in annual revenue — to publish frameworks for identifying "catastrophic risk" and report qualifying incidents within 72 hours, or 24 hours for imminent threats. Lawmakers estimate the three states together represent roughly 40% of the US AI market, effectively creating a de facto national standard in the absence of federal rules.
That absence is the story at the national level. TechCrunch's reporting on how the government actually vetted GPT-5.6 before release found that nobody outside a small circle can say. A senior research analyst at Georgetown's Center for Security and Emerging Technology said she has no visibility into the process, and former White House AI adviser Sriram Krishnan has said plainly: "There will not be an FDA for AI." Six cabinet agencies are due to define a final evaluation process by early August.
SK Hynix's $26.5bn listing underscores the AI infrastructure boom — and its risks
South Korea's SK Hynix priced its Nasdaq listing to raise $26.5bn, oversubscribed more than seven times, beating out Saudi Aramco's 2019 debut. It's the latest sign of capital pouring into AI infrastructure: Amazon is separately planning to raise at least $25 billion in bonds to fund AI investment, and Meta confirmed it will put its own "Iris" AI chip into production in September as it targets 14 gigawatts of compute next year. Not everyone is convinced the payoff is guaranteed: Taiwan's central bank chief has warned of AI bubble risk, and chip stocks have dropped sharply from June highs as investors question whether hyperscaler spending will translate into revenue.
Fidji Simo steps back from OpenAI as IPO planning continues
OpenAI's second-in-command, Fidji Simo, is stepping down from her full-time role as chief executive of AGI deployment, transitioning to a part-time advisory position after a medical leave for a chronic neuroimmune condition proved longer than expected. Simo had consolidated much of OpenAI's business and product operations after joining from Instacart in 2025, and her departure leaves a leadership vacuum just as the company reportedly eyes a $1 trillion IPO valuation.
The Hexalink view
Look at the shape of this week: OpenAI shipped a model family and a workplace agent within hours of a government "greenlight" nobody can fully explain, Meta undercut prices before anyone had properly tested the last release, and a state legislature had to write its own catastrophic-risk rules because Washington hasn't. Capability and commercial deployment are moving on a weekly cycle; verification, whether legal, regulatory or journalistic, is still moving on a quarterly or annual one. The NYT sanctions motion matters beyond one lawsuit because it's the clearest evidence yet that "we can't technically check" is not always true — it can also be a negotiating position.
For technology leaders, the practical takeaway is to stop assuming government approval or a vendor's safety card is a substitute for your own due diligence — build in independent verification of any frontier model's data provenance and safety claims before wiring it into critical workflows. We'd also flag that with Illinois joining California and New York, a 40%-of-market state patchwork now functions as a national AI standard whether or not Congress acts, so compliance planning should treat it as binding, not optional.
Come back tomorrow for the next briefing, or catch the five-minute audio version on the AI Storm Daily podcast.