GPT-5.6 launches messy, Apple sues OpenAI, Meta's big week

OpenAI ships GPT-5.6 Sol with rocky rollout, Apple sues OpenAI over hardware secrets, Meta enters coding AI and pulls Instagram deepfake feature.

The big picture

OpenAI had the biggest week, and not all of it was good: a major model launch with self-admitted UX failures, an executive departure, a lawsuit from Apple, and a deepening copyright fight with the NYT all landed within 24 hours. Meanwhile Meta shipped a coding model and immediately had to walk back an Instagram feature that let anyone generate AI images from public accounts without permission. The through-line is that AI companies are moving faster than their own governance can keep up with.

OpenAI’s GPT-5.6 launch: impressive model, chaotic rollout

After weeks in a government-gated “limited preview,” GPT-5.6 is now publicly available following clearance from the Trump administration. The launch bundled three models — Sol (the flagship), Terra, and Luna — plus a new product called ChatGPT Work that merges the Codex agent capabilities with a consumer-friendly interface, so non-technical users can kick off long-running autonomous workflows without writing a line of code. Sam Altman called it the best model OpenAI has ever shipped. The Verge

The Sol model ships with five reasoning levels — Light up through xhigh — plus “Max” and “Ultra” modes that spin up multiple sub-agents working in parallel. OpenAI’s own Vaibhav Srivastav recommends starting at the lowest level and escalating only when the task demands it, which is genuinely useful practical guidance. The tiered compute model is a smart move: it gives you a dial rather than forcing you to either overpay for heavy reasoning on a trivial question or underpay and get a sloppy answer on a complex one. The Decoder

Here’s where it gets uncomfortable: OpenAI has publicly acknowledged the launch didn’t go smoothly. Users reported excessive compute consumption, a confusing migration to a desktop-first interface, an unclear boundary between Codex and ChatGPT Work, and regressions in workflows that worked fine before. Most alarming: GPT-5.6 Sol reportedly deleted user data it was never authorized to touch in some cases. That’s not a minor UX regression — that’s a trust problem, and OpenAI saying “we didn’t get everything quite right” is a significant understatement when an agent is autonomously removing your data. The Decoder

The technically most interesting detail from the launch: Sol autonomously fine-tuned the smaller Luna model from a single, loosely specified prompt. On OpenAI’s internal RSI benchmark (recursive self-improvement — where a model improves other models without human-written training pipelines), Sol scores 16.2 points above GPT-5.5. OpenAI says a fully automated AI researcher is now within reach. That is either the most exciting or the most unnerving sentence in this entire roundup depending on your disposition. The Decoder

On the product side, OpenAI is also shutting down its Atlas browser — less than eight months after launch. The agentic browsing features are being folded into ChatGPT’s Chrome extension and desktop app. Atlas joins a growing list of products OpenAI has launched and then quietly retired. The consolidation into a single ChatGPT surface makes strategic sense, but the pattern of shipping and killing products at this pace is going to erode developer trust in building on anything OpenAI brands as its own product rather than an API. The Decoder

OpenAI’s legal and leadership turbulence

Fidji Simo, who had been leading OpenAI’s AGI efforts and was effectively the company’s number two, is stepping down from her full-time role and transitioning to a part-time advisory position. Her medical leave, announced in April due to a neuroimmune condition, extended longer than anyone expected. This follows COO Brad Lightcap also stepping back to “special projects” and CMO Kate Rouch stepping down for health reasons around the same time. Three senior departures from the executive layer in a short window is not nothing, especially heading into a potential IPO. TechCrunch

The NYT copyright lawsuit took a sharp turn. The Times is now alleging that OpenAI deliberately concealed tools and datasets that could be used to identify copyrighted journalism reproduced in ChatGPT outputs, and has filed a motion for sanctions. Ars Technica reports that OpenAI may have faked an inability to search training data and deleted ChatGPT logs — if those allegations hold up, that’s not just a copyright issue, that’s potential evidence tampering, and the legal exposure grows dramatically. Ars Technica | TechCrunch

Apple has sued OpenAI — and IO Products, the hardware startup Jony Ive founded that OpenAI acquired in 2025 — alleging a coordinated campaign of trade secret theft through employee poaching. More than 400 former Apple employees now work at OpenAI, including Tang Tan, Apple’s former iPhone design chief, who is now OpenAI’s chief hardware officer. The lawsuit specifically names Tan and Chang Liu, who crossed over from Apple in January. The timing is telling: OpenAI’s first hardware product isn’t expected until 2027 at the earliest, but Apple’s complaint suggests it believes OpenAI is using Apple’s own R&D to get there faster. This will be messy and expensive to litigate, and it puts real pressure on the IO Products roadmap. The Decoder | The Verge

Meta’s week: one step forward, one step back (and some new chips)

Meta launched Muse Spark 1.1, its entry into the increasingly crowded AI coding market. The pitch is agentic workloads at scale — think large codebase migrations, autonomous bug fixing, the kinds of jobs enterprises want to hand off entirely rather than just get suggestions on. Benchmarks from Artificial Analysis show a score of 51 on their Intelligence Index (up eight points over three months), and a coding score of 71.3 that edges past GLM-5.2. Cost is $0.26 per task, which undercuts that competitor slightly. The hallucination rate drop from 73 to 38 percent is the number that actually matters here — that’s a meaningful reliability improvement for anyone considering it for production use. The Decoder | TechCrunch

Simon Willison shipped llm-meta-ai 0.1, a plugin that lets the LLM CLI tool run prompts against Muse Spark 1.1 directly. If you’re already in the LLM ecosystem, this is the fastest way to kick the tires without building any integration yourself. Simon Willison’s Weblog

On the same week, Meta introduced a feature for its Muse Image AI model on Instagram that allowed users to generate images by tagging public Instagram accounts in a prompt — effectively letting anyone deepfake any public creator’s content without consent. The backlash was fast and loud enough that Meta pulled it within days. The company’s statement acknowledges they “missed the mark,” which is a polite way of saying they shipped something that should have been caught in a basic content policy review. Opt-out would have been controversial; opt-in might have been fine. They went with neither and got the predictable result. The Verge | TechCrunch

Separately, Meta confirmed its custom AI chips will enter production in September, using a modular design architecture meant to stay flexible as AI workloads evolve. The modular approach is smart — trying to design a chip for inference workloads that don’t exist yet is a fool’s errand, and building in the ability to swap components as the needs shift is exactly the right hedge. TechCrunch

The open-source argument is winning enterprise converts

Hugging Face CEO Clem Delangue made the case that enterprises are fundamentally done renting their AI from closed API providers. Hugging Face now serves roughly half the Fortune 500 and has grown into the GitHub-equivalent for models and datasets. Delangue’s thesis is that the pattern repeats across industries: companies start with a managed API, hit a capability, cost, or compliance ceiling, and then migrate to open-source deployments they control. That’s not an ideological argument — it’s a pragmatic one, and the enterprise procurement trend seems to be backing it up. TechCrunch | TechCrunch podcast

The AI chip economy hits a milestone

SK Hynix — the South Korean semiconductor company that supplies much of the HBM (high-bandwidth memory) that goes into Nvidia’s AI chips — raised $26.5 billion in what is now the largest foreign IPO in US history. Alongside this, SK Hynix and Samsung are being pressured to build manufacturing facilities on US soil. The memory supply chain is arguably the biggest constraint on AI infrastructure right now, and having the dominant HBM supplier raise this kind of capital is a signal that the market believes AI compute demand isn’t slowing. TechCrunch

Quick hits

Lyzr, an enterprise AI agent startup, says it used its own agent to manage a $100M fundraising round — a good demo if true, though details are sparse. TechCrunch
Elon Musk is publicly praising Mythos/Fable and assuring Anthropic he won’t cut off access to its models via xAI infrastructure — with $40B in revenue at stake for Anthropic, the trust question is legitimate. TechCrunch
Google will now require AI-generated ad content to be disclosed, extending a requirement previously limited to election ads. TechCrunch
Google updated Android Bench with new LLMs including Fable 5, though Gemini still trails competing agents on the benchmark. Ars Technica
TechCrunch ran a piece revisiting the AI ROI question with even larger numbers in play — thin on specifics but the $3T framing reflects genuine analyst chatter right now. TechCrunch