AI Weekly Roundup #2: GPT-5.6, Apple vs OpenAI, and China AI Overtakes the US (July 7-13, 2026)
AI Weekly Roundup #2: GPT-5.6 Launch, Apple Sues OpenAI, and China’s AI Overtakes the US (July 7-13, 2026)
This was the biggest week in AI so far in 2026. OpenAI launched GPT-5.6 — a three-tier model family with multi-agent capabilities that had been held up by US regulators for two weeks. Apple sued OpenAI for allegedly stealing hardware trade secrets. Chinese AI models overtook US models in global usage for the first time. GLM-5.2, an open-source Chinese model, was adopted as the default engine by Databricks. Anthropic’s revenue surpassed OpenAI’s.
If you only read one AI news summary this week, this is the one. Here are the 7 stories that actually matter — not the noise, not the press releases, just what changed the landscape.
Browse previous editions in our AI blog.
1. GPT-5.6 Officially Launches — Three Models, Multi-Agent Capabilities, and the Codex Merger
The story: On July 9, OpenAI launched GPT-5.6 to the public after a two-week regulatory review by the US government. This is not a single model — it is a family of three, each targeting a different tier of use. The naming scheme (Sol, Terra, Luna — Latin for Sun, Earth, Moon) signals that OpenAI is adopting a permanent tiered release strategy going forward: every generation will have flagship, balanced, and lightweight versions.
The three models:
| Model | Positioning | Input (per 1M tokens) | Output (per 1M tokens) | Compared to GPT-5.5 |
|---|---|---|---|---|
| Sol | Flagship — coding, science, cybersecurity | $5 | $30 | Same price, stronger performance |
| Terra | Balanced — daily work, general tasks | $2.50 | $15 | Matches GPT-5.5 at half the price |
| Luna | Lightweight — speed and cost priority | $1 | $6 | Lowest cost option in the lineup |
The key number: Terra delivers GPT-5.5-level performance at half the price. This is the most commercially significant detail — it directly lowers the cost for businesses and developers who need reliable AI without flagship pricing. For individuals and small teams, Luna at $1/$6 per million tokens makes API access genuinely affordable for the first time.
Ultra mode and multi-agent coordination: The Sol model introduces two new reasoning modes. Max mode allows deeper, longer reasoning — the model “thinks” more before responding. Ultra mode is the real breakthrough: it coordinates four AI agents simultaneously by default, scaling up to 16. Each agent handles a different part of a complex task in parallel — one researches, one writes code, one reviews, one tests — and Sol orchestrates the results into a single output. In practical terms, Ultra mode turned a complex coding benchmark (Terminal-Bench 2.1) from 88.8% to 91.9%, pulling ahead of Claude Mythos 5’s 88.0%.
GPT-5.6 Sol also became the first frontier model to officially pass the ARC-AGI-3 benchmark — a notoriously difficult test of abstract reasoning that most models score near zero on. The score of 7.8% may sound low, but it represents the first verified breakthrough on a test designed to resist memorization-based solutions.
Codex merges into ChatGPT: In a strategic move targeting Anthropic’s Claude Code dominance, OpenAI merged its standalone Codex coding app directly into the ChatGPT desktop application. The new ChatGPT app now has three unified tabs: Chat, Work, and Codex. The Work agent — powered by GPT-5.6 — can accept a high-level goal, break it into steps, and operate across email, calendar, Slack, cloud storage, and CRM tools to produce finished deliverables. This is a direct shot at Anthropic, whose Claude Code product has been growing 10x month-over-month.
Why this matters for you: If you use ChatGPT’s free tier, you now have access to GPT-5.6 Luna for everyday tasks — it is faster and cheaper to run, which means fewer rate limits. If you use the API, Terra gives you last-generation flagship performance at half the cost. And if you need the absolute best, Sol’s Ultra mode represents the new state of the art in multi-agent reasoning. For a broader look at how GPT-5.6 compares to other tools, see our DeepSeek review and AI writing tools comparison.
2. Apple Sues OpenAI for Trade Secret Theft — 400+ Former Employees Now Work There
The story: On July 11, Apple filed a lawsuit against OpenAI in US federal court, alleging systematic theft of hardware trade secrets. The complaint claims over 400 former Apple employees now work at OpenAI — a number that suggests a deliberate talent drain rather than ordinary hiring. Apple alleges that former iPhone design VP Tang Tan, whose startup io Products was acquired by OpenAI for $6.5 billion, instructed Apple employees he was interviewing to bring physical iPhone parts and CAD design files. Another former engineer is accused of using an unreturned Apple laptop to access internal networks and download confidential files.
OpenAI’s response: “We have absolutely no interest in anyone else’s trade secrets.”
Why this matters: This is not a standard patent dispute. Apple rarely files lawsuits of this nature — the company prefers quiet settlements and cross-licensing deals. The public filing suggests Apple believes the damage is ongoing and severe enough to warrant legal escalation. It also reveals that OpenAI is actively building hardware capabilities, likely for AI-specific devices or data center infrastructure, and has been aggressively recruiting Apple’s hardware engineering talent to do it. This is the first clear signal that OpenAI sees hardware as a strategic priority, not just a side project.
3. Chinese AI Models Overtake US Models in Global Usage
The story: Data from OpenRouter — the largest independent API gateway for AI models — shows that Chinese AI models now account for over 30% of global weekly token volume, up from 11% in November 2025. Domestic Chinese models handled 23.45 trillion overseas API calls in a single week. For the first time, Chinese models are generating more usage outside China than US models.
What is driving this: Price. Chinese models are 60-90% cheaper than their Western equivalents on a per-token basis. DeepSeek V4-Flash costs $0.14 per million input tokens. GPT-5.6 Luna — OpenAI’s cheapest new model — costs $1. The price gap is not narrowing; it is widening. For developers and businesses running high-volume AI workloads, the math is simple: if a Chinese model delivers 90% of the quality at 10% of the price, that is a rational business decision, not a political one.
The implications: This marks a structural shift in the AI market. The era when “the best AI” automatically meant “the most-used AI” is over. A new dynamic is emerging where US labs compete on frontier capability and Chinese labs compete on cost and accessibility. Both positions are defensible, but they lead to very different business models. OpenAI and Anthropic charge premium prices for premium models. DeepSeek, GLM, and Kimi are building market share through volume. The next 12 months will determine which strategy wins. Our DeepSeek review covers one of the key players in this shift.
4. GLM-5.2 Becomes the First Chinese Model Adopted as Default by a Major US Tech Company
The story: Databricks — one of the largest enterprise data platforms — announced it is switching its default AI coding engine to GLM-5.2, an open-source model from Chinese AI lab Zhipu. Internal testing showed GLM-5.2 performing within 1% of Anthropic’s Claude Opus 4.8 on the FrontierSWE coding benchmark, while costing significantly less. Cryptocurrency exchange Coinbase also switched to a combination of GLM-5.2 and Kimi 2.7, cutting its AI costs by approximately 50%.
This is a milestone: It is the first time a major US enterprise has adopted a Chinese open-source model as its default, not as an experiment or a fallback. When Databricks — a company that sells data infrastructure to the Fortune 500 — publicly endorses a Chinese model for production workloads, it signals that the quality gap has closed to the point where cost becomes the deciding factor. For a broader comparison of open-source AI models, see our productivity tools roundup.
5. Anthropic’s Revenue Surpasses OpenAI’s — $47 Billion Annual Run Rate
The story: Anthropic has reached an annualized revenue run rate of $47 billion, surpassing OpenAI for the first time. The primary growth driver is Claude Code — Anthropic’s AI-powered development environment — which is growing 10x month-over-month. Enterprise adoption of Claude for coding workflows has accelerated dramatically, and Anthropic’s focus on safety and reliability appears to be winning over risk-averse corporate buyers.
The broader picture: Five major AI labs — OpenAI, Anthropic, Google DeepMind, Meta, and xAI — have collectively borrowed $350 billion over the past five years to fund model training and infrastructure. The revenue numbers are enormous, but so is the spending. The industry is in a high-stakes race where winning requires both the best models and the best business model. OpenAI’s GPT-5.6 launch and Anthropic’s Claude Code growth represent two different answers to the same question: how do you turn world-class AI into a sustainable business? For a comparison of how these tools work in practice, see our beginner’s guide to choosing the right AI.
6. Quick Hits: Stories Worth Knowing
Gemini 3.5 Pro launches July 17. Google announced that Gemini 3.5 Pro will be released this Thursday (July 17) with a 2-million-token context window and aggressive pricing aimed at winning back developers who have migrated to cheaper alternatives. This will be the first major post-GPT-5.6 response from a competitor.
NVIDIA Vera CPU targets the AI agent era. NVIDIA launched its Vera CPU architecture on July 6, designed specifically for the high-concurrency, low-latency workloads that AI agents demand. Traditional multi-core CPUs struggle with the rapid tool-calling and reasoning chains that multi-agent systems like GPT-5.6 Ultra require. Vera is NVIDIA’s answer to that bottleneck.
China’s “human-like AI” regulation takes effect July 15. A new regulation governing AI services that present themselves as human goes into effect tomorrow. Platforms including Doubao (ByteDance) and Qianwen (Alibaba) have already removed or modified AI companion features to comply. This is part of a broader global trend — AI regulation is shifting from principles to enforcement.
Zhipu announces “Touch High” plan: no monetization for 2 years. Zhipu AI founder Tang Jie published an internal letter declaring the company will abandon short-term revenue goals for the next two years and focus entirely on AGI research. The lab completed a funding round reportedly in the tens of billions of yuan, giving it the runway to compete with Western labs on research rather than revenue.
The Big Picture
Three patterns from this week that will shape the rest of 2026:
- AI is becoming multi-agent by default. GPT-5.6’s Ultra mode, Claude Code’s workflow automation, and NVIDIA’s Vera CPU all point in the same direction: the next phase of AI is not about better single-model output, but about orchestrating multiple models working together. If you are building anything with AI, start thinking in terms of agent workflows, not single-prompt responses.
- The price war is accelerating, not slowing. GPT-5.6 Terra cut prices 50% while matching last generation’s performance. Chinese models are 60-90% cheaper than US equivalents. Google’s Gemini 3.5 Pro is positioning on price. The cost of running AI workloads is dropping faster than anyone predicted — and that is good news for creators and small businesses who can now access frontier AI at a fraction of last year’s cost.
- Open-source is winning the enterprise. Databricks adopting GLM-5.2, Coinbase switching to open-source Chinese models, DeepSeek’s MIT-licensed V4 — the enterprise trend is clear. Companies want models they can audit, self-host, and fine-tune, even if they sacrifice a few points on benchmarks. The most commercially successful AI models of 2026 may not be the ones with the highest benchmark scores.
That is it for this week. See you next Monday. Want more AI news and analysis? Check out our guide to AI content rules across platforms and our prompts for making AI remember your projects.
