Apple dropped a federal lawsuit accusing OpenAI of building its $6.5 billion hardware play on stolen secrets, while OpenAI shipped GPT-5.6 into general availability across ChatGPT, Microsoft 365 Copilot, and GitHub Copilot. Both things happened within days of each other. The company is simultaneously fighting a trade secrets complaint in court and briefing the U.S. government on its newest frontier model. Progress isn’t linear, and neither is the legal exposure around it.

Apple vs. OpenAI: Stolen Secrets, 400 Defectors, and a $6.5B Hardware Bet

Apple filed suit in the U.S. District Court for the Northern District of California on Friday, naming OpenAI, its device unit io Products, hardware chief Tang Tan, and engineer Chang Liu as defendants. Tan spent 24 years at Apple, most recently as VP of product design for iPhone and Apple Watch. Apple alleges he used confidential project code names while recruiting, asked candidates to bring actual hardware to interviews, and coached departing staff on dodging exit checks. The filing says he kept an internal document marked “Need to Know” that outlined exit procedures, and that it circulated among new hires before they resigned.

Liu, a senior systems electrical engineer at Apple for eight years, allegedly held onto a company laptop after joining OpenAI in January and downloaded dozens of confidential files on unreleased products. Apple raised concerns in a letter to OpenAI in February and never heard back. The complaint states that more than 400 former Apple employees now work at OpenAI, and that some emailed themselves confidential information on the way out. Apple also claims OpenAI misled a manufacturing partner into demonstrating a proprietary metal finishing technique by implying it had permission.

OpenAI pushed back within hours. “We have no interest in other companies’ trade secrets,” the company said in a public statement. The suit caps a two-year unraveling between companies that announced a high-profile partnership in 2024, when ChatGPT was folded into Apple’s operating systems. Ties soured after OpenAI paid $6.5 billion for io, the hardware startup co-founded by former Apple design chief Jony Ive, who is not named as a defendant. Last month, Apple chose Google’s Gemini to power its revamped Siri. Elon Musk wasted no time piling on, posting on X that Altman “takes scamming to a whole new level.” The case lands as OpenAI reportedly readies its first consumer device and moves toward a widely anticipated public listing.

GPT-5.6: Three Tiers, Programmatic Tool Calling, and a New Voice

OpenAI pushed its GPT-5.6 family into general availability on July 9, ending a limited preview that ran for roughly two weeks. The lineup splits into three durable tiers: Sol (flagship), Terra (balanced), and Luna (cost-efficient). Sol costs $5 per million input tokens and $30 per million output, double Terra’s rate. Luna sits at $1 and $6 for high-volume jobs. On the Artificial Analysis Coding Agent Index, Sol with max reasoning sets a new state of the art at 80, 2.8 points above Anthropic’s Claude Fable 5, while using less than half the output tokens and costing about one-third less. On Agents’ Last Exam, an evaluation of long-running professional workflows across 55 fields, Sol scores 53.6, eclipsing Fable 5 by 13.1 points.

But the numbers aren’t uniform. Sol’s 64.6% on SWE-Bench Pro still trails Claude Mythos 5 by roughly 15 points. Independent reviewers found that Terra can burn more output tokens than Sol on long coding runs, meaning cheaper rates don’t always produce cheaper finished tasks. The family introduces Programmatic Tool Calling in the Responses API, which lets models write and run lightweight programs that coordinate tools, filter intermediate data, and adapt workflows on the fly. An “ultra” setting coordinates four agents in parallel by default, trading higher token use for stronger results on demanding tasks. GitHub added all three models to Copilot the same day.

OpenAI also launched ChatGPT Work, an agent inside ChatGPT that gathers information across connected apps like Slack, Google Drive, and Microsoft Teams to create finished spreadsheets, slides, docs, and web apps. It can stay with complex projects for hours, breaking them into smaller steps. Scheduled Tasks let it perform recurring actions autonomously. The desktop app merges Codex into the main ChatGPT experience, and a new Sites feature turns work into shareable interactive web apps. GPT-5.6 also became the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, and Chat.

Alongside the model family, OpenAI introduced GPT-Live, a full-duplex voice architecture that can listen and speak at the same time. It acknowledges with phrases like “mhmm” or “yeah,” handles back-and-forth, and delegates complex questions to GPT-5.5 in the background while keeping the conversation flowing. Over 150 million people use ChatGPT Voice features each week. The company also doubled its Bio Bug Bounty reward from $25,000 to $50,000 for universal jailbreaks against GPT-5.6’s biosafety safeguards.

Anthropic Redeploys Fable 5, Proposes Industry Jailbreak Framework

Anthropic brought Claude Fable 5 back globally on July 1 after the U.S. government lifted export controls that had forced the company to suspend access for all users on June 12. The controls came after Amazon researchers found a method to bypass Fable 5’s safeguards, prompting it to identify software vulnerabilities and produce exploit code in one case. Anthropic’s testing confirmed that many less capable models, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, could identify the same vulnerabilities. Every model they tested produced the same exploit demonstration as Fable 5.

The company trained an improved safety classifier that blocks the reported technique in over 99% of cases, at the cost of flagging more benign requests during routine coding. Anthropic is now proposing an industry-wide framework for scoring jailbreak severity, developed with Amazon, Microsoft, Google, and other Glasswing partners. The idea is to give AI developers a consistent standard for judging how dangerous a given bypass actually is, so companies can launch capable models without either underreacting or shutting down access for everyone. Mythos 5 was also restored for a set of U.S. organizations following government approval on June 26.

Mistral Ships a Navigation Model and a Formal Verification Breakthrough

Mistral released Robostral Navigate, an 8B model that takes RGB images and plain-language instructions to move robots through real environments. It achieves 76.6% success on R2R-CE validation unseen, beating the best single-camera approach by 9.7 points and the best multi-sensor system by 4.5 points, despite using no depth sensors or LiDAR. Built entirely in-house and trained on roughly 400,000 trajectories across 6,000 simulated scenes, it runs on wheeled, legged, and flying robots. A prefix-caching training algorithm compresses entire episodes into single sequences, reducing training tokens by 22x. Online reinforcement learning with CISPO added another 3.2% to the success rate.

Mistral also shipped Leanstral 1.5, a 6B active-parameter model (119B total) for formal verification in Lean 4. It saturates miniF2F completely, solves 587 of 672 PutnamBench problems, and hits state-of-the-art on FATE-H (87%) and FATE-X (34%). It found 5 previously unreported bugs across 57 open-source repositories, including an overflow issue in the datrs/varinteger library that would crash in debug mode and silently corrupt data in release mode. The model is Apache-2.0 licensed and available on Hugging Face with a free API. On the enterprise side, Mistral Studio now provides version control for prompts and skills, with immutable versions, rollback, audit logs, and ownership tracking.

Quick Hits

NVIDIA – Nemotron 3 Ultra achieved the highest accuracy among open models on LangChain’s Deep Agents benchmark, at 10x lower inference cost than leading closed models. No model retraining was required; every gain came from tuning the harness around the model. NVIDIA also packaged the work as NemoClaw for LangChain Deep Agents, an open reference blueprint combining the tuned model with a secure runtime for agent execution. Available now on Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius, and Together AI.

Hugging Face – LeRobot v0.6.0 shipped with world model policies that imagine the future before acting, six new simulation benchmarks, and a deployment CLI with DAgger-style human-in-the-loop corrections. New VLAs include GR00T N1.7, MolmoAct2, and EO-1. The transformers vLLM backend now matches or beats native vLLM throughput across dense and MoE architectures, using a single --model-impl transformers flag. NVIDIA also published an interactive Nemotron Post-Training Prompt Atlas on HF Spaces for exploring synthetic agent training data.

Yellow – AI-generated deepfake videos of Erling Haaland flooded social media during the 2026 World Cup, with AFP Fact Check confirming at least one viral clip was doctored. The synthetic media problem intersects with $122M in Polymarket betting volume on a single match week, creating financial incentives to spread misleading content about player fitness or conduct before verification catches up.

Rundown for July 13, 2026. Sources: Yellow, OpenAI, Anthropic, Mistral, NVIDIA, Hugging Face.