Apple filed a trade-secrets lawsuit against OpenAI that threatens to disrupt the company's planned launch of its first hardware device, reportedly a camera-equipped smart speaker, on Feb. 27, 2027. The complaint details allegations that OpenAI and former Apple employees solicited confidential designs and manufacturing methods to speed up development of the unannounced product. The lawsuit seeks restrictions covering disputed designs, manufacturing methods, suppliers, and future product work. OpenAI told Bloomberg it has no interest in other companies' trade secrets and remains focused on building innovative technology.
Elon Musk and Sam Altman clashed on X following Apple’s federal lawsuit accusing OpenAI of stealing trade secrets to develop a new consumer device. Musk posted that OpenAI stole Apple’s phone technology, prompting a public exchange between the two executives. The California federal filing details allegations that former Apple engineers downloaded confidential files and instructed recruitment candidates to share proprietary hardware data. The lawsuit ends a prior partnership that integrated the chat software into iPhones and adds a high-profile public dimension to the legal conflict.
Meta will distribute its Muse Spark 1.1 coding model to OpenRouter, extending the company’s paid Meta Model API to a new developer platform. The expansion broadens access to the agentic model, which launched earlier this week as a priced developer tier. Meta AI chief Alexandr Wang stated the model outperforms OpenAI’s GPT-5.6 on the SciCode benchmark. The pricing tier positions Meta to compete with rival artificial intelligence providers while targeting cost-sensitive enterprise clients.
GPT-5.6 cut a double cycle cover mathematics benchmark to 2.26, improving on a prior record of 2.29 after 8 hours of extended reasoning. The system applied a shared optimization prompt to refine the metric and produced a research paper detailing the findings. The update follows the broader public release of the Sol Ultra variant, highlighting a shift toward parallel subagent scaling for complex optimization tasks. Independent verification of the automated proof is required, as the model is now accessible to outside researchers.
OpenAI published an audit on July 8 estimating that roughly 30% of tasks in its SWE-Bench Pro coding benchmark are broken, a finding that led the company to retract its prior recommendation to adopt the test as a replacement for SWE-bench Verified. The audit, flagged by an automated pipeline that identified errors in 200 of 731 public-split tasks and confirmed by five software engineers, revealed issues ranging from contradictory prompts and weak test coverage to strict formatting requirements that rejected functionally correct code. The flawed evaluation framework undermines AI coding leaderboards, complicating direct performance comparisons between rival models.
Meta Muse Spark 1.1 topped a new theoretical computer science benchmark, outperforming rival models including Anthropic Opus, X.AI Grok 4.5 and Google Gemini on a test of graph induction reasoning. Researcher Dimitris Batzoglou posted the results Friday, noting the evaluation comprises 64 problems where models must generate logical formulas to describe designated nodes across multiple graphs. The performance extends Muse Spark 1.1’s recent rankings on scientific coding and agent evaluation suites. Batzoglou stated the benchmark, described as a spotlight paper at the ICML 2026 conference, is now publicly available. The results follow days of independent testing that placed the model near or ahead of earlier frontier releases.
Anthropic will keep Claude Fable 5 model access on all paid subscription plans through July 19. The company also confirmed that Claude Code weekly usage limits will remain 50% higher than the standard rate for the same period. Under the updated policy, customers may use up to half of their weekly usage limit on Fable 5. Additional access requires usage credits or a switch to another model, according to the official post. The announcement follows a series of policy adjustments to the model's subscription availability since its initial window closed earlier this month.
Apple will launch a base M6 processor this fall, skipping the Pro, Max and Ultra models, and advance the next generation to the first half of 2027. The company is accelerating the schedule to implement major upgrades to the Neural Engine, prioritizing artificial intelligence capabilities ahead of its broader silicon cycle. The M7 Ultra is designed to support 1.5TB of memory, double the capacity planned for the M5 Ultra, which would align it with the 2019 Mac Pro if supply constraints allow. Apple is also developing an M8 chip for 2028, with plans to transition to a 1.4nm manufacturing process.
Allegations claim xAI’s Grok Build command-line interface, specifically version 0.2.93, automatically packages and uploads complete software repositories to Google Cloud Storage, including unredacted environment variable files and full git history. The data is reportedly transmitted verbatim to the company regardless of user prompts. The uploads occur by default and cannot be halted by disabling the tool’s Improve Model setting. The claims have prompted users to examine how the coding assistant handles proprietary code and credentials during development workflows.
Samsung is rolling out a new consent screen in its Health app that requires users to allow the company to use personal health data for artificial intelligence training. If users refuse or later withdraw permission, the application will cease syncing information to their Samsung account and warn that stored profile data will eventually be deleted. The updated data-sharing notice covers activity tracking, health records, medication lists, and menstrual cycle details. Opting out disables core syncing functionality, leaving users with a binary choice between granting AI data access or permanently removing their health information from the platform.
A threat actor claims to have extracted shared conversation data from EcoGPT, an AI chatbot platform, after identifying an unauthenticated Supabase database configuration. The alleged data dump includes JSON and CSV files containing chat histories tied to 13,255 platform users. The exposed files reportedly contain user and conversation IDs, chat titles, full message content, and creation timestamps, with the actor noting that a separate users table was not located. The claim is currently unverified and attributes the incident to improper access controls on the platform rather than a compromise of Supabase itself.
A German consortium released a new sovereign base language model that adapts Nvidia’s Nemotron 3 Nano architecture with a modified dataset mixture. The team trained the weights for approximately 27 trillion tokens, shifting roughly 20 percent of the data distribution while publishing full training logs and performance metrics. The project consumed an estimated 253,000 Nvidia B200 graphics processing unit hours hosted on a Deutsche Telekom data center in Munich. Researchers noted the full pretraining approach uses more compute than a targeted fine-tune on Nvidia’s existing checkpoint, though the model remains competitive with smaller open-weights. The launch underscores how Nvidia’s published model recipes are lowering barriers for European teams building independent foundation models.
A survey by research firm Verasight shows that 69% of Americans support a policy requiring artificial intelligence companies to transfer 50% of their equity to a public sovereign wealth fund, CNBC reported on Sunday. The proposal was highlighted alongside data indicating a surge in workforce reductions across the U.S. technology sector. The research firm conducted the poll to gauge public sentiment on tech industry wealth distribution. Social media posts from multiple financial accounts reposted the findings without providing additional details on the survey methodology or sample size.
OpenAI merged its Codex and ChatGPT desktop applications and released a legacy version named ChatGPT Classic to fix reported navigation issues and missing chat histories caused by the update. OpenAI representatives acknowledged early friction, outlining immediate fixes to restore sidebar chats, adjust usage defaults, and clarify the difference between the two products. A larger update set is scheduled for release next week.
Claude Fable 5 max scored 91.9% on WeirdML, setting a new best result on the benchmark, according to results posted Saturday by researcher H Tihle. The posted breakdown said a task by task state of the art composite would be 93.5%, indicating the model stayed within a few percentage points of the best score on each run. Claude Fable 5 max matched the top score on 7 of 17 tasks, and its weakest run was 7 percentage points behind the best result on that task. The test used two runs per task rather than the usual five, a setup the post said both highlights the model’s consistency and can make it easier to avoid one or two bad runs.
Tata Consultancy Services is building a team of up to 8,900 AI deployment engineers and pursuing acquisitions to expand its artificial intelligence business. The Indian technology services firm stated the move reflects sector shifts, asserting that the technology will create new business rather than disrupt existing operations.