AI News Roundup — July 4, 2026
Even on US Independence Day, the AI industry kept moving: Grok 4.5 entered private beta at SpaceX and Tesla with 1.5 trillion parameters and near-Opus claims, China's Meituan open-sourced LongCat-2.0 — a 1.6 trillion parameter model trained entirely on domestic chips that had quietly topped OpenRouter usage as the anonymous 'Owl Alpha' — and Anthropic moved to close loopholes that let Chinese companies like Ant Financial and ByteDance access Claude through Singapore subsidiaries and VPN-reimbursed personal subscriptions.
Grok 4.5 Enters Private Beta at SpaceX and Tesla — 1.5 Trillion Parameters, Near-Opus Claims
Elon Musk announced on X that Grok 4.5, built on xAI's 1.5-trillion-parameter V9 foundation model, has entered private beta at SpaceX and Tesla. The model incorporates supplemental training data from Cursor (whose parent company SpaceX acquired for $60 billion in June), exposing it to vast real-world coding workflows. Early internal evaluations claim performance "close to, perhaps exceeding" Anthropic's Opus — though no independent third-party benchmarks exist yet. xAI is using SpaceX and Tesla as production evaluation environments, testing on actual rocket trajectory calculations and vehicle software development — harder tasks than any standard benchmark. Musk confirmed that SpaceX plans to release completely new V9-based models every month through the end of 2026, and the Grok Build coding harness is showing daily advancements during the beta period. On the longer horizon: Grok 5 is targeting 6–10 trillion parameters and is currently training on Colossus 2 alongside six other concurrent training runs. The monthly V9 variants (Grok 4.5, and future 4.6–4.9 through Q4 2026) are stepping stones in the reinforcement learning and capability improvement process feeding Grok 5.
LongCat-2.0 Open-Sourced by Meituan — 1.6 Trillion Parameters Trained Entirely on Chinese Chips
Meituan, China's largest food delivery platform, released LongCat-2.0 under an MIT license — a 1.6-trillion-parameter Mixture-of-Experts model trained entirely on a 50,000-card cluster of domestic Chinese ASICs with no US-restricted hardware. Benchmark results: 59.5% on SWE-bench Pro (narrowly above GPT-5.5 at 58.6%), 70.8% on Terminal-Bench, with a native 1M-token context window and dynamic parameter activation averaging 48 billion per token (33–56B depending on query complexity). The model had already been deployed anonymously as "Owl Alpha" on OpenRouter, where it topped developer usage rankings for weeks before the reveal. Meituan's strategy: deploy anonymously first to get merit-based adoption, then reveal attribution once track record was established. The result: developers who had dismissed Chinese models on geopolitical grounds discovered they'd already been using one successfully. With an MIT license and no regional restrictions, LongCat-2.0 joins GLM-5.2 and DeepSeek V4-Pro as a credible open-weight alternative for teams needing frontier-adjacent coding capability without US-origin model dependency.
Anthropic Moves to Block Chinese Access Loopholes — Ant Financial, ByteDance, Singapore Subsidiaries
Anthropic has stepped up efforts to close loopholes allowing Chinese companies to access Claude in violation of its terms of service, the Financial Times reported. Documented workarounds: Ant Financial provided employees corporate Claude accounts linked to Singapore-based subsidiaries, routing Chinese employees through a legally separate overseas entity; ByteDance reimbursed engineers for personal Claude subscriptions accessed via VPN, disguising corporate deployment as individual consumer usage. Detection methods Anthropic is now deploying include monitoring account timezone indicators and targeting transfer station services that relay requests through overseas Claude accounts — which connects directly to the hidden steganographic tracking system discovered in Claude Code this week, and disclosed as an "experiment" to prevent abuse and distillation. Anthropic's terms prohibit Chinese companies and entities under their control from using its models, but the geographic and legal fragmentation of modern multinationals creates substantial gray area. The enforcement escalation follows Anthropic's accusation that companies affiliated with Alibaba's Qwen lab ran 28.8 million fraudulent API exchanges — the largest known distillation attack — as industrial-scale capability extraction continues intensifying.
RAISE US Launches With Over $500 Million to Retrain American Workers for the AI Economy
RAISE US — the Reskilling American Industry, Science, and Enterprise for the United States — launched with over $500 million committed from Amazon, Anthropic, Microsoft, OpenAI, Bank of America, General Motors, and other corporations. Co-founded by former Commerce Secretary Gina Raimondo, the nonprofit will partner with states and community colleges to retrain workers in industries most exposed to AI displacement. The initiative is a direct response to the political and economic backdrop of 2026: US jobless claims have been rising, enterprise AI budgets are maxing out as automation accelerates, and the populist backlash against AI wealth concentration — evident in both the Sanders 50% ownership proposal and Trump's 5% government stake overture — reflects growing public anxiety about who benefits from the AI economy. RAISE US frames retraining not as charity but as a business necessity: AI companies need an expanded workforce capable of building, deploying, and managing AI systems, and the current talent gap is a bottleneck on revenue growth.
Menlo Ventures Raises $3 Billion After 'Betting the Firm' on Anthropic Pays Off
Menlo Ventures closed a $3 billion fund — the largest in the firm's 50-year history — on the back of its Anthropic investment, which the firm made when few others believed in the company. Managing Partner Shawn Carolan had described the Anthropic bet as "betting the firm," and the fund's oversubscription reflects LP appetite for the Anthropic-to-IPO return story. Menlo will use the new capital to back AI startups across all stages, from seed through growth. The fund signals that the private VC ecosystem sees Anthropic's trajectory — $47B annual revenue run rate, first operating profit on track, October 2026 IPO targeting $965B valuation — as an anchor return that validates the broader AI startup thesis. Menlo's success also illustrates how concentrated AI venture returns are becoming: a single Anthropic bet positioned a mid-size firm to raise its largest fund ever, while firms without frontier AI portfolio exposure are finding fundraising increasingly difficult.
US Government AI Model Pre-Release Review — Why the New Precedent Matters
With the Fable 5 ban resolved and the White House finalizing voluntary frontier AI standards ahead of the August 1 deadline, analysis pieces are crystallizing why the precedent matters beyond one model. Under the emerging framework, the US government can designate models as "covered frontier models" — triggering a 30-day pre-release national security review before any trusted partner access. The framework is nominally voluntary; in practice, a lab that skips it faces the risk of an emergency directive like the one that took Fable 5 offline for 19 days. Both OpenAI and Anthropic have publicly endorsed the general concept while calling specific implementations "not the long-term default." The deeper questions analysts are now asking: what happens when a foreign ally's cybersecurity agency needs access to a covered model for legitimate defense work? What happens if a well-resourced adversary can simply train equivalent open-source models domestically? And does the classification of an AI model as a controlled export — subject to citizenship-based access restrictions — set a precedent that applies to future AI outputs, safety research publications, or training data?