AI News Today — July 4, 2026:
The 5 Biggest Stories You Need to Know
Independence Day, but the AI industry isn’t taking a break. Grok 4.5 enters private beta at SpaceX, Meta reveals a GPT-5.5-class model codenamed Watermelon, OpenAI offers the US government a 5% stake ahead of its IPO, court documents reveal the Pentagon demanded Anthropic accept autonomous weapons, and China open-sources a 1.6-trillion-parameter model trained entirely on domestic chips. Here’s everything that moved today.
Grok 4.5 Enters Private Beta at SpaceX and Tesla
Elon Musk announced on X that Grok 4.5 has entered private beta testing inside SpaceX and Tesla as of June 28. The model is being evaluated internally before a broader release — no public access timeline has been confirmed yet. xAI’s previous model, Grok 4, scored competitively against GPT-5.5 and Claude Opus 4.8 on reasoning benchmarks, making 4.5 one of the most-watched upcoming releases in the frontier race.
Meta’s Secret “Watermelon” Model Matches GPT-5.5 Performance
Business Insider reported that Meta’s Chief AI Officer Alexandr Wang told a closed briefing that Meta’s model currently in training — internally codenamed Watermelon — matches GPT-5.5 performance on current evaluations. Critically, Wang said Watermelon uses “an order of magnitude more” compute than Meta’s previous frontier model. If Meta’s last frontier run used approximately 100,000 H100 equivalents, Watermelon is training on roughly 1 million GPU-equivalents — consistent with Meta’s announced plans for a 1-million-GPU cluster in 2026.
Pentagon Demanded Anthropic Accept Autonomous Weapons — Court Emails Reveal
Court documents released this week from Anthropic’s lawsuit against the Department of Defense include email exchanges between CEO Dario Amodei and Pentagon Undersecretary Emil Michael. The Wall Street Journal first reported the emails. Michael’s position was blunt: the ethical guardrails Anthropic wanted were “just not workable.” He offered Anthropic “one more chance to align on core principles” before announcing the talks were over. The dispute centers on a fundamental question: can an AI lab set ethical limits on a government customer?
OpenAI Offers US Government 5% Stake Ahead of IPO
OpenAI is reportedly offering the US government a 5% equity stake in the company as part of its IPO strategy. Governance scholars immediately flagged the structural problem: a regulator with an equity stake in the company it regulates cannot enforce rules impartially. OpenAI’s proposal also explicitly asks competing labs — Anthropic, Google, and xAI — to cede the same stake. Whether the White House accepts and whether Congress would authorise such an arrangement remain open questions. The IPO timing creates urgency: if announced before the S-1 becomes effective, it shapes the roadshow narrative entirely.
China Open-Sources 1.6-Trillion-Parameter Model on Domestic Chips Under MIT License
A Chinese research team has open-sourced a 1.6-trillion-parameter model trained entirely on domestically manufactured chips — no NVIDIA hardware involved. Released under an MIT license, the model is freely available for commercial use worldwide. This is significant for two reasons: it demonstrates that US chip export controls have not prevented China from training frontier-scale models, and it adds a massive open-weight model to the global ecosystem that any developer can run and fine-tune without licensing restrictions.
That’s today’s wrap. The Meta Watermelon revelation and the Pentagon-Anthropic court emails are the two stories that will shape the conversation this week — one about compute scale, the other about the limits of AI ethics in government contracts. Follow NeuralPaws daily for the next AI news drop.