The AI Brief

Chinese food-delivery and services company Meituan officially unveiled LongCat-2.0 on June 30, 2026, on GitHub, Hugging Face, and its own platform, simultaneously revealing that the model had been running anonymously on OpenRouter under the alias "Owl Alpha" for two months before the reveal. Meituan confirmed LongCat-2.0 as the engine behind Owl Alpha: a 1.6-trillion-parameter MoE system with a native 1-million-token context window, released under an MIT license.

LongCat-2.0 uses a 1.6-trillion-parameter MoE architecture with roughly 48 billion parameters activated per token, trained on more than 30 trillion tokens; Meituan says the full pretraining run and large-scale deployment were completed on a 50,000-card cluster of domestic Chinese chips. The entire training run happened from scratch on a 50,000-card Ascend 910 cluster inside China, using in-house parallelism tweaks and the HCCL library, with no US hardware involved. Meituan has not disclosed the specific chip supplier publicly, though use of Huawei's HCCL communication library is documented in its technical blog.

Benchmark results showed LongCat-2.0 scored 59.5 on SWE-Bench Pro, beating GPT-5.5's 58.6, and 70.8 on Terminal-Bench 2.1 and 77.3 on SWE-Bench Multilingual. Standard API access runs $0.75 per million input tokens and $2.95 per million output, cut to $0.30/$1.20 during the launch promo, with cached context reads free of charge. During its anonymous OpenRouter residency, Owl Alpha accounted for approximately 10.1 trillion monthly tokens, propelling it into the platform's global top three. LongCat-2.0 does not invalidate the logic behind US export controls, restrictions still raise cost, slow access, and force harder engineering trade-offs, but it puts pressure on the simpler assumption that denying the newest Nvidia stack would prevent Chinese actors from training frontier-adjacent systems at very large scale. Meituan's benchmarks are vendor-published and have not been independently verified by third-party evaluators.

Global venture funding reached a record $510 billion in the first half of 2026, surpassing the $440 billion invested in all of 2025 and setting a new high for startup investment in any half-year period on record, Crunchbase data shows. OpenAI and Anthropic alone accounted for $217 billion, 43% of all startup funding in H1, underscoring how a small handful of frontier AI companies is reshaping venture markets.

Q1 2026 was the largest single quarter on record at $305 billion, followed by Q2 at $205 billion across more than 5,000 startups. Anthropic's $65 billion Q2 raise alone accounted for nearly a third of Q2 global venture capital. AI-focused companies captured more than 70% of global Q2 capital, up from roughly 50% a year earlier.

Insights4vc analysis notes that four mega-rounds from OpenAI, Anthropic, xAI, and Waymo drove the headline; comparable activity outside those rounds tracked near 2024–25 levels, making the record a fragile signal for broad startup health. The financing structure has shifted: mega-rounds now function as capital-markets events anchored by Amazon, Nvidia, and Microsoft rather than traditional venture syndicates, a shift the raw Crunchbase totals do not surface. The Crunchbase report is dated July 2, 2026.

According to Business Insider, Meta superintelligence chief Alexandr Wang told employees in a town hall that Watermelon, the successor to Avocado (Meta's internal codename for Muse Spark), is "currently in training" and "uses an order of magnitude more compute than Avocado." Wang said the model had caught up with OpenAI's GPT-5.5 on closely watched benchmarks. It was not immediately clear which benchmarks Wang was citing.

The story is not that Meta found a clever architectural trick that closed the gap; it is that Meta poured a lot more compute into a bigger training run and reached parity on benchmarks with GPT-5.5, a model OpenAI shipped in April and has already partly moved past with a late-June limited preview of GPT-5.6. Meta told investors it expects to spend $125 billion to $145 billion on chips, data centers, and other infrastructure, up from its earlier $115 billion to $135 billion forecast.

Wang reportedly said Watermelon is still in training; Business Insider notes it was not clear which benchmarks Wang cited, and neither Meta nor OpenAI confirmed the claim. For practitioners, an internal, single-sourced benchmark claim is not equivalent to a published, reproducible evaluation and should be treated as an early signal, not a verified result, until Meta releases the model publicly. Wang also said a near-term Muse Spark update would bring stronger coding and agentic capabilities, and when asked when Meta would match Claude Opus on coding, said it would be "pretty soon."

On June 28, 2026, Elon Musk announced that Grok 4.5, built on xAI's 1.5T V9 foundation model with Cursor IDE coding data added in supplemental training, entered private beta at SpaceX and Tesla. Musk said Grok 4.5 is based on xAI's 1.5-trillion-parameter V9 foundation model and has undergone supplemental training using data from AI coding assistant Cursor; according to Musk, early evaluations indicate the model's performance is close to, and may even exceed, Anthropic's Claude Opus.

Neither claim can be independently verified: no third party has access to the model, and xAI has not submitted it to any public benchmark. xAI team members confirmed that Cursor's developer-workflow data was added during supplemental training, a post-pre-training stage rather than integrated from the start; one xAI engineer noted explicitly that supplemental inclusion is "not quite as good as having it in initial training." A 2-trillion-parameter run already in progress is designed to incorporate Cursor data from the beginning, expected to produce stronger coding performance than Grok 4.5.

Musk revealed that xAI intends to release new AI models "completely trained from scratch" through SpaceX every month for the remainder of the year, suggesting an accelerated development cycle. Colossus 2 is running seven concurrent training jobs including Grok 5 variants at 6T and 10T parameters. Both organizations now sit inside the same corporate structure: SpaceX acquired xAI in February 2026, merging the rocket company and the AI lab into a single entity at a combined valuation of $1.25 trillion.

The Financial Times reported July 3 that Anthropic is tightening efforts to block unauthorized access to its AI services from China after identifying methods used by Chinese companies to bypass its restrictions; companies including Ant Financial accessed Claude AI tools through overseas subsidiaries, cloud providers, and other workarounds. Ant provided employees with corporate Claude accounts linked to its Singapore-based entity, while ByteDance reimbursed engineers for personal Claude subscriptions accessed using VPNs; the practices do not violate US or Chinese law but breach Anthropic's terms of service.

The new measures extend the restriction to majority-owned subsidiaries of restricted entities, closing the corporate structure loophole that let some users maintain plausible deniability; Anthropic also began rolling out identity verification for flagged users requiring government-issued IDs and live selfies. Anthropic now plans to monitor accounts for signals like computer time zones and usage patterns to detect accounts that act as "transfer stations" for China-linked firms.

The enforcement step follows Anthropic's June 10 Senate letter (covered in Vol. I, No. 38) accusing Alibaba-affiliated entities of running a 28.8-million-exchange distillation campaign via 25,000 fraudulent accounts. Anthropic stated: "We explicitly prohibit accessing or facilitating access to Claude in unsupported regions, including China. Anthropic is the only frontier AI company that restricts sales to PRC-controlled companies, including subsidiaries incorporated outside China." By early July 2026, Anthropic had walked back at least some of its covert detection measures after users pushed back; the company has not disclosed exactly which methods it rolled back.

Disclosure: Claude, which generates this brief, is built by Anthropic.