The last few weeks in frontier AI have been dominated by access fights rather than pure capability jumps.

US export controls on Anthropic’s newest models just ended.

On June 30 the Department of Commerce lifted the emergency controls it had placed on Claude Fable 5 and Mythos 5 roughly three weeks earlier. Anthropic confirmed the models are redeploying.

Fable 5 is now rolling out globally across the Claude platform, Claude.ai, Claude Code, and Claude Cowork starting July 1. Pro/Max/Team plans get inclusion for up to half of weekly usage through the first week, then credits. Enterprise via the usual cloud partners (AWS, Google Cloud, Microsoft Foundry) will follow as fast as they can re-enable.

Mythos 5 is back first for approved US organizations, with Glasswing cybersecurity partners next in line.

Primary source: Anthropic’s official redeploy post.

This matters for anyone running serious agentic coding workflows. Fable 5 was the model pushing the highest numbers on production-mergeable benchmarks before the ban. Teams that had it hard-wired into long-horizon agents or internal harnesses can stop scrambling for workarounds.

Meanwhile, China just demonstrated it can train frontier-scale models without Nvidia.

Meituan (the massive food delivery company) open-sourced LongCat-2.0, a 1.6-trillion parameter Mixture-of-Experts model trained start-to-finish on a 50,000-chip cluster of domestic Chinese processors.

Key specs from their announcement:

• 1.6T total params, ~48B active per token
• 1 million token context
• MIT license (weights “coming soon” per repo notes)
• Claims performance comparable to Google’s Gemini 3.1 Pro on relevant evals
• Positioned explicitly for agentic coding and long-context reasoning workloads

Sources: Reuters coverage, VentureBeat reporting, and the company’s statements.

This is the clearest public proof yet that US chip export controls are not a hard blocker for Chinese labs at the very top end. They trained the full pre-training run plus inference on home silicon. Previous Chinese models often did inference domestically but still leaned on restricted hardware for the heavy training.

For open-source and indie agent builders this is huge. A permissive license at that scale changes the playing field for anyone who wants to fine-tune or run long-horizon agents without sending every token to a US frontier provider.

xAI continues its internal-first strategy.

Grok 4.5 (V9 architecture, 1.5T params) entered private beta at SpaceX and Tesla around June 28. Elon noted Cursor developer data was added in supplemental training. No public API, no independent benchmarks yet. The plan is monthly new-from-scratch foundation models through the end of the year.

It’s the same pattern: dogfood the biggest model inside the companies that actually stress it on real engineering before wider release.

The pattern across these stories is clear.

Export controls can slow things down and create friction, but they don’t stop capability from spreading. Chinese labs are proving they can close the hardware gap on their own silicon. Meanwhile the US labs are playing a combination of “controlled release” and “internal dogfooding” while open releases (even at extreme scale) keep landing under permissive licenses.

For people actually shipping agentic coding systems right now, the practical takeaway is wider model choice at the high end than we had 30 days ago — both through restored access and through new open weights you can actually run or adapt.

The wall isn’t capability. It’s still orchestration, cost control, and verification of what the agent actually shipped.

Keep building. The tool selection just got marginally less constrained.