OpenAI unveiled GPT-5.6 — its most powerful model to date — in three tiers: Soul, Terra, and Luna. Before it could reach the public, the US government stepped in and blocked access. Only vetted partners with approved credentials can use it. This isn't the first time: Claude Fable faced the same restriction five weeks earlier.

The pattern is becoming clear. The most capable frontier models are no longer being released to the public. They're being gated behind government review, partner agreements, and identity verification. The era of "anyone can try the best AI" is quietly ending.

Anthropic released Claude Sonnet 5 — and it's the most interesting model launch of the month. It delivers Opus-class reasoning at Sonnet pricing: $2/M input tokens, $10/M output tokens (introductory pricing until August 31, then $3/$15). For developers doing agentic coding, multi-step debugging, and complex workflows, this is the model to switch to right now.

The significance here isn't just the price — it's the compression of capability. Six months ago, Opus-level reasoning cost Opus-level money. Now it doesn't. This is what healthy model competition looks like.

Here's the story nobody's connecting the dots on: AI data centers are consuming memory chips at a scale that's creating a global shortage. The result? MacBook Pro prices jumped ₹70,000 (~$840) overnight in India. Xbox prices were raised globally. Micron — a memory chip manufacturer — became the most-traded stock in America.

The mechanism is straightforward: AI training and inference require enormous amounts of HBM (High Bandwidth Memory). Data centers are buying it faster than it can be manufactured. That leaves less supply for consumer electronics, which drives up prices for everyone.

Moonshot AI's Kimi K2.7 Code is now available in GitHub Copilot's model picker — making it the first open-weight model to appear inside the world's most popular AI coding assistant. Enterprise admins need to enable it manually, but once enabled, developers can choose Kimi K2.7 alongside GPT and Claude models.

This is a significant moment for the open-weight movement. It's one thing to run an open model locally. It's another to have it available inside a tool that 5+ million developers use every week. Kimi K2.7 is now competing directly with closed frontier models on their home turf.

OpenAI filed its S-1 on June 8, targeting a valuation between $830 billion and $1 trillion. To put that in context: that would make it one of the most valuable companies in American history at IPO. The numbers backing it up: ChatGPT has 1 billion monthly active users. Codex has 5 million weekly users. And in a notable strategic move, OpenAI is offering the US government a 5% stake as part of the IPO structure.

That government stake is worth paying attention to. It's not charity — it's alignment. An OpenAI where the US government is a shareholder is an OpenAI with a very different set of incentives than the nonprofit research lab it started as.

Reflection AI secured $6.3 billion in compute contracts running through 2029. This isn't a funding round — it's a reservation. They've locked in GPU capacity years in advance, betting that compute will be the scarce resource that determines who can build frontier AI and who can't.

They're probably right. The infrastructure race is no longer just about who has the best researchers or the most data. It's about who has guaranteed access to the chips needed to train and run the next generation of models. Compute is becoming the new oil.

Google shipped two notable releases this month. Gemini OmniFlash brings conversational video editing at $0.10 per second of video output — a genuinely new capability. Nano Banana 2 Lite delivers rapid image generation in approximately 4 seconds. Both include C2PA credentials and SynthID watermarks by default — Google is baking AI content provenance into the product rather than treating it as an afterthought.

While OpenAI and Anthropic are focused on text and code, Google is pushing hard on multimodal: video, images, and audio. The $0.10/sec video pricing is aggressive — it makes AI video editing accessible to individual creators, not just studios.

xAI released Grok 4.5, positioning it as a challenger to Opus-class models. Available to X Premium+ subscribers, Grok 4.5 continues xAI's pattern of shipping at an aggressive pace — faster than most labs would be comfortable with. Whether it actually matches Opus-class performance on real-world tasks is something the community is still benchmarking.

What's notable is the distribution strategy: Grok is tied to X (Twitter) Premium+, which means its user base is defined by social media subscribers rather than developers or enterprise customers. That's a different market than Anthropic or OpenAI is targeting.

Following OpenAI's lead, Anthropic filed for IPO in July 2026. This means both of the most important AI safety-focused labs — the two companies most vocal about responsible AI development — are now on the path to becoming publicly traded companies answerable to shareholders.

This is a structural shift in the AI industry. Public companies have quarterly earnings calls. They have growth targets. They have activist investors. The "safety first, revenue second" ethos that both labs were founded on will face its most serious test when Wall Street starts asking about monetization timelines.

Meta announced plans to build a cloud business selling its excess AI compute capacity. The backstory: Meta has been stockpiling GPUs at scale for Llama model training. They now have more capacity than they need for internal use, and they want to monetize the surplus by offering it to external developers and companies.

This is significant because Meta's compute costs are already amortized across their core business. They can afford to undercut AWS, Google Cloud, and Azure on price. If they follow through, this could create a genuinely competitive low-cost option for AI developers who are currently paying premium cloud prices.