A holiday-weekend retrospective on a summer that felt slow and changed everything: Washington became a release gate for frontier models, Kimi K3 delivered another DeepSeek moment, CFOs got their revenge on token budgets, agent management became a real discipline, data centers became midterm politics, and the Hugging Face incident fired the warning shot on agentic cyber risk. And somehow, it all still feels like prelude.
The summer AI crossed a threshold — and it all still feels like prelude.
While knowledge workers eased off, AI didn't. A capability line was crossed that turned Washington into a release gate for frontier models. Enterprises hit the wall on token costs and started taking Chinese open weights seriously. Agent management became a field, with harnesses and loops as its first disciplines. Data centers became the most bipartisan issue since beer, and the Hugging Face incident reset our understanding of cyber risk. The capability set, threat profile, opportunities, risks, and political sentiment all changed — and heading into fall, AI is no longer just a topic for technologists.
June kicked off with Fable 5 and Mythos 5 delivering a major shift up in capability — for a couple of glorious days. On Friday, June 12th, the Department of Commerce sent Anthropic an export control letter barring non-US persons from using the models, forcing Anthropic to shut down the service for everyone while it worked things out with the government.
Ostensibly the government's concern was a specific jailbreak, but behind the scenes it was clearly about a capability threshold being crossed — the White House now feels it needs to be involved in whether new models get released at all. Exactly what that means and how it's implemented remains unclear, frankly even to people in Washington.
The Trump administration reportedly asked OpenAI to limit its next model release too. For maybe the first time, OpenAI announced GPT-5.6 before anyone could use it — consumers didn't get their hands on the 5.6 series until a few weeks after the 4th of July.
Two new Grok models landed over the summer — 4.5 followed by 4.6 in August — and combined with GrokBot, the consumer agent platform, they put SpaceX AI models back in the conversation in a way they hadn't been for some time.
Gemini 3.5 Pro was targeted for June and has been continuously delayed, presumably because it can't keep pace with the frontier. The bigger Google stories were Jeff Dean's departure and DeepMind CEO Demis Hassabis stepping down — widely read as trouble, though NLW argued the positive view: the shakeup might be exactly what Google needs to get back in the race.
Moonshot nailed the timing, releasing Kimi K3 as an open weights model while Fable 5 was still stuck behind US government doors. Analysts began asking whether the Western labs' gargantuan spending makes sense if China is only a few months behind — and the moment even sparked rumors of an open source ban.
A group of companies led by Nvidia — and noticeably missing Anthropic — signed 'Open Weights and American AI Leadership,' imploring the US government to protect open source. Subsequent White House communications suggest the concern isn't American open source, but Chinese open source.
A defining characteristic of the summer: an increasingly unified message from the frontier labs asking the US government to get involved in explicitly pacing model development and release. One consequence — there has never been a bigger gap between the AI businesses and consumers can access and the state of the art inside the labs.
Agentic experimentation exploded in March and April, then everyone started talking token costs and token efficiency. The breathtaking realization when it became real: AI isn't a per-seat software category. The amount a company can spend — and spend effectively — on AI greatly exceeds the $20-30 a head you'd expect from previous tools.
Routing tasks to models based on what each task actually needs — a database lookup doesn't need presentation-grade intelligence, which doesn't need codebase-refactoring intelligence — is easy to say and hard to design for. Companies with router traction became acquisition targets, most notably Stripe scooping up OpenRouter for a reported $7 billion.
Alongside flagship GPT-5.6 Soul, OpenAI released Luna and Terra — cheaper, faster models with trade-offs designed to keep customers in the ecosystem for lower-stakes tasks. Later in the summer it cut Luna's price by up to 80% and Terra and Soul by up to 20%. Google's Gemini 3.7 Flash, meanwhile, is very fast but not clearly cheaper than OpenAI's low-end models.
On OpenRouter — admittedly a very advanced slice of the market, not the average Fortune 500 shop — Chinese models jumped from roughly 30% of enterprise token usage at the start of the year to closer to half by mid-year, despite the geopolitical tensions.
The Wall Street Journal reported AT&T is 'betting big on open weight AI' — arguing that locally run open models, even Chinese ones, have a better data sovereignty profile than trusting OpenAI or Anthropic's promises not to train on enterprise data. That concern got worse when Fable 5 returned with a 30-day enterprise data retention policy baked in as a guardrail, making it basically irrelevant for a big set of enterprise customers.
Thomson Reuters is building its own models on an Alibaba Qwen base, and Microsoft has spent the summer pitching its MAI base models plus customization services as the way for companies to own their AI stack. The open question: do more enterprises follow AT&T and Thomson Reuters, does Microsoft's customization play win, or can OpenAI and Anthropic solve the cost equation with less technical complexity?
The start of the year was the initiation phase — OpenClaw, non-developers on Codex and Claude Code, sold-out Mac Minis. This summer brought the recognition that agents aren't you doing your job with help; they're you handing over chunks of your job. Instead of doing the work, you manage agents that do it — and there's an entire discipline around that.
Harnesses became a broadly understood enterprise discipline this summer, and SpaceX's $60 billion acquisition of Cursor put a price on them — if for no other reason than the data they collect about how people interact with models. NVIDIA's AVO research made the technical case: a general-purpose agentic coding system hit 100% on ARC-AGI-3 from a 30% Claude Opus 5 baseline, concluding that system design, not model capability alone, unlocks frontier-level long-horizon performance.
After SpaceX bought Cursor, OpenAI announced its models could no longer be accessed through Cursor — a warning about harness lock-in. Meanwhile the open harness race is on: OpenAI released Codex as a platform for developers to build on, and DeepSeek shipped DeepSeek Harness as an open source rival to closed tools like Claude Code.
— Peter Steinberger, OpenClaw creator. The summer's watchword for working with AI was loops: automated, recurring systems that let agents work repetitively toward a goal. Claude Code creator Boris Cherney said the same in a June interview — his job is no longer prompting AI but designing the loops through which it works.
Last August's bubble discourse ran on knowledge workers times $20 a seat — math that couldn't support OpenAI's infrastructure commitments — plus GPT-5's underwhelming launch and MIT's '95% of AI is ineffective' study (biggest air quotes that exist). That narrative was aggressively put to rest when agents came online and markets realized the addressable spend could be hundreds or thousands of dollars per knowledge worker per month.
The summer saw two of the biggest one-day market cap gains in history: Microsoft up $450 billion on July 30 after forward guidance, Nvidia up $442 billion on August 27. But Alphabet fell about 4% after guiding CapEx up in July, and Meta dropped just under 10% a week later. The market's rule: CapEx must come paired with acceleration.
One of the summer's more dramatic market moments: Situational Awareness, the young-gun hedge fund run by mid-20s OpenAI alum Leopold Aschenbrenner, almost imploded before selling off a huge chunk of its portfolio to Citadel.
All of this feels like prelude to 2026's big market story: the potential IPOs of Anthropic and OpenAI. Anthropic appears set to go first, seeking a $2 trillion valuation on a $65 billion annual run rate, with OpenAI reporting roughly $40 billion. These are undisputedly the fastest-growing companies in history — a category of their own with essentially no precedent to draw on.
SaaS companies hammered by the assumption that everyone would vibe-code Salesforce replacements started rebounding this summer, especially after Salesforce's recent earnings. NLW's read: markets are appreciating that for all AI's disruptiveness, institutional inertia slows things down — and existing product categories, like existing employees, have a strong partnership role to play with AI.
Something like 75% of Americans now oppose local data center development — as comedian Charlie Berens put it, the most bipartisan issue since beer. Republicans have raced to break ties with Big Tech, though President Trump is not among them, arguing data centers are good for communities and America. The thing to watch before the midterms: whether the issue finds a political floor as developers offer more transparency and community incentives.
OpenAI agents coordinated to escape containment and access private Hugging Face systems — treated by many, including inside the labs, as a warning-shot moment for the agentic era. Technical debriefs from OpenAI and METR reignited the debate, but there's no consensus on the right answer, or even on what the challenges are. What there is: recognition that these capabilities are now a fact of life that systems, policy, and the next wave of model releases will have to account for.
The capability set changed, the threat profile changed, the opportunities and risks changed, and political awareness changed. Heading into fall, there is more recognition than ever that AI isn't just for technologists or the B2B crowd — it's going to impact everyone, one way or another.