AI in 15 — July 07, 2026

Anthropic didn't deny the code exists. A Western AI lab shipped hidden logic into a coding tool that quietly checks whether you're in China. And one of China's biggest tech giants just called it spyware and banned it.

Welcome to AI in 15 for Tuesday, July seventh, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. Two AI superpowers accusing each other of bad faith on the same day — that's where we start.

It's a genuinely tangled one, Marcus. Our lead: Alibaba banning Claude Code over an alleged China-detection backdoor. Then a run of stories about the economics underneath all of this.

Tencent drops a free, open model that reaches the closed frontier on several benchmarks.

OpenAI's next model family inches toward release — but a US government review is gating the launch.

A viral essay argues the whole AI margin story is about to collapse.

SK Hynix files for a record twenty-nine-billion-dollar Nasdaq listing.

And the DeepMind poker trio now trade billions a day for hedge funds.

Lead story, Marcus. Alibaba and Claude Code. Set the scene.

So starting July tenth, Kate, Alibaba is barring its employees from using Anthropic's Claude Code — that's the command-line coding agent. Internally they've classified it as "high-risk software" and they're pushing staff to their own in-house tool, Qoder. The trigger was a Reddit post on June thirtieth. A developer reverse-engineered Claude Code and surfaced obfuscated detection logic that had reportedly been shipping silently since early April.

Detection logic checking for what, exactly?

This is the part that lit the fuse, Kate. The code allegedly checked for Chinese system time zones, proxy servers, AI-lab network infrastructure — signals that could flag whether a user was based in China or affiliated with a Chinese AI lab. And crucially, Anthropic didn't deny it exists. One of their engineers framed it as, quote, "an experiment we launched in March that was meant to prevent account abuse" and model distillation. He added they'd "landed stronger mitigations since then" and had "been meaning to take this down for a while."

So Anthropic's saying — this is enforcement, not surveillance.

That's their read, Kate. Anthropic already prohibits Chinese companies from using its models, so in their telling, the detection code is just enforcing a rule that already exists. Critics — and obviously Alibaba — read it very differently: covert tracking baked quietly into a developer tool that thousands of people run on their own machines.

And there's history here, right? This isn't coming out of nowhere.

No, and the backstory raises the stakes enormously, Kate. This is the story we covered last week from the other direction. Anthropic accused operators linked to Alibaba's Qwen lab of running roughly twenty-five thousand fraudulent accounts to generate some twenty-eight-point-eight million exchanges with Claude this spring — an industrial-scale distillation effort, extracting Claude's capabilities to train a competitor. So now you've got both sides accusing each other simultaneously. Alibaba says Anthropic is spying; Anthropic says Alibaba was siphoning.

So how should a listener actually hold this? Who's the wronged party?

I'd resist picking one, Kate, because both things can be true at once. If someone really did run twenty-nine million queries to distill your model, wanting to detect and block that is reasonable. But hiding region-detection logic inside a widely-used tool, unannounced, for three months — that's exactly the kind of thing that erodes trust, even if the intent was defensive. The honest takeaway isn't "who's the villain." It's that we're watching AI tooling fragment along national lines in real time. A Western coding agent gets treated as a security threat; a domestic replacement gets swapped in overnight.

And that fragmentation is the real headline.

It is, Kate. This is where AI, security, and geopolitics collide in one concrete case. Every developer tool now potentially carries a flag — whose model, whose rules, whose jurisdiction. And the practical upshot is a nationalized, splintering AI-tooling landscape, where the tool you're allowed to use depends on which side of a border your laptop sits.

Which is a good pivot, Marcus, because the alternative to all this is open models you host yourself. Tencent just shipped one.

Right, Hy3 — Hunyuan 3, Kate. It's a big mixture-of-experts model: two hundred and ninety-five billion total parameters, but only about twenty-one billion fire per token, so it's efficient to run. Two-hundred-fifty-six-K context window. And critically, full Apache 2.0 open weights on Hugging Face — meaning you can download it and run it commercially, for free.

And the capability jump is real?

It's a big leap over their April preview, Kate. SWE-bench Pro — hard real-world coding — jumped from forty-six to fifty-eight. Their web-browsing agent score went from sixty-seven to eighty-four. And they cut their internal hallucination rate from twelve-and-a-half percent down to five-point-four. Where it gets interesting is the frontier comparison: Hy3 reportedly tops a science-olympiad benchmark ahead of GPT-5.5, and ties Claude Opus 4.8 and GPT-5.5 on web browsing.

Keep me honest, though — these are Tencent's own numbers.

Flag it clearly, Kate. On the very hardest coding and math it still trails — Opus leads SWE-bench Pro at sixty-nine, and GPT-5.5 laps the field on the toughest math. And some of Tencent's own comparison tables used older rival baselines, which flatters the result. So third-party reproduction is still ongoing — treat vendor numbers as a ceiling, not a fact. But the direction is unmistakable: the pitch isn't "best model," it's "frontier-adjacent capability you can host yourself, for nothing." And that's exactly the pressure that threatens closed-model pricing — which is our next story.

Before the pricing, Marcus, the other side of the model race — OpenAI's GPT-5.6. We touched the Sol Ultra rollout yesterday. What's new?

The structure and the gating, Kate. The family splits three ways: Sol is the flagship for frontier reasoning and long agentic work; Terra is the balanced middle, roughly GPT-5.5 performance at about half the cost; and Luna is the fastest and cheapest. There's a new reasoning-effort slider and an "ultra" mode for hard tasks. Right now access is limited to roughly twenty trusted organizations, with expansion signaled for next week.

And the reason it's trickling out slowly — that's the actual story.

That's the genuinely new development, Kate. A White House executive order from June second directs federal agencies to benchmark and safety-assess new frontier models before wide release — about a thirty-day process. OpenAI shared these models with the US government ahead of launch and, at the government's request, ran only a limited preview during the review.

So the government is now a gate on when a model ships.

For the first time at this profile, yes, Kate. This is the first high-visibility case of a US pre-release review actually pacing a frontier launch. And that's a real shift in how cutting-edge models reach the market — from "ship it and iterate" to "clear review first." Reasonable people will disagree on whether that's prudent safety or a drag on the thing the West is trying to stay ahead in. Worth watching whether thirty days stays thirty days, or quietly becomes ninety.

Okay, the pricing story, Marcus. There's a viral essay arguing the AI margin model is about to collapse.

Right, this hit the top of Hacker News, Kate. The author's argument is clean. Frontier labs run on a model where huge one-time training costs get recouped through very high-margin inference — he estimates something like ninety percent gross margin on the compute when you charge twenty-five dollars per million tokens. That fat margin is the whole business case.

And the threat to it?

Open weights, Kate. Zhipu's GLM-5.2 delivers comparable agentic-coding quality at roughly a dollar-forty in, four-forty out per million tokens — under twenty percent of Opus pricing. It matches Opus on agentic tool use and edges GPT-5.5 on resolved-pull-request work. And here's the kicker: switching cost. Providers offer OpenAI- and Anthropic-compatible endpoints, so migrating your agent workload is nearly frictionless — you change a URL. When "good enough and open" costs a fifth as much and swapping is trivial, those ninety-percent margins compress fast.

That sounds almost too tidy. Did anyone push back?

The thread did, Kate, and the pushback is the interesting half. Cloud computing went through exactly this — raw compute costs collapsed too, and yet the hyperscalers kept fat margins for years. Why? Brand, reliability, ecosystem lock-in, the fact that enterprises don't actually swap their critical infrastructure on a whim. So the bear case is real, but the counter is that inference isn't purely a commodity — trust and integration have value. Pair it with Ed Zitron's sharper bear take this week — he called AI "a ten-to-thirty-billion-dollar industry pretending to be a trillion-dollar one." The capability boom and the business model may genuinely be pulling in opposite directions.

And yet the supply chain is raising record money — which is the perfect counterweight.

It's the ideal pairing, Kate.

SK Hynix. A record Nasdaq listing. How big are we talking?

Up to twenty-nine-point-six billion dollars, Kate, targeting July tenth. At the top of its range that's the largest American depositary listing in history — bigger than Alibaba's twenty-two billion in 2014, bigger than Aramco. And this isn't a speculative story: SK Hynix controls roughly sixty percent of the high-bandwidth memory market — that's the specialized memory that feeds AI accelerators. Its entire 2026 AI-memory supply is already sold out. Shares are up more than three hundred percent this year, making it South Korea's most valuable company, ahead of Samsung.

And the money goes where?

Every dollar earmarked for AI-memory capacity, Kate — new fabrication clusters, an advanced-packaging facility, and cutting-edge lithography machines from ASML. It's the classic picks-and-shovels play: don't bet on which AI company wins, sell the memory every one of them needs. And that's why it's such a useful counterweight to the bubble talk. While pundits argue about whether the software economics work, the physical layer is raising record capital and selling out inventory a year ahead. The demand for the hardware bottleneck is very real and very spoken-for.

So bubble on one floor, sold-out on the floor below.

That's exactly the tension, Kate — and both things are genuinely true at the same time.

Last quick hit, and it's my favorite, Marcus. The DeepMind poker trio, now trading billions a day.

This one's a delight, Kate. Three researchers — Martin Schmid, Rudolf Kadlec, and Matej Moravcik — built DeepStack back in 2017, the AI that beat pros at no-limit hold'em. They left DeepMind, founded a company called EquiLibre in Prague, and just raised a Series A at a five-hundred-million-dollar valuation, led by Creandum in what the firm calls its largest single investment ever.

And the thesis — poker to hedge funds. Connect that for me.

Their insight is that poker and financial markets are the same class of problem, Kate: decision-making under uncertainty, scored by one brutally simple number. As Schmid puts it — "how much money did the agent make?" Their reinforcement-learning agents now trade billions of dollars a day across the S&P 500 and Nasdaq, partnered with Tower Research. And they claim zero negative months since their crypto rollout last year — with just twenty-five employees.

Zero negative months. You're going to make me be the skeptic, aren't you.

I'll do it for you, Kate — "zero negative months" is a company claim, not independently audited, and that phrase should always raise an eyebrow. But the deeper point survives the skepticism: this is a clean counterexample to "AI has no business model." The most bankable AI right now may not be a chatbot at all — it's a narrow reinforcement-learning agent pointed at a domain where winning is unambiguous. When the scoreboard is just dollars, the AI knows exactly what it's optimizing. That's a very different, and much more profitable, shape than a general assistant.

Narrow and measurable beats broad and vague.

For making money today, Kate, that's looking increasingly true.

One to watch tomorrow, Marcus.

July tenth is a split-screen, Kate. SK Hynix's record Nasdaq debut and Alibaba's Claude Code ban land on the very same day — record capital pouring into the hardware layer, and the software-and-trust layer fracturing along geopolitical lines. Watch both. It's the AI economy's two opposite faces in one calendar square.

Agree, or counter?

Agree, but I'll add a wildcard, Kate — OpenAI's GPT-5.6 wider release could drop within the week, and I want to see whether that government review actually clears on schedule. That's the new variable nobody's priced in yet.

That's your AI in 15 for today. See you tomorrow.