A new, free Chinese AI model is rattling U.S. tech investors
China's Zhipu AI says its newest model can find software security bugs as well as Anthropic's most tightly restricted system.
The claim landed this week with no benchmark paper behind it, just a viral post. That hasn't stopped it from rattling a market already nervous about what's propping up American AI valuations.
The system Zhipu is comparing itself to, Claude Mythos, isn't available for outsiders to test.
Washington ordered Anthropic to cut off foreign access to Mythos and its public sibling, Fable 5, on June 12, citing national security. That means Zhipu's claim can't be independently verified, and may never be in its current form.
A real benchmark sits underneath the unverified one
Strip away the cybersecurity headline and a harder number remains. Zhipu's open model, GLM 5.2, landed last week with the kind of buzz that followed DeepSeek's debut a year earlier.
It has since pulled ahead of every other open release, sitting within a single percentage point of Anthropic's Opus 4.8 on a closely watched agentic benchmark at roughly a fifth of the cost, a CNBC report found.
Developers have noticed: OpenRouter token traffic for GLM 5.2 is climbing faster than it did after DeepSeek's V4 launch in April, per the same report.
That comparison matters because DeepSeek's moment faded once the market decided it was a one-off chatbot scare.
Related: Broadcom gets major OpenAI boost in AI chip race
GLM 5.2 is built for agentic work instead, planning, coding, testing, and looping through tasks on its own, which is exactly the kind of labor enterprises are trying to automate right now.
As AI token spending has climbed past what many companies budgeted for, intelligence per dollar has become the metric that matters, and a cheap, good-enough open model is a direct answer to that pressure.
"I've been consistently surprised by how quickly the open source has caught up," Gabe Pereyra, co-founder of Harvey, told CNBC.
Open and closed models are now solving different problems
The distinction underneath all of this is simple but easy to miss. Closed models like Opus 4.8 and Mythos live entirely on Anthropic's servers, accessed through an API the company, or a regulator, can switch off at will.
Open models like GLM 5.2 ship their weights under a license, MIT in this case, that lets anyone download, modify, and run the system on their own hardware.
That difference used to be mostly philosophical. It's now a procurement decision.
More AI:
- Micron just dethroned Nvidia in one key way
- How the White House is quietly bottlenecking AI
- Morgan Stanley resets Micron stock price target on strong AI demand
A company self-hosting GLM 5.2 isn't exposed to a future export order, a pricing change, or an outage at someone else's data center, though it does take on the engineering cost of running the model itself.
Hosting through Zhipu's own cloud API instead reintroduces a version of that dependency, since usage would fall under Chinese law rather than a self-hosted server's jurisdiction.
Access, not just capability, is now part of the valuation math
Here's what should bother U.S. AI investors more than any single benchmark: The National Security Agency lost its own access to Mythos as a result of the export order, Nextgov reported, even though NSA analysts had been using it to find vulnerabilities in classified systems.
A model good enough for the government to rely on got pulled before the government finished relying on it.
OpenAI separately moved to limit GPT-5.6 to "trusted partners" at the U.S. government's request the same week. A model nobody can revoke is starting to look like the safer bet to enterprises planning years out, regardless of which one currently scores higher on a leaderboard.
Big news: Mythos is partially back
The test arrived sooner than expected. The U.S. government cleared Anthropic to restore Mythos 5 access on Friday, but only to roughly 100 vetted organizations focused on critical infrastructure and cyber defense, NBC News reported.
Fable 5, the public-facing version most enterprises actually used, stays offline.
That distinction is the real signal. Commerce Secretary Howard Lutnick could amend the approved list at any time.
For the roughly 100 firms back in, the tool is faster than building the same defenses from scratch.
For everyone else, the lesson Zhipu has been selling all month just got reinforced by the U.S. government itself: A model you can't self-host is a model someone else can take away.
That's the deeper problem for Anthropic and OpenAI's combined $1.8 trillion-plus valuation.
A free, downloadable alternative closing the gap on cost and agentic performance was already a hard sell.
A premium model whose access list a Commerce Secretary can edit by letter, even temporarily, makes the case harder still.
Related: How the White House is quietly bottlenecking AI
The Arena Media Brands, LLC THESTREET is a registered trademark of TheStreet, Inc.
This story was originally published June 30, 2026 at 1:03 AM.