AI This Week: Open Models Close the Gap
GLM 5.2 takes the design-taste crown at a sixth of the cost, Claude Fable stays offline by government order, and Anthropic becomes the most valuable lab. What it means for the tools you choose.
The big picture
Open models are catching the closed leaders
Free-to-run models, many from China, now rival the top US models on many tasks at a fraction of the cost. The gap is months, not years.
The money is enormous and unproven
Record losses and trillion-dollar valuations in the same week. Anthropic just became the most valuable of them all. The bubble question is live.
Your tools are being bought, and can be pulled
Giants now own the major AI tools, and one model line vanished overnight by government order. Vendor dependence is a real operating risk.
The question is now where is the return
Almost everyone has adopted AI. The winners this year are measuring outcomes, not running more pilots.
The Models and the Open-vs-Closed Race
Operator lens: Which model for which job, and how fast the cheap options are catching up.
Open weightsGLM 5.2 just took the design-taste crown from the US labsIn simple terms: A free-to-run Chinese model now tops the leaderboards for front-end and UI generation, at roughly one-sixth the cost of the leading US models.
GLM 5.2, released June 13 by Zhipu AI (Z.ai), is the clearest sign yet that open-weight models have caught up where it counts. It took the number-one spot on Design Arena for web and UI generation, and reviewers are specifically calling out its taste for clean front-end design, not just raw correctness.
- Tops Design Arena for front-end and UI; strong coding scores (around 62 on SWE-bench Pro).
- Roughly 5x to 7x cheaper on output than Claude Opus 4.8, under a permissive MIT license you can self-host.
- On the hardest long-horizon engineering tasks, the top closed models still lead.
Good enough at a sixth of the price changes the math for anything high-volume. Reserve the premium closed models for the genuinely hard, multi-step work.
Sources: Latent Space, llm-stats
GeopoliticsChina has nearly closed the gap, and leads on open modelsIn simple terms: The performance gap between the best US and Chinese models has shrunk to a couple of points, and Chinese labs now hold four of the top five open-model spots.
Despite the US spending far more on AI, the measured capability gap between the best American and Chinese models has narrowed to roughly 2.7 percent on Stanford's AI Index, down sharply from a few years ago. China now leads specifically on open-weight models and their global adoption.
- By some open-model leaderboards, Chinese labs hold four of the top five open-weight spots (GLM, Kimi, DeepSeek, and Qwen among them).
- The US still leads on frontier capability, chips, and capital; China leads on open adoption, cost, and research output.
- Caveat: Chinese models carry built-in restrictions on topics sensitive to the Chinese government, which matters for some uses.
You no longer have to buy frontier-priced US models to get strong results. But open and cheap comes with its own due diligence on data handling and content restrictions.
Sources: The New Stack, Understanding AI
Platform Risk and Vendor Lock-In
Operator lens: What happens when the model you depend on is not yours to control.
Live riskA whole Claude model line vanished overnight by government orderIn simple terms: The US government ordered Anthropic to pull its Fable 5 and Mythos 5 models for export-control reasons, and they went offline worldwide within hours. Nearly two weeks later they are still down.
On June 12 the US Commerce Department directed Anthropic to place Fable 5 and Mythos 5 under export controls, barring access by any foreign national anywhere. Because Anthropic could not verify nationality per user in real time, it took both models fully offline for everyone. As of June 23 they remain down, with Anthropic saying only that they will return in the coming days.
- Reporting (Fortune, TechCrunch) says Amazon's CEO had raised security concerns about the models to the White House beforehand; the order itself cites export-control grounds.
- Anthropic's other models (Opus 4.8, Sonnet, Haiku) are unaffected and fully available.
- This is the clearest real-world example of platform risk: a capability you built on can disappear for reasons unrelated to you.
If a single model is load-bearing in your business, you need a fallback. The lesson is not avoid Anthropic, it is do not be undiversified on any one model.
The Money and the Financials
Operator lens: How stable are the companies you are building on, and is this a bubble.
FinancialsOpenAI lost about $38 billion last yearIn simple terms: The most famous AI company made roughly $13 billion but lost around $38 billion. Even setting aside one-time items, it spent about $21 billion more than it earned.
Leaked, unaudited 2025 financials show roughly $13B revenue (ahead of their own target) against a $38.5B net loss. Much of that headline figure is a one-time, non-cash charge tied to restructuring; the underlying operating loss was about $20.9B. OpenAI paid Microsoft around $17B for the year.
The economics of frontier AI are brutal. This is the core of the bubble debate and a reason to assume pricing, access, and even providers could shift under you.
Sources: Quartz, Where's Your Ed At
ValuationsAnthropic is now reportedly the most valuable AI companyIn simple terms: Anthropic raised a giant round at a valuation that just passed OpenAI's for the first time.
Anthropic reportedly raised $65B at a roughly $965B post-money valuation, edging past OpenAI's private valuation for the first time. Separately, SpaceX (which now includes xAI) is targeting a $1.75 trillion-plus valuation at IPO.
The capital flooding in is staggering and the pecking order is shifting fast. Weigh a provider's momentum and staying power, not just today's benchmark scores.
Sources: Build Fast with AI
Consolidation: Who Owns Your Tools
Operator lens: The independent-tool era is closing. Know whose ecosystem you are entering.
MegadealSpaceX is buying Cursor for $60 billionIn simple terms: With this deal, every major AI coding tool is now owned by a tech giant. The independent-startup era for these tools is basically over.
SpaceX agreed to acquire the popular AI coding tool Cursor in a $60B stock deal (closing expected Q3). That completes the pattern: Microsoft owns Copilot, Anthropic owns Claude Code, OpenAI owns Codex, and now SpaceX and xAI own Cursor.
Picking a tool increasingly means picking a camp. Factor in who owns it and how locked in you would be before you standardize a team on it.
Sources: CNBC, TechCrunch
Adoption, Infrastructure and Rules
Operator lens: Where the rest of the market actually is, and the constraints forming around it.
Business realityThe conversation shifted from should we use AI to what is the returnIn simple terms: Almost every company now uses AI in some form. The hard part is no longer adopting it, it is proving it actually pays off.
Surveys show the vast majority of companies have deployed AI tools or agents in the past year, with a growing share of enterprise apps embedding task-specific agents. But most also report struggling to scale and to demonstrate clear ROI, alongside real governance gaps.
This is the gap you can win in. Measured outcomes and disciplined implementation now separate leaders from the crowd more than access to the latest model.
Sources: Business Standard, First Page Sage
Rules and policyEurope is delaying its toughest AI rules to 2027In simple terms: The EU pushed several of its strictest AI requirements from this August out to late 2027, giving businesses more breathing room, but the direction is still toward more regulation.
The EU's Digital Omnibus deal, provisionally agreed and not yet formally adopted, would move the toughest high-risk obligations from August 2026 to December 2027, while adding new bans (compliance by December 2026) on AI-generated abuse imagery. In the US, the picture is a patchwork of state laws plus a federal push toward a national framework, and, as Fable showed, the government will use export-control powers directly on models.
If you were racing an August EU deadline, that pressure eased. But do not mistake a delay for a reprieve, and remember governments can move on specific models quickly.
Sources: Inside Privacy, Holland & Knight
Sources and further reading
- GLM 5.2 / open models -- Latent Space
- US / China AI race -- The New Stack
- Fable 5 suspension -- explainX
- OpenAI financials -- Quartz
- Valuations / roundup -- Build Fast with AI
- SpaceX / Cursor -- CNBC
- EU AI Act -- Inside Privacy
Several figures this week (OpenAI's losses, Anthropic's and SpaceX's valuations, GitHub traffic volumes, model benchmark scores) come from leaked, reported, or provisional sources rather than official filings, and benchmark results vary by test setup. They are described as reported or provisional above and should be verified before being quoted as hard fact. Adoption statistics vary by survey and are directional.
Get The Skim every week.
One email, every week, on where AI is heading and what it means for operators.
The setup is the missing piece.
Most owners have the tools. The 30-minute call is about whether we can build the system around them.
30 minutes · No preparation needed