The Best Model in the World is Now Free

Open weights stopped chasing the frontier this week and started setting it — even as Washington reminded the closed labs who really controls the off switch.

The open-source narrative flipped this week. The most-discussed model on the planet right now isn't from OpenAI, Anthropic, or Google — it's an MIT-licensed model from a Chinese lab that anyone can download for free. Zhipu AI's GLM-5.2 took the top open-weight spot on independent benchmarks, Nvidia shipped its most capable open model yet, and Mistral kept Europe's sovereign AI bet alive. Meanwhile, the U.S. government export-controlled two of Anthropic's frontier models overnight — a reminder that "frontier" and "in your control" are not the same thing.

Chinese lab Zhipu AI released GLM-5.2 to its coding subscribers on June 13 and then published the full open weights under an MIT license on June 16. It's a 744B-parameter mixture-of-experts model (about 40B active) with a one-million-token context window — five times longer than its predecessor. According to benchmark tracker Artificial Analysis, GLM-5.2 is now the leading open-weights model on its Intelligence Index, scoring 51 and edging out MiniMax-M3 and DeepSeek V4 Pro.

The numbers that matter to a business are the coding and cost figures. On long-horizon software-engineering benchmarks, GLM-5.2 beats GPT-5.5 and trails Anthropic's Claude Opus 4.8 by about a single point — at roughly one-sixth the cost of the closed leaders. And because it's plain MIT, there are no acceptable-use restrictions, no regional limits, and no vendor lock-in: a company can host it on its own infrastructure and freely commercialize the output.

For most of the last two years, open models arrived looking strong on benchmarks and then quietly faded. This one feels different — even skeptical AI watchers say GLM-5.2 "passes the vibe check", and Zhipu AI is already forecasting an open "Fable-class" model by year-end. The takeaway for leaders: the question is no longer whether open models can compete, but which of your routine, high-volume workloads still justify paying frontier prices.

The license is the real disruption. VentureBeat called GLM-5.2 a "Pure Open" system because the unmodified MIT terms allow enterprises to run frontier-class AI on sovereign infrastructure with no royalties and no governance strings attached. That's a sharp contrast with the "open-ish" community licenses that cap users or restrict commercial use — and a direct challenge to the per-token economics of the closed labs.

Read it here →

The average time it takes for the best open-weight models to catch up with state-of-the-art closed models, according to Epoch AI's May 2026 Capability Index (ECI) update. While the open model performance gap reopened slightly in early 2026, the arrival of models like GLM-5.2 shows that the trailing distance is now measured in mere months, not years — forcing enterprise leaders to rethink whether the frontier premium is still worth paying for routine workloads.

Source: Epoch AI

The U.S. government issued an export-control directive suspending all foreign-national access to Anthropic's top models, forcing the company to disable Fable 5 and Mythos 5 for hundreds of millions of users outside the country. Anthropic disputed that a narrow jailbreak justified the move, but complied. The episode is a vivid reminder that access to closed frontier models can change with a policy memo — and a quiet argument for why some enterprises are hedging with open weights they actually control.

Nvidia released Nemotron 3 Ultra, a 550B-parameter (55B active) open model shipped with open weights, training data, and recipes and built for fleets of long-running agents. It's the smartest open model from a U.S. lab to date on the Artificial Analysis index — though Chinese models like GLM-5.2 and Kimi K2.6 still sit ahead. For enterprises standardizing on Nvidia hardware, an efficient open reasoning model tuned for their GPUs reduces the cost of scaling from dozens to hundreds of agents.

Mistral's latest, Medium 3.5, powers agentic "Work mode" in Le Chat and is available with open weights on Hugging Face under a modified MIT license. The release reinforces Mistral's pitch as Europe's sovereign AI alternative to U.S. providers — open weights, on-prem deployment, and data residency, aimed squarely at CIOs who can't or won't route sensitive workloads through American clouds.

DeepSeek's MIT-licensed V4 Pro keeps topping community coding leaderboards, scoring around 80.6% on SWE-bench Verified — the best open-weights entry — at roughly $0.87 per million output tokens versus $25 for Claude Opus. Independent reviewers now rate it the strongest open-weight model on real-world "vibe coding" tasks. The pattern across this week's releases is consistent: open-coding models are landing within a point or two of the closed frontier at one-tenth to one-thirtieth the price.

The open-model hub now hosts just under 2.95 million models — and the second million arrived in roughly a third of the time the first took. The flood is both the opportunity and the problem: there's a free, task-specific model for almost anything, but discoverability and vetting are now the hard part. For teams, the lesson is to standardize on a short, evaluated shortlist rather than chase every new release.

AnythingLLM — Point an open or cloud model at your own documents and get a private, citable assistant. Built for teams that need retrieval over internal files without shipping data to a vendor.

OpenRouter — One API key for 500+ open and closed models, with live price and latency comparisons. Makes it trivial to route routine work to the cheapest model that clears your quality bar.

eve — Vercel's open-source (Apache-2.0) framework for building production AI agents — "Next.js for agents." Define each agent as a directory of files and get durable execution, sandboxed compute, and human-in-the-loop approvals out of the box.

Azeem Azhar, Exponential View · June 18, 2026 · 8 min

This is the economic autopsy behind the Big Story. Azhar traces how MIT-licensed models from Zhipu AI, DeepSeek, and Alibaba are collapsing the performance premium that closed labs have charged for two years — and argues that the real disruption isn't the benchmark scores; it's where the margin goes next. His case is concrete: when the model is free and self-hostable, value migrates to whoever controls the integration layer, the fine-tuning pipeline, and the operational infrastructure. That's a direct challenge to the per-token business model of every closed provider. Read it alongside this week's GLM-5.2 coverage to pressure-test your current vendor contracts before renewal season.