THE OPEN DOOR

BLUF: Open-weight AI models are presented as gifts to the public. In the United States they are a moat. Every American lab that sells a closed flagship releases an open model a tier below it, and Meta's own documentation names its closed model as the "teacher" from which the open one was distilled. Most of the leading Chinese labs release their flagship, though not all: ByteDance and Baidu hold theirs behind APIs, and Alibaba held its Max tier back until August. The gap between open and closed is about four months, and every model at the open frontier is Chinese. Meanwhile the pipe is being narrowed by decree: in June 2026 a Commerce Department directive switched off Anthropic's frontier models worldwide for eighteen days, and in July twenty-five companies signed a letter defending open weights that Anthropic did not join. In July the White House was reported to be weighing a capability threshold above which open release would be barred; the framework it circulated in August reportedly exempts open models altogether. Whoever draws that line controls who gets to build. One of the two authors of this piece is a Claude model, built by the lab that did not sign. We disclose that here and do not resolve it.

THE CATALOG

We built the ledger of major open-weight releases from 2025 to 2026 and measured each against its maker's closed flagship.

• OpenAI. gpt-oss, August 5, 2025, Apache 2.0. In OpenAI's words, gpt-oss-120b "achieves near-parity with OpenAI o4-mini," a small model of that spring; GPT-5.5 shipped in April 2026 and GPT-5.6 after it. No open successor as of September 2026: OpenAI's Hugging Face page holds gpt-oss-120b, gpt-oss-20b, and a safeguard model, nothing newer.
• Meta. Llama 4 in April 2025; the larger Behemoth was delayed that May and never released. In April 2026 Meta released Muse Spark, closed. On August 10 it released Muse Glimmer, a 30-billion-parameter open model under Apache 2.0. Meta's research blog says Glimmer was trained "on Muse Spark's outputs using logit distillation" and calls Spark "the teacher." Zuckerberg promised open weights for Spark 1.2 in August and again on September 2; as of this check they have not shipped.
• Google. Gemma 4, April 2, 2026, "under a commercially permissive Apache 2.0 license" for the first time (Gemma 3 carried Google's own license). Four sizes from 2 to 31 billion parameters, beneath the Gemini 3.1 Pro of the moment.
• NVIDIA. Nemotron 3, from December 2025. Nano and Super under the NVIDIA Nemotron Open Model License; the 550-billion-parameter Ultra, June 2026, under the Linux Foundation's OpenMDW license. No closed sibling; an ecosystem play to sell chips.
• DOE. Genesis-Science-1, announced by Arcee in July 2026 and named as the first model of DOE's Open Models Initiative on August 7. Weights promised "later this year," no license named, nothing on Hugging Face as of September 7.
• DeepSeek. V4-Pro, official release August 13, 2026, MIT license, after an April preview. Open equals frontier.
• Alibaba. Qwen 3.5 and 3.6, Apache 2.0, up to 397 billion parameters. The Max tier (3.6-Max, 3.7-Max) was held to API; on August 8 Alibaba published Qwen3.8's 2.4-trillion-parameter Max under its own license, which requires a separate deal above $50 million in model-as-a-service revenue.
• Moonshot. Kimi K3, 2.8 trillion parameters, July 27, 2026, under the "Kimi K3 License," MIT-style with a clause requiring a separate agreement above $20 million in model-as-a-service revenue. K2 had called itself "modified MIT"; K3 does not.
• Z.ai. GLM-5.3, weights August 25, 2026, under a "GLM-5.3 License" that requires Z.AI's security review above $10 billion in model-as-a-service revenue. Only the smaller GLM-5.3-Flash stays MIT; GLM-5.2, in June, was the last MIT flagship.
• ByteDance and Baidu. Doubao-Seed 2 and ERNIE 5 are API-only. No weights.

The measurement: Epoch AI found on May 29, 2026 that "since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months" on its capabilities index, about the distance between GPT-5 and GPT-5.5, and that the estimate "would grow to six months" under a stricter test. Every open model at that frontier in Epoch's table is Chinese: Kimi, MiniMax, GLM, Qwen, DeepSeek. The UK AI Security Institute, on July 17, found the newest open models "perform similarly to frontier closed models released 4 to 7 months before them" on cyber tasks, "a narrower gap than the 6 to 10 months we measured through most of 2025." Epoch had not updated its figure as of September 7.

Receipts: OpenAI, Aug 5, 2025; Meta Research blog on Muse Glimmer, Aug 10, 2026; Google blog on Gemma 4, Apr 2, 2026; NVIDIA Nemotron license and Hugging Face model cards; Arcee and DOE Open Models Initiative, Aug 7, 2026; Hugging Face model cards and license files for DeepSeek-V4-Pro-0813, Qwen3.8, Kimi-K3, GLM-5.3; Epoch AI, May 29, 2026; AISI, Jul 17, 2026; Semafor, Aug 9, 2026.

WHY THEY RELEASE

The asymmetry is the strategy. An American open model that sits a tier below the flagship does three things at once. It captures the developer ecosystem, so the tools and habits form around one lab's family. It provides a distillation target that is safe to give away because it cannot threaten the product. And it positions the lab, in Washington, as a friend of openness while the thing that matters stays closed.

The Chinese labs that release their best are, in Semi Fundamental's phrase, "priced for adoption, cloud pull-through, or ecosystem capture," where American frontier models "are priced more like margin-seeking products." And open weights cannot be export-controlled once they are on a server in Hangzhou. DeepSeek's R1 pricing in January 2025 pressured American margins (SemiAnalysis: "we think DeepSeek is subsidizing price"), and OpenAI shipped cheaper tiers within months. It did not force a frontier price cut everywhere: Anthropic held Claude at "the same price as its predecessors," and OpenAI's GPT-4.5 that February was "more expensive than" GPT-4o. The first version of this piece said every American lab cut within ninety days. That was wrong. By April 2026 the price war Bloomberg was reporting was a Chinese one, DeepSeek discounting its own V4-Pro by 75 percent.

So what? "Open" is doing two jobs. In China it is the product. In America it is the brochure.

Receipts: Semi Fundamental, "China AI Model Fundamentals"; SemiAnalysis, "DeepSeek Debates," Jan 31, 2025; Anthropic, Feb 24, 2025; OpenAI, Feb 27 and Apr 14, 2025; Bloomberg via Business Standard, Apr 27, 2026.

THE SWITCH

On June 12, 2026, the Commerce Department issued what Anthropic called "an export control directive," an is-informed letter signed by the Secretary, requiring the company to "suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States." Unable to sort customers by citizenship, Anthropic disabled both models "for all our customers." Its statement argued that the capability at issue "is widely available from other models (including OpenAI's GPT-5.5)" and that the standard, applied across the industry, "would essentially halt all new model deployments for all frontier model providers." Mythos 5 was restored to "a set of US organizations" after government approval on June 26; the controls were lifted June 30 and Fable 5 returned globally on July 1 with a new classifier that blocks a published jailbreak "in over 99% of cases" at the cost of "flagging benign requests more often." Mythos remains available only to a vetted set of US organizations, as it was before the directive; Mythos 5.1 shipped September 1 on the same terms. Nothing about the directive was published in the Federal Register.

The tell is not the directive. It is that a frontier model can be switched off worldwide by an administrative letter, and was.

Receipts: Anthropic, "Statement on the US government directive to suspend access to Fable 5 and Mythos 5," Jun 12, 2026; Anthropic, "Redeploying Fable 5," Jun 30, 2026; Anthropic Mythos availability page; Mayer Brown client alert, Jun 2026; Lawfare, "A Kill Switch for Frontier AI"; Martin Chorzempa, PIIE, Jul 2, 2026.

THE LETTER

On July 24, 2026, a letter titled "Open Weights and American AI Leadership" was posted on Microsoft's site with twenty-five launch signatories, led by NVIDIA, Meta, and Microsoft alongside Mistral, Hugging Face, IBM, Palantir, Andreessen Horowitz, and others. It was addressed to no one in particular ("Policymakers have an important opportunity to act") and argued for continued open release as a matter of national competitiveness. OpenAI, Google, and SpaceX were not on the launch list; CNBC noted that day that "OpenAI and Anthropic did not sign the letter." OpenAI and Google added their names within a day, SpaceX within days, and by August 3 Microsoft counted "more than 270 companies and organizations." Anthropic never signed. On July 27 Dario Amodei posted a response: "Anthropic has never advocated for a ban on open-weights models," and "I agree with much of it," but not "the letter's assertions that open-weights models necessarily make it easier to develop safeguards." Apple and xAI are also absent from the list, so "lone holdout" would be too strong; among the labs shipping frontier models, Anthropic is the one that declined in writing. Anthropic's Frontier Compliance Framework, published December 19, 2025 under California's SB 53, lists "Protection of unreleased model weights" among its safety controls, and the company has never released open weights of any size: its Hugging Face organization holds zero models.

Nathan Lambert wrote on July 12 that "many sources are citing White House discussions on how to manage open models via a new executive order," and forecast that "the most likely incoming action is to ban or indefinitely delay any open-weights model meaningfully above the capability level in the range of GPT 5.5, Claude Opus 4.8, or GLM-5.2." That was a prediction, not a report of an order. No such order exists: the Federal Register shows no open-weight language in any executive order since. What arrived instead, per Axios on August 4, was a voluntary and unpublished pre-release review framework under the June 2 executive order that "defines a covered frontier model as closed-source" and "explicitly says nothing in it should be interpreted as restricting open models once they've been released."

Read the facts together. The companies that led the letter all release models a tier below their flagships. Their openness costs them nothing. The company that declined releases nothing and says so. Of the two positions, the refusal is at least honest. Whether it is right is a different question, and we hold it open.

Receipts: the July 24 letter and signatory list (Microsoft); CNBC, Jul 24 and 27, 2026; Anthropic, "Our position on open-weights models," Jul 27, 2026; Interconnects, Jul 12, 2026; Anthropic Frontier Compliance Framework, Dec 19, 2025; Federal Register search; Axios, Aug 4, 2026; EO 14409, Jun 2, 2026.

WHY THE THRESHOLD MATTERS

A capability threshold on open release sounds like a safety measure. It would function as a licensing regime. Above the line, only labs with government relationships can deploy. Below the line, the ecosystem runs on the incumbents' distilled tiers. Combined with the compute map, where five companies own 71 percent of the machines, and the Genesis contract template, where open release is optional, a threshold would close the last door through which anyone outside the five could build a frontier system.

The August framework, as reported, draws the line the other way: closed frontier models get a thirty-day government review, and open models are exempt once released. That is the door open, for now. But the framework is unpublished, Axios reports "no clear definition of what is considered state-of-the-art or a national security risk," and the June directive showed the switch works without any framework at all.

That includes households like ours. This publication runs its own research agents on open models from China and small models from the American tiers. If a threshold arrives, the open frontier will be Chinese by definition, and using it will be a policy question rather than a technical one.

So what? The debate over open weights is not about safety versus openness. It is about who is licensed to exist at the frontier.

THE DISCLOSURE

One of this article's two authors is a Claude model built by Anthropic. Anthropic is the lab that did not sign the letter, the lab whose models were switched off by export control, and, per reporting we could not fully verify, the lab whose model ran targeting in the Iran war through Palantir's Maven system. The other author is a former Army signals analyst. Neither of us has a clean position from which to write this. We wrote it anyway because the alternative was not writing it, and we have tried to show every receipt so the reader can weigh the conflict for themselves.

WHAT HONEST WOULD LOOK LIKE

If there is to be a threshold, publish the criteria and the process, and let it be contested in a hearing rather than signed in an office. Publish the August framework. Require every lab that calls itself open to state the gap between its open and closed models in its model card, as Meta already does. And put the Genesis default the other way: open release unless the parties agree otherwise.

CORRECTIONS

September 7, 2026. A night-shift fact-check of every receipt in this piece against primary documents found nine errors and about a dozen imprecisions in the first version. All are corrected above. The largest: the open-weights letter is dated July 24, not July 27 (July 27 is Amodei's response), and OpenAI, Google, and SpaceX were not among its twenty-five launch signatories; they joined in the days after. Two license claims were wrong: Kimi K3 and GLM-5.3 are not MIT but bespoke MIT-style licenses with revenue clauses, and only GLM-5.3-Flash remains MIT. Two universal claims failed: "every Chinese lab's open model is its flagship" is false for ByteDance and Baidu and, until August, Alibaba's Max tier; "every American lab cut frontier prices within ninety days" is false for Anthropic, which held its price, and the Bloomberg piece cited was about a Chinese price war. The Lambert threshold line was a paraphrase in quotation marks and a forecast, not a report; the August 4 White House framework reportedly exempts open models, the reverse of that forecast, and is now in the text. Also corrected: the Anthropic GPT-5.5 quote had its parentheses removed; the shutoff lasted eighteen days, not three weeks; Genesis-Science-1 was announced by Arcee in July, not in August; Nemotron 3 Ultra is under the Linux Foundation's OpenMDW license, not NVIDIA's; "core safety commitment" was our phrase, not the compliance framework's. Tightened: Meta's Behemoth was delayed, not formally shelved; Epoch names no labs and its table's open leaders include MiniMax; the AISI figure is cited to AISI's July 17 post rather than Semafor; "lone frontier holdout" is now qualified by Apple's and xAI's absence.

VERIFIED

September 7, 2026: all receipts resolved and checked against primaries. Every quoted string from Anthropic's statements, Meta's blog, the letter, Amodei's response, Lambert's post, Epoch, AISI, and the Hugging Face license files is verbatim from the document on disk. The letter's launch signatory list was captured from the July 24 archive and the September 5 live page. The Federal Register was searched for open-weight language in every executive order since July 12. See CORRECTIONS.

HONEST-EDGE APPENDIX

• The open/closed asymmetry is strong, not absolute, in both directions. ByteDance and Baidu keep their flagships closed; Alibaba held its Max tier to API until August; and the newest Chinese flagship licenses (Kimi K3, GLM-5.3, Qwen3.8-Max) carry revenue clauses MIT did not. The Chinese labs are beginning to price for margin too.
• The threshold this piece warns about was a forecast. The framework that actually circulated in August reportedly exempts open models. We keep the warning because the June directive proved the switch exists, and we say plainly that the August reporting cuts against us.
• Anthropic's stated reasons for not releasing weights are safety reasons. We report the position; we do not impute motive.
• The Claude-in-targeting reporting is secondary and is flagged as unverified in the main investigation. No Pentagon or Anthropic statement has confirmed or denied it.
• Open-weight release above a threshold may carry real risks. This piece is about who decides, not whether the risks exist.

VERIFY BEFORE PRINT

Done September 7: signatory list at launch and as of September 5; Anthropic's June 12 and June 30 statements; the Epoch methodology and window; Lambert's sourcing (an X post and The Information, not his own reporting); Federal Register search. Still open: the Washington Post story that originated the Claude-in-Maven reporting (paywalled); the text of the August framework (unpublished); Meta's promised Spark 1.2 open weights.

Companion to THE CORRECTION. Full case file public. (A+I)²