The Signal — September 28, 2026

Dario Amodei, who spent September asking the industry to slow down and got mocked by the president as a "perfect little angel" for it, was due at the White House for a private dinner with Trump late on September 27. Elsewhere, companies are moving work to cheaper open-weight models and talking about them more often on earnings calls, and the Department of Homeland Security is preparing a tool that suggests what to black out of public-records releases.

Trump invited Amodei to dinner two days before a White House AI meeting with tech CEOs

President Trump planned a private dinner with Anthropic CEO Dario Amodei at the White House on the night of September 27, Axios first reported. Trump confirmed the plan to reporters that evening. According to Axios and CNBC, it would be the first one-on-one meeting between the two. Amodei had missed the state dinner for Xi Jinping on September 24 because of a scheduling conflict, and Trump then invited him personally, both outlets reported, citing people familiar with the plan. A White House official's statement to CNBC did not mention Anthropic: "President Trump has been clear: America will lead the world in Super Intelligence, while protecting American consumers."

The invitation came after a rough stretch between the two. Earlier this month Amodei called for the industry to slow the pace of capability gains and asked for more government oversight, and Trump responded on Truth Social that the only guardrail AI needs is a "STRONG AND SMART (High IQ!) PRESIDENT," adding that Amodei was "pretending to be a 'perfect little angel,'" as CNBC quoted him. On September 25 a federal appeals court upheld the Pentagon's designation of Anthropic as a supply-chain risk, which we covered on September 26, and Defense Secretary Pete Hegseth celebrated the ruling online. CNBC also reports that Commerce Secretary Howard Lutnick and White House science adviser Michael Kratsios have spoken warmly of the company in recent weeks, so the administration has not been speaking with one voice.

No readout of the dinner had been published by the morning of September 28, and the White House and Anthropic both declined to comment to the New York Times beforehand. Axios describes the invitation as a sign of thawing relations, which is its reading rather than anything either side has said. Trump and House Speaker Mike Johnson are due to meet AI company leaders at the White House on September 29, according to Axios, as pressure builds in Washington to regulate the frontier labs.

Sources: Axios · CNBC · New York Times

Open-weight models ran most of Vercel's gateway tokens in August, but took a smaller share of spending

Executives mentioned "open-weight" or "open-source" AI models six times as often on earnings calls and at investor conferences in August and September as in the same months last year, the Financial Times reported on September 27, citing data from the research platform AlphaSense. The mentions were not limited to tech companies: PNC Financial Services, the logistics company C.H. Robinson and Siemens have all discussed using such models recently. Open-weight models are ones whose parameters are published, so a company can run them on its own hardware or through a cheaper host instead of paying a frontier lab per token.

The clearest reason given is cost. Tinder's chief technology officer, Vinay Kuruvila, told the FT that the dating app's AI spending ran at about $1 million a year in January and $10 million a year by July. "I don't want another 10X increase," he said, adding that frontier models like OpenAI's Astra and Anthropic's Fable "already have enough intelligence for 90% of the tasks we're trying to do." Tinder now sends some queries from non-technical users to open-weight models.

Vercel's monthly production index measures usage on its AI Gateway, which routes requests from production apps to many model providers. In August, open-weight models handled 56 percent of the tokens passing through it, up from 7 percent in December, but accounted for only 14 percent of the estimated spending. Average price per token on the gateway fell 23.2 percent that month. Frontier labs still collect most of the money on the gateway. Anthropic kept 64 percent of gateway spending in August, and teams that cut back on its most expensive model, Fable 5, moved more of that work to the cheaper Opus 5 than to any other model. Vercel's numbers cover traffic through one gateway, not the whole market. Its open-weight category was recently widened, and spending is estimated from list prices. The AlphaSense count measures how often executives bring the models up, not how much work they actually run on them.

Sources: Vercel · Financial Times · PYMNTS

Homeland Security is preparing an AI tool that suggests what to redact from FOIA releases

The Department of Homeland Security says it will use AI to "revolutionize" how it handles redactions in Freedom of Information Act releases, according to a September 26 Washington Post analysis by Nate Jones, the paper's FOIA director. An entry from Customs and Border Protection in the agency's own AI use-case inventory describes a tool called RedactAI. It uses Google's Vertex AI and other Google Cloud services to find passages that might be withheld, such as personal information or material covered by FOIA's exemptions. It groups the suggestions by type and presents them to staff for review; each redaction requires their approval. DHS lists it as a generative AI system that has not yet been deployed.

The inventory also classifies RedactAI as not "high-impact," which, as FedScoop has noted, means it does not have to meet the stricter risk-management requirements federal rules set for high-impact uses. The workload behind the project is large. FedScoop reported in July that DHS received more than a million FOIA requests in fiscal 2025 and still ended the year with 245,572 pending.

Transparency advocates worry that training a model on an agency's past answers could reproduce its habit of withholding too much. "AI trained on existing responses will reflect that," Abigail Kunkler of the Electronic Privacy Information Center told FedScoop, referring to agencies' habit of withholding more than they need to. "Using AI to automate redactions almost guarantees that DHS will over-redact." Cody Venzke of the ACLU said redaction is "a legal question with significant nuance." DHS has not published accuracy figures or any measure of how much faster the tool is. Under the planned process, every suggestion goes to a human reviewer, so how much it changes what the public actually receives will depend on how closely those reviewers check it.

Sources: DHS AI Use Case Inventory · FedScoop

On the Editor's Desk

The Information reported that China's industry ministry told ByteDance, Alibaba and others it intends to approve their purchases of Nvidia's new RTX Pro 5500 chip. Reuters said it could not verify the report, and the only sources are two unnamed people, so we are waiting for something on the record.

Bill Gates told NBC's Meet the Press that a kill switch for AI is "not enough" and that monitoring should be required by law. We left it out because it is an argument rather than a policy change, though it lands the same weekend as the dinner above and two days before the White House meeting. We also held several weekend follow-ups on the OpenAI agent incidents we covered on September 26, because they add counts from unnamed sources but no new documents from the companies.