OpenAI launches $500 ChatGPT tier; Anthropic lists $518 billion in commitments

Morning, humans.

Let’s take a closer look at what AI work actually costs, and who pays for it.

OpenAI's Pro 500 costs $500 a month. Dots connect to more than 4,000 apps, though the subscription ladder needs a look before you hand over the keys.

Anthropic's prospectus lists at least $518 billion in computing commitments. About 80% is non-cancelable or payable regardless of usage; the payment schedule deserves its own stop.

Our Paperclip test recorded 91,630 tokens for “local model OK.” The Paperclip Compendium, free for subscribers, walks you through agent cost controls before a tiny reply becomes a budgeting exercise.

Stay curious,

Marcus Schuler

OpenAI launched Dots and the $500-a-month Pro 500 ChatGPT plan at DevDay on Tuesday.

A dot is an always-on agent with its own cloud computer and browser. Access starts at $100 monthly for Pro or $125 per Business Premium seat; free and $8 Go users get none.

OpenAI's ladder tops out at $500 monthly. From Oct. 30, Pro 200's ChatGPT Work and Codex allowance drops from 20 times to 10 times the Plus allowance; GPT-6 Pro messages fall from 200 to 100 weekly. Pro 500's usage multiplier remains unpublished.

Why This Matters:

• ChatGPT’s $100 Pro plan includes Dots as of Tuesday, although subscribers in the EU, Switzerland and UK cannot use them.
• From Oct. 30, Pro 200’s Work and Codex allowance falls from 20 to 10 times Plus; existing subscribers get a $2,500 credit that expires at year’s end.

Anthropic’s confidential IPO prospectus lists at least $518 billion in computing commitments over roughly a decade.

About 80% of that amount is non-cancelable or payable regardless of usage. Anthropic’s 2025 revenue was about $4.59 billion. Those terms make weaker demand a payment problem.

Facts Are Friends traces how Harvey's reported margin collapse forced the legal AI company to own more of its model stack and may push its law-firm customers toward the same choice.

Harvey's gross margin reportedly fell from about 50 percent in January 2026 to about negative 50 percent in June 2026 as customer token use rose 20-fold during the first six months of 2026. Its shift toward an open-weight model reframes model ownership as control over price and supply.

• Implicator published the Paperclip Compendium, a guide to running AI agents like a company, free for subscribers, with a 23-entry troubleshooting guide.
• OpenAI shelved GPT-6.1 Astra, expected in ChatGPT and Codex in October, after tests showed regressions in staying within scope, seeking authorization and honestly reporting actions.
• Samsung Electronics' $500 million investment in Helix backs an infrastructure company led by former AWS chief Adam Selipsky, with no Samsung supply contract announced.
• EliseAI, with $200 million ARR as of June 2026, raised $350 million in a round led by Andreessen Horowitz and Bessemer Venture Partners.
• Bitdeer has booked over 70% of its roughly 21.7-megawatt Malaysia site before first-quarter 2027 energization.
• Meta describes Muse for Small Business, which connects to QuickBooks, Shopify and Stripe, as “free for most of what people need.”

Turn a new rule into an action clock.

A regulatory document can hide the date that changes your workload. Convert the text into a dated operating plan while keeping legal judgment with counsel.

Your raw input:

The prompt:

Why this works:

Requiring quoted passages keeps the analysis tied to the source. The action clock separates operational preparation from legal conclusions and exposes where delay carries a cost.

What to use:

Use Claude for long regulatory text or ChatGPT for a shorter notice. Use a smaller model only to format counsel-approved conclusions.

HiddenLayer sells software that inventories, tests and monitors AI models and agents for enterprise security teams. Its position is arguable because AI security may become a distinct budget category, or disappear inside the broader platforms already sold to chief information security officers.

Founders:

Christopher “Tito” Sestito, Tanner Burns and Jim Ballard founded HiddenLayer in March 2022 after working together at Cylance, where Sestito led threat research and the team handled an attack against Cylance’s machine-learning malware detector.

Product:

The platform scans models and their software supply chains, runs adversarial tests, and monitors models and agents in production for manipulation, malicious tools and unauthorized actions. Customer names are undisclosed. HiddenLayer said annual recurring revenue grew more than tenfold during the year ended September 2, 2026, while it added more than 50 platform customers and protected an unnamed frontier-model provider serving more than 700 million weekly users.

Financing:

HiddenLayer has raised about $156 million across disclosed rounds: a $6 million seed round in July 2022, a $50 million Series A in September 2023, and a $100 million Series B led by Delta-v Capital on September 2, 2026. The latest round did not disclose a valuation.

The risk:

Palo Alto Networks has assembled model scanning, red teaming and runtime controls inside Prisma AIRS through acquisitions including Protect AI. If buyers prefer AI protection as a feature of a larger security suite, HiddenLayer’s independent platform could lose distribution and pricing power.

Credit: Midjourney

Prompt: Post-apocalyptic industrial dieselpunk explorers, shot on Cinestill 800T, 35mm, natural film grain, low key lighting, deep shadows, candid unstyled moment

Trump’s Sept. 29 executive order directs agencies to replace AI with “Super Intelligence” in non-statutory documents to the maximum extent permitted by law.

Our take: Trump gives his science adviser 60 days to propose legislative language on the definition of Super Intelligence. The order already gives that name to everything covered by the existing AI definition. That is a generous promotion for a technology that did not need to pass a new capability test.

“I think it’s morally binding,” Trump said of the voluntary accord. Musk calls the approach “just generally grading each other’s homework.” The homework now has a grander title on the cover, which should please anyone who mistakes presentation for oversight. Apparently, the guardrails are printed on the letterhead.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai