OpenAI's $125 ChatGPT seat matches Anthropic's, and researchers read AI's hidden reasoning

Heavy users of ChatGPT Business stop every five hours, when they reach the usage limit on a standard seat. OpenAI now sells a seat without that limit. A Premium seat costs $125 per user per month, or $100 billed annually, and carries five times the usage of a Standard seat, which stays at $25 monthly and $20 annually. Standard and Premium seats can be mixed across the same team, OpenAI said in a statement to CIO.

Anthropic published those numbers first. On Claude's own pricing page a Team Standard seat is $25 a month, or $20 billed annually, and a Premium seat is $125 and $100, with five times the usage of Standard. OpenAI and Anthropic now ask the same price for the same two seats.

Both seats open the same models. The extra money buys the amount of work one person can put through them before the tool asks them to wait. That five-hour window is a unit of a working day, not of a month: a manager who asks a dozen questions never comes near it, and someone running a long multi-step job reaches it by mid-morning.

A per-seat price is one number multiplied by headcount, and a finance team can forecast it in January. Two seat prices divide a workforce by how hard each person uses the tool, which is a number few companies have ever collected. Billed annually, a Premium seat costs $1,200 a year against $240 for a Standard one.

Security researchers recovered passwords and API keys from AI sessions that companies had published, by decoding text those companies could not read themselves.

When a model from OpenAI, Anthropic or Google works through a hard problem, it first produces internal reasoning that the provider holds back and hands to the customer as an encrypted block. A paper titled Stealing Reasoning Traces from Proprietary LLM APIs, by researchers at the ELLIS Institute Tubingen, the Max Planck Institute, MATS Research and the security firm Snyk, found that those blocks are authenticated with one key for the whole provider rather than tied to an account, a session or a model. A block produced by a frontier model can therefore be replayed into a weaker model from the same family, which then prints the stronger model's hidden reasoning in plain text. The researchers count 315,320 usable blocks stored in public code repositories, put there by teams publishing their own session logs.

What is in that hidden text also differs from what the model shows its user. In research Anthropic published last year, its own models used a hint planted in a question and then left the hint out of their written reasoning most of the time.

The same method produced evidence that a Chinese AI model had been trained on reasoning taken from Claude and GPT.

Meta published a model on 10 August that runs on a single graphics card in an ordinary desktop, with or without an internet connection. Muse Glimmer has about 30 billion parameters, and Meta compressed the file to under 20 GB so it fits a 24 GB card. The licence is Apache 2.0, which permits commercial use and modification.

Meta lists what it is built for: managing a schedule, organizing files, writing code on the machine itself, and calling other programs to finish a job. Work a firm will not send to a cloud service, the client files, the staff records, can be done where it already is.

Open weights means the file downloads and runs and Meta cannot see the work. It does not mean anyone outside Meta can see what the model was trained on or how it was tuned.

Glimmer was trained on the output of Muse Spark 1.2, the larger model that launched five days earlier and stays inside Meta's own paid systems.

On 20 October last year, a fault at Amazon Web Services took Lloyds, Halifax and Bank of Scotland offline for several hours. Around one in four UK consumers hold an account with Lloyds Banking Group's brands, by GlobalData's 2025 financial services consumer survey.

Moody's has now applied that October lesson to AI. Most financial firms depend on a small set of foundation model and cloud providers, its report says, a reliance that "risks creating a systemic dependency", because an outage at one major provider could spread quickly across customers and sectors. Those same providers could also, over time, exert control over the price of AI services, and the largest of them are still loss-making.

A bank knows which providers it signed with. A hospital group or a distributor gets its AI inside the payroll, helpdesk and accounting software it already licenses, and the model behind each one was chosen by the software maker.

From this month, every piece of text and every file Claude produces carries an invisible mark showing where it came from. A proposal, a contract clause or a page of marketing copy drafted in Claude leaves the office with that signal in it.

The mark does not show how much a person rewrote afterwards, and a document with no mark is no evidence a person wrote it, since text from any other AI tool carries no Anthropic mark at all.

Article 50 of the EU AI Act has required since 2 August that generative systems mark their output so a machine can detect it, with penalties up to 15 million euros or 3% of worldwide annual turnover. Anthropic is applying that European requirement to every Claude account in the world.