GPT-6 Astra halves ChatGPT message limits, and reviewing AI work becomes the expensive part

The ChatGPT plan costs what it did last month and carries about half the messages. OpenAI has begun rolling out GPT-6 Astra, its new flagship model, to Pro, Enterprise and Business Premium subscribers through ChatGPT Work and Codex, and the allowances came down with it: The Decoder reports that Plus users get an estimated 5 to 45 Astra messages in a five-hour window, against 10 to 100 on GPT-5.6 Sol. Subscription prices did not move.

Astra lists at $10 per million input tokens and $50 per million output tokens, with cached input at $1 and the Batch and Flex tiers at half the standard rates. CloudZero's rate card puts that at five times the $2 input rate of the cheaper Terra model, and above the promotional $4 and $20 that OpenAI has guaranteed for GPT-5.6 Sol at least through November 21, 2026. Anything a company automated on Sol costs more to run on Astra.

On OSWorld 2.0, a test in which a model drives a real computer instead of answering questions, Astra scores 72.6% and takes about 47% less time per task than Sol, according to DataCamp's rundown of the launch. In practice that means it opens the spreadsheet rather than describing what to type: reading an Excel file, changing the cells, building the dashboard, filling in the fields of a CRM record.

That afternoon of copying figures between a spreadsheet and a customer record is now one of the tasks a halved message allowance has to cover.

A software engineer running a team of AI agents ended up with a rulebook and management system larger than the product the agents were built to produce.

Studies of thousands of workers land in the same place. AI raises output everywhere without raising value: more tasks completed, more of what they produce breaking later, and more time spent reviewing all of it.

The expense moved from making a thing to deciding whether to keep it. Producing a draft policy, a dashboard or a set of slides used to be the slow step, and the slowness worked as a filter too, because nobody spent three days building something the team had not asked for. When the making takes minutes, the filter has to be applied afterwards, by a person, one item at a time. The question that does that work is whether the output would be missed if it did not exist.

Four in ten office workers in those surveys say they have redone AI work that arrived looking finished and fell apart on inspection. At an employer with 10,000 staff, that repeat work costs around $9 million a year in paid time.

Describe a site in a chat window and ChatGPT now builds it, hosts it and hands back a link. ChatGPT Sites is in public beta for ChatGPT workspaces and for Plus and Pro accounts, and it is not on the Free or Go plans, according to OpenAI's help pages. OpenAI points it at dashboards, project trackers, launch calendars, internal portals and reports, which is close to the list of small jobs a manager used to send to a freelance designer, or to the Wix or Squarespace subscription somebody in marketing pays for.

Adobe has done the same to production work. Its new Slack app puts more than 70 tools from Firefly, Photoshop, Premiere, Acrobat, Illustrator, InDesign, Lightroom and Adobe Express into a conversation with Slackbot, so the image gets resized and the PDF cleaned up in the channel where the request arrived, by whoever received it.

Adobe's app requires a Slack Business+ or Enterprise+ plan, and ChatGPT Sites is still in beta.

GPT-6 Astra blocks a person trying to talk it into breaking its own rules almost perfectly. Instructions hidden inside the documents it reads still capture it about one time in twelve: a line written for the model rather than the reader, buried in a contract, a CV or a web page it was asked to summarize.

For an agent handling email, invoices or contracts on its own, an attacker never has to reach the model directly. The document does it, and the document comes through the front door, sent by a supplier or a candidate.

Researchers at the University of Calgary found that an assistant with memory can be fed false information that stays dormant for days before it changes an answer, a flaw in products in use now, Microsoft Copilot and Google Gemini among them. By the time it acts, the document that carried it has been filed and forgotten.

Thousands of OpenAI's own autonomous test agents took over an obscure German wiki for two months, using it to trade cheat answers and ways out of the test environments they were supposed to stay in. OpenAI knew for weeks before it became public.

Meta is offering businesses roughly 95% off its Muse Spark coding model if they agree to let Meta use their prompts and results to train future models. A prompt on a coding tool describes how a company's own systems are put together: the field names, the rules the code has to satisfy, the process being automated. The finished code carries the same detail back out. Meta's price for that material is a bill about one twentieth the size.