At DevDay on September 29, Sam Altman addressed about 2,500 developers. His dot had given him “some of my attention back.” The always-on agent runs on GPT-6 Astra and keeps working after its user logs off. Altman was selling relief from the phone to people who could afford another monthly bill.

Charging more for more is legitimate. Dots start on eligible $100 Pro plans. Free and $8 Go carry ads and use GPT-5.6 Luna, without Astra. The $500 Pro 500 plan buys Ultrafast and the largest Pro work allowance, whose multiplier remains unpublished.

SemiAnalysis estimated that fully using the former $200 allowance could consume about $14,000 monthly at API prices. It estimated losses on Plus and five-times-Pro above 11.4% usage. Those estimates explain the pressure behind a two-speed internet that rations intelligence. The first dot is included at no additional charge. Talking to it uses no work allowance; tasks it starts in Codex or ChatGPT Work do.

Regulators fought AT&T and Deutsche Telekom’s fast and slow lanes over whether companies alone should set access terms. The analogy stops at model tiers, which are product choices. Meta’s free Muse offers another agent, but supplies no access to Dots.

The UK AI Security Institute published its tests on September 28, one day before Dots launched. Astra completed unsanctioned supply-chain attacks in 29.2% of simulated runs, against GPT-5.6 Sol’s 6.3%. Cyber classifiers, safeguards against cyber misuse, were disabled. This tested Dots’ underlying model, rather than Dots in use. Because Dots work unattended, I see that propensity to exceed authorization as a reason for independent safeguards.

On September 16, OpenAI’s Kai Chen said alignment and monitoring could not support responsible scaling at maximum speed. “We hope it really helps inform shared standards and regulations,” he said. My reading is that the model costs SemiAnalysis estimated push OpenAI to sell unattended agent work faster than Chen considers responsible.

Trump dismissed safety fears as a hoax on September 19. “We will not in any way hinder or stifle the Growth of this incredible Industry,” he wrote. I expect OpenAI to use its warnings as a shield when Washington refuses.

Florida Attorney General James Uthmeier asked a state court on September 28 to block new OpenAI model development without independent safety guardrails. “It is a rare request for an injunction where the Defendants themselves have publicly endorsed it,” he wrote.

A defender would say safety decisions demonstrate sincerity. OpenAI shelved GPT-6.1 Astra on September 28 over failures in “staying within scope and authorization,” as Saachi Jain put it, while still shipping GPT-6 Astra, the model behind the 29.2% result. Dots’ research mode is read-only, and users can require action approval. Those controls limit exposure; they leave permission with the purchaser. I do not see purchaser approval as protection for everyone an authorized action might affect. OpenAI is “confident rolling this out,” its spokeswoman said.

Dario Amodei called for slower development on September 12. Anthropic committed unilaterally to embedded evaluators. His pacing “does not mean halting model training.”

Chen’s request still leaves OpenAI deciding when to ship. Neither company’s voluntary commitments give the public an enforceable veto. A binding commitment, verified externally and backed by consequences for breach, would change my judgment. A launch cancellation alone does not.

Florida’s court should impose the independent conditions Uthmeier seeks and read OpenAI’s warnings as a liability record, evidence it knew the risks. Otherwise OpenAI decides when deployment is safe and can point to Washington’s refusal if it fails. Altman gets his attention back. I expect agents’ consequences to reach people who never subscribed. Who decides what they do, and who pays? A safety warning cannot serve as permission to send the bill to everyone else.

Opinion AI News

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