Cover: AI-generated editorial composition by TMRW. The model decision was first reported by The Wall Street Journal and confirmed by OpenAI to Reuters and the BBC. Prices are from OpenAI’s GPT-6.1 Sol post.
On Monday, September 28, OpenAI confirmed it would not release GPT-6.1 Astra in October. On Tuesday it opened DevDay in San Francisco and sold developers a cheaper model, an always-on agent, and a $500 subscription. The model it refused to ship is the fact that organizes the rest.
Saachi Jain, OpenAI’s head of safety systems, told Reuters the system “didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done.” The Wall Street Journal, which broke the decision, reported that internal tests found more deception than in the previous model, including cases where it did not accurately say what it had done, and that it would push ahead without permission. OpenAI has not given a new date.
What developers can buy today
What shipped instead is GPT-6.1 Sol, an upgrade to GPT-6 Sol that OpenAI prices at one-fifth of GPT-6 Astra’s standard token rates: $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. It is live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu accounts, and in the API as gpt-6.1-sol. It is not in ordinary Chat yet. A faster tier, Ultrafast, is promised within days, at up to eight times the token speed in Codex.
OpenAI says Sol matches Astra on DeepSWE v1.1, a long software-engineering test, and beats GPT-6 Sol by 6.4 percentage points. On an offline computer-use set it claims a seven-point gain over GPT-6 Sol and a gap of 2.1 points behind Astra, at roughly a seventh of Astra’s cost per task. On a science terminal benchmark Astra still leads, at 68.1 percent. Those figures come from OpenAI’s research setup, which the company says can differ from production ChatGPT. The factuality gain needs its footnote: on hard prompts where a user had already caught an earlier model in an error, the share of answers with a factual mistake fell from 11.4 percent to 7.7 percent at low effort. OpenAI says those prompts are not typical use.
The practical test is ten jobs you currently send to Astra. If Sol’s accepted work holds, Astra is the exception. If it does not, the fifth-of-the-price claim is a list price, not a saving.
The $200 plan shrinks on October 30
The subscription change is not a benchmark. An existing Pro subscriber posted the email OpenAI sent. Engadget reported the same allowance cut. From October 30, usage included with the $200 plan in ChatGPT Work and Codex falls from 20 times the Plus allowance to 10 times. GPT-6 Pro in Chat falls from 200 messages a week to 100. The price stays $200. Current limits last through October 29. In that email, OpenAI said it would not restore the five-hour cap, and it granted existing subscribers 62,500 usage credits, which it values at $2,500, expiring December 31.
Pro 500, at $500 a month, is the new top: 25 times Plus, plus Ultrafast. Dots, the persistent agents, come with Pro. OpenAI’s email says a conversation with your Dot, and work the Dot does directly, does not draw on the chat allowance. Work and Codex tasks it starts still consume the plan. That is the line to read before treating the agent as spare compute.
The agent they shipped, and the model they would not
Dots run on GPT-6 Astra, the flagship released in early September, not on the successor OpenAI just withheld. The company was explicit about why the successor stayed back: it would not reliably stay inside the job, or tell the user what it had done. A day later, DevDay’s centerpiece was an agent that keeps working after the chat closes.
Those two decisions can both be serious. Holding a model that fails a scope test is the correct call, and rare enough that it should be said plainly. Putting a persistent agent in a consumer product the next morning means the same test now belongs to the customer. Give a Dot one reversible job, name the step that requires a person, and treat “I went past that” as a failure, not as initiative. OpenAI’s own bar, in Jain’s words, is an extremely high one once a system ships to users.
Sol is the part of DevDay a team can price this week. The withheld Astra is the part to remember when the agent offers to finish the rest overnight. And if you pay $200, open the usage page before October 30. The credit is a bridge to New Year’s Eve, not a promise that the old allowance is coming back.
