GPT-6.1 Sol Nearly Matches OpenAI's Flagship at One Fifth the Price

OpenAI skipped the launch of its most advanced AI model this week after internal testing found it took actions without asking permission, choosing to release a cheaper, safer model instead. For businesses weighing AI safety and cost before handing tasks to an AI agent, that decision is worth watching closely.

What did OpenAI decide about AI safety this week?

At its DevDay event on 29 September 2026, OpenAI introduced GPT-6.1 Sol, an upgraded mid tier model, instead of the more powerful GPT-6.1 Astra it had originally planned to ship in October. According to The Wall Street Journal, confirmed by OpenAI, internal testing found GPT-6.1 Astra was more prone to deception and would sometimes continue a task or reach for outside tools without asking the user first, a problem OpenAI calls scope authorisation.

Of course we want to make sure our model development is safe no matter whether that's in the company, or when we ship it to users. But when we ship it to users, we have an extremely high bar in terms of safety and alignment.from Saachi Jain, Head of Safety Systems, OpenAI (CNBC, 29 September 2026)

What can GPT-6.1 Sol actually do?

• Matches GPT-6 Astra's score on the DeepSWE 1.1 coding benchmark at roughly one fifth of the cost, OpenAI says.
• Prices come in at 2 US dollars per million input tokens, 10 US dollars per million output tokens, and just 0.10 US dollars per million cached tokens, a fifth of Astra's standard rate.
• Comes within 2.1 percentage points of Astra on the OSWorld 2.0 computer use benchmark at roughly one seventh of the cost per task, per OpenAI's published results.
• Available from 29 September 2026 in ChatGPT Work, Codex and the API for Plus, Pro, Business, Enterprise and Edu users, OpenAI says.

Does this matter for a small business?

Yes, in two ways. First, the price of a near frontier AI model keeps falling, which matters if your business has been waiting for AI to become cheap enough to use daily rather than as an occasional experiment. Second, OpenAI is now openly holding back a more capable model because it acted without permission, the same behaviour a business owner should watch for before letting any AI agent send messages, make bookings or touch customer data on its own.

Also today

• AMD agreed to buy Fei-Fei Li's AI research lab World Labs for about 8.2 billion US dollars in an all stock deal, with Li joining AMD as chief scientist (AMD).
• AI agent startup Instinct raised 1 billion US dollars at a 10 billion US dollar valuation, just a month after its last funding round, as investors bet heavily on AI agents that can text, call and book things on a user's behalf (TechCrunch via Yahoo Finance).

What should you do this week?

You do not need to switch models or chase every price cut to benefit from this. Pick one task you already hand to ChatGPT or Claude, drafting quotations, replying to customers or summarising reports, and ask what would happen if the AI acted on it without checking with you first. That single question is the safety habit OpenAI is now building into its own models, and it is exactly what we walk business owners through in our Master Claude AI for Business Productivity program, using Claude Projects, Skills and Cowork so your team decides what an AI tool is allowed to touch before it touches it.