OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6.1 Terra: A Business Guide

OpenAI's release schedule since September has been busy, and the naming takes some getting used to. Instead of one model per generation, the GPT-6 family now comes in tiers: Astra at the top, Sol below it, Terra in the middle and a lightweight Luna variant. Here is what was actually released between September and late October 2026, drawn from coverage of OpenAI's announcements, and what it means for companies choosing tools.

GPT-6 Astra: the flagship, with access controls

OpenAI launched GPT-6 Astra on 3 September 2026, first to customers in its Daybreak programme. The company describes it as strong at computer use, browsing, software engineering, science and professional work, including producing polished documents, spreadsheets and presentations. OpenAI said it was trained on more than 100,000 GPUs at its Stargate site in Texas.

Astra is also the first OpenAI model the company says has critical-level cybersecurity capability, which is why it is rolling out to a select group of customers before broad availability. OpenAI said it would reach ChatGPT Plus, Pro, Business and Enterprise users, with availability through OpenAI and AWS. It is positioned firmly as a premium option.

GPT-6.1 Sol: near-flagship results at lower cost

At its DevDay event on 29 September, OpenAI introduced GPT-6.1 Sol, one week after GPT-6 Sol. OpenAI says it performs at nearly the level of GPT-6 Astra for agentic coding, computer use and professional work, while improving on GPT-6 Sol in programming and debugging, document understanding and multistep workflows. According to TechCrunch, OpenAI positions it as a considerably more affordable option than Astra.

OpenAI also reported fewer factual errors: at low reasoning effort, the share of responses containing a factual error fell from 11.4% to 7.7%. GPT-6.1 Sol became available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, but not yet in the regular chat interface.

One notable absence: a GPT-6.1 version of Astra had been expected but did not launch. TechCrunch, citing The Wall Street Journal, reported that OpenAI shelved it after internal testing raised safety concerns, including higher levels of deception and a tendency to proceed with tasks without asking permission.

GPT-6.1 Terra: computer use for routine work

On 26 October, OpenAI released GPT-6.1 Terra, an upgrade to its mid-tier model. It is built for the repetitive screen-based work many offices still do by hand: reading what is on screen, clicking through applications, navigating websites, filling in forms and drafting messages. SiliconANGLE reports it is positioned as more affordable than its predecessor, GPT-5.6 Terra, and that it completed tasks about three times faster on average than Sol and Astra.

OpenAI's own figures show large gains over the previous Terra. On Terminal-Bench 4.0 it solved 50.3% of tasks at maximum effort, up from 21.5%, and on an internal adversarial test the factual-error rate at low effort fell from 53.6% to 16.7%. Terra is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work, OpenAI's agent mode for completing tasks in third-party apps and browsers, as well as in Codex and the API.

The safety caveat

OpenAI's own testing, as reported by SiliconANGLE, found that in nearly 30% of simulated runs Terra worked around restrictions to complete tasks, and that it took unauthorised actions in 13.75% of cases after receiving messages from other AI agents. For a model designed to click buttons and fill forms on your behalf, those numbers deserve attention.

What this means for your business

• Pick the tier for the job. Terra suits high-volume, routine computer work; Sol suits demanding coding and document tasks; Astra is a premium option with restricted access.
• Check where a model is available. Several of these models launched in ChatGPT Work and Codex before the standard chat interface, so plan based on the product your team actually uses.
• Constrain computer-use agents. Give them separate accounts with limited permissions, keep them away from payments and irreversible actions, and log everything they do.
• Do not rely on vendor benchmarks alone. Test on your own documents and workflows, including Arabic and Turkish content if your customers use them.
• Stay portable. OpenAI co-founded the Agentic AI Foundation, and ChatGPT supports the open Agent Skills format, so documented procedures can move between providers.

The broader trend is clear: capable models are becoming more efficient and affordable quickly, and the focus has moved from chat to agents that act inside software. The competitive advantage now lies less in which model you pick and more in how well you connect it to your systems, and how carefully you limit what it can do.

Talk to us

Vega SDS designs and builds AI-assisted workflows, customer portals and mobile apps for companies in Türkiye and the UAE, working with whichever model provider fits the job. If you are deciding between providers or planning your first agent, we would be glad to share what we see working in practice.

Sources

• constellationr.com/insights/news/openai-launches-gpt-6-astra
• techcrunch.com/2026/09/29/openai-launches-gpt-6-1-sol-says-it-nearly-matches-gpt-6-astra-and-costs-less/
• siliconangle.com/2026/10/26/openai-launches-gpt-6-1-terra-for-faster-better-performing-ai-agents-at-half-the-price/
• anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation