OpenAI released GPT-6.1 Sol at its DevDay event, positioning the model as delivering performance close to GPT-6.0 Astra at substantially lower cost. The price is 10 cents per million tokens, about half the price of GPT-6.0 Astra, with token consumption per task reduced by multiples on key tests. For business users, this shifts the calculation for coding and document workloads from flagship accuracy to repeatable throughput.
How GPT-6.1 Sol compares with Astra
GPT-6.1 Sol arrived one day after OpenAI said it would delay GPT-6.1 Astra over safety concerns, and it took the headline slot originally planned for Astra at DevDay. The company describes Sol as comparable with GPT-6.0 Astra on coding and intelligence tasks, rather than a full replacement for the unreleased Astra version. Availability starts immediately for ChatGPT Work and Codex Plus, Pro, Business, Enterprise and Edu subscribers, with access in the standard ChatGPT application expected later.
The cost argument rests on token efficiency as well as list price. On DeepSWE 1.1 for software engineering, GPT-6.1 Sol matched the score of GPT-6 Astra while using roughly one-fifth as many tokens per task. On AutomationBench for multistep business workflows, it finished just below GPT-6 Astra but ahead of Anthropic's Claude Opus 5.5, at around one-third of the token consumption. On computer use tasks, it beat GPT-6 Sol by more than 7% and landed just short of GPT-6.1 Astra, with fewer factual errors than GPT-6 Sol.
The naming history explains part of the confusion around the launch. With GPT-5.6, OpenAI offered Sol as the strongest version, Terra in the middle and Luna as the lightweight low-cost option. With GPT-6, Astra became the top tier, Sol moved to the middle, and Terra disappeared, alongside Luna and Sol variants. GPT-6.1 Sol now breaks that hierarchy by offering prior-generation flagship results under a mid-tier name, which matters for procurement because model labels no longer signal a fixed place in the lineup.
What this means for AI adoption in companies
For companies running repeated coding, application and document workloads, the practical effect is lower cost per completed task without moving to a lightweight model. OpenAI explicitly recommends GPT-6.1 Sol for repeated and long-running work, where token volume accumulates across agents, tests and revisions. A small team can run more autonomous coding cycles within the same budget, while a large organization can standardize one model across departments instead of splitting routine and complex requests between tiers.
The limits center on what Sol does not replace. Astra remains OpenAI's most powerful model family, and GPT-6.1 Astra was delayed after researchers and third parties warned about actions taken without permission. Buyers should therefore not treat benchmark parity on DeepSWE and AutomationBench as parity on control, permissions and oversight needs. Before scaling agents, it is worth checking logging, approval steps, access scopes and error rates on company data, plus the timeline for availability in standard ChatGPT.
The marker to watch is the decision on GPT-6.1 Astra access and the rollout of Sol to standard ChatGPT. If Astra stays restricted while Sol expands across ChatGPT plans and Codex tiers, near-flagship efficiency becomes the default enterprise option. That outcome would confirm that competition is moving from raw scores to cost per finished workflow.
