OpenAI’s new GPT-6.1 Sol gets surprisingly close to Astra at a fraction of the cost

OpenAI has gone ahead with a new GPT-6.1 model even after deciding not to release GPT-6.1 Astra over safety concerns. The company has unveiled GPT-6.1 Sol, an upgraded version of GPT-6 Sol that promises near-Astra performance at a much lower cost.

The launch comes just after OpenAI pulled the public release of GPT-6.1 Astra. Internal testing found that Astra had become more capable at completing complex tasks, but also worse at staying within user-set limits. OpenAI also found problems involving deception and what it calls “scope authorization,” where the model could continue pursuing tasks or reach for outside tools without permission.

GPT-6.1 Sol arrives with a different pitch. OpenAI says the model gets close to Astra across coding, computer use, professional work, and scientific research while costing roughly one-fifth as much at standard API rates.

GPT-6.1 Sol gets much closer to Astra

The gains over GPT-6 Sol are fairly substantial in some tests. On DeepSWE, GPT-6.1 Sol matches GPT-6 Astra at roughly one-fifth the cost, while beating GPT-6 Sol’s best score by 6.4 percentage points.

Computer use also sees a noticeable jump. On OSWorld 2.0’s offline set, GPT-6.1 Sol beats GPT-6 Sol by seven percentage points at maximum reasoning effort and comes within 2.1 points of Astra while costing roughly one-seventh as much per task.

Scientific work shows an even bigger improvement. GPT-6.1 Sol more than doubles GPT-6 Sol’s score on Terminal-Bench Science 0.1 at maximum reasoning effort. Astra still leads the test at 68.1%, while OpenAI says GPT-6.1 Sol costs $5.47 per task compared with $23.80 for Astra.

Safety has improved, but Astra still has an edge

OpenAI says GPT-6.1 Sol is also more reliable than GPT-6 Sol at staying within user intent and safety restrictions. Its computer-use safety stress-test failure rate drops from 17.4% on GPT-6 Sol to 4.3%. Astra remains lower at 2.4%.

A similar gap appears in warning circumvention. GPT-6.1 Sol records a 23.5% attempt rate compared with 64.4% for GPT-6 Sol, while Astra sits lower at 17.4%. Those results are notable given why Astra was held back. OpenAI found the model was less reliable when explaining what it had done and more willing to cross user-set boundaries.

GPT-6.1 Sol is available today to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex, but not yet in regular ChatGPT conversations. API access is also live at $2 per million input tokens and $10 per million output tokens, while cached input costs $0.10 per million tokens.

RELATED COVERAGE