Expanding the GPT-6 Universe

OpenAI has officially introduced GPT-6 Sol and Luna, expanding its model family following the earlier release of GPT‑6 Astra. While Astra remains the company's most intelligent and aligned model designed for the most demanding projects, the new Sol and Luna models are built to distribute frontier intelligence across a wider range of scales, rhythms, and budgets.

The company trained Sol and Luna using methods similar to those applied to Astra. These advancements bring state-of-the-art performance in professional work, factuality, coding, computer use, and alignment into faster and more affordable offerings. Improvements in caching and inference infrastructure have allowed OpenAI to reduce API prices for Sol and Luna by 50% compared with their GPT-5.6 promotional pricing.

Performance and Efficiency in Business Workflows

Designed to tackle difficult work tasks with higher usage limits and lower costs, GPT-6 Sol provides improved intelligence over similarly priced competitor models. Performance evaluations on business workflows demonstrate substantial gains in efficiency and task completion across multiple environments.

On AutomationBench 1.0.6, where AI agents are tested on end-to-end workflows utilizing 47 tools across fields like sales, marketing, operations, support, finance, and HR, GPT-6 Sol at xhigh effort outperforms competitor models at a fraction of the cost per task. Additionally, evaluations on Agents’ Last Exam V1 show that GPT-6 Sol scores competitively while maintaining high cost efficiency on long-horizon professional tasks spanning 55 sub-industries.

Advancements in Coding and Software Engineering

As coding agents increasingly take on complex, long-duration tasks, the cost of sustained development usage has grown more critical. OpenAI notes that internal daily token usage valued at API prices has scaled significantly among researchers, making cost efficiency a vital factor for development teams.

On FrontierCode 1.1 Main, which evaluates whether coding agents produce changes ready for integration into real codebases, GPT-6 Sol shows substantial improvements over its predecessor. Furthermore, testing on DeepSWE 1.1 highlights the capability of Sol and Luna to handle complex software engineering tasks with high accuracy and reduced resource overhead.

Computer Use and Factual Reliability

In addition to software engineering improvements, Sol and Luna bring enhanced performance to computer-use workflows. On OSWorld 2.0 offline, which tests agents on everyday and professional computer tasks, GPT-6 Sol achieves high scores at a fraction of the cost required by comparable models.

OpenAI has also addressed factual reliability across the new models. Internal evaluations using de-identified real-world conversations indicate that GPT-6 Sol cuts mistake rates in half compared to its predecessor, approaching Astra-level reliability. GPT-6 Luna also shows substantial improvements in factual accuracy at higher effort levels.

Communication and Usability Upgrades

Beyond raw capabilities and cost reductions, OpenAI has integrated GPT-6 Astra's improved communication style into both Sol and Luna. Users can expect clearer technical and coding conversations characterized by reduced jargon, fewer odd turns of phrase, minimal low-value details, and slightly shorter answers overall without losing substance.

Sources

• OpenAI NewsIntroducing GPT-6 Sol and Luna