OpenAI Unveils GPT-6 Sol & Luna, Supercharging ChatGPT Voice with Plugins and Advanced AI

OpenAI's latest additions, GPT-6 Sol and GPT-6 Luna, are strategically positioned as more accessible yet powerful alternatives to the top-tier GPT-6 Astra. Released on September 22, 2026, these models are designed to bring the advanced intelligence generation capabilities of the GPT-6 series to a broader range of applications, focusing on optimizing for both cost and speed. This release is particularly exciting for developers, as it directly addresses the often-cited challenge of balancing high-performance AI with practical operational costs.

GPT-6 Sol emerges as a balanced model tailored for interactive and 'agentic' coding tasks. It excels in multi-step workflows such as building new features, reviewing and refactoring code, debugging complex issues, and analyzing data. A key feature of Sol is its granular control over reasoning effort, allowing developers to specify levels from none to max. This flexibility enables a trade-off between latency, cost, and the depth of analysis per request, making it highly adaptable to varying task complexities.

On the other hand, GPT-6 Luna is presented as a lightweight, cost-efficient model optimized for high-volume, focused, and repeatable tasks. Its strengths lie in summarizing documents, extracting specific information, classifying inputs, and handling quick, routine questions across numerous users or applications. Luna is the most economical option within the GPT-6 family, making it ideal for scenarios where efficiency at scale is paramount and small differences in latency and cost can significantly impact overall expenses.

Both Sol and Luna were trained using methods similar to Astra, resulting in notable improvements over their GPT-5.6 predecessors across various metrics, including professional work, factuality, coding accuracy, and computer utilization. OpenAI reports that GPT-6 Sol, for instance, exhibits approximately half the factual mistakes of GPT-5.6 Sol. Furthermore, the models demonstrate improved alignment, leading to lower rates of misleading claims, particularly in coding-related tasks. These advancements mean developers can rely on more accurate and trustworthy outputs, reducing the need for extensive post-processing and validation.

Beyond the new models, OpenAI has delivered a significant overhaul to ChatGPT Voice, transforming it into a much more versatile and integrated conversational agent. Rolled out globally starting September 23, 2026, these upgrades are set to enhance user productivity and open new avenues for developers to create voice-enabled applications.

The most impactful update is the introduction of plugin support for ChatGPT Voice. Users can now connect their voice interactions with external services like email, calendar, and Slack. This means a spoken request can now trigger actions in connected applications, such as drafting an email, scheduling a meeting, or summarizing recent Slack messages. This integration bridges a crucial gap, allowing voice commands to extend beyond mere conversation to tangible, multi-application workflows. Developers can leverage this to build more sophisticated voice assistants that interact with a user's digital ecosystem seamlessly.

Another major enhancement is the ability for ChatGPT Voice to be powered by the entire GPT-6 model family: Astra, Sol, and Luna. Previously, voice interactions might have been tied to a single model; now, users and developers have the flexibility to select the underlying GPT-6 model that best suits the task at hand. For instance, a developer might choose GPT-6 Astra for highly complex reasoning tasks, Sol for intricate coding queries, or Luna for quick, high-volume information retrieval, all through a natural voice interface. This allows for a more tailored and efficient voice experience, optimizing for both performance and cost.

Furthermore, ChatGPT Voice is now fully integrated within ChatGPT Work, accessible on both web and mobile platforms. This integration empowers users to create documents, prepare presentations, and tackle other complex tasks in the browser or on the go, simply by speaking. This repositioning of Voice from a conversational novelty to a robust work interface signifies OpenAI's commitment to making AI a fundamental part of daily productivity. For developers, this opens doors to building voice-driven enterprise solutions and agentic systems that can interpret and execute complex commands across various business applications.

OpenAI has also refined the communication style of voice conversations, promising richer text output and the ability for users to easily switch between text and voice modes. This focus on a more natural and fluid interaction aims to reduce friction and make AI assistance more intuitive and omnipresent.

The release of GPT-6 Sol and Luna, coupled with the enhanced ChatGPT Voice capabilities, carries significant implications for developers, particularly concerning economics and performance. A standout feature of Sol and Luna is their significantly reduced API pricing – approximately 50% lower than their GPT-5.6 predecessors. GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens, while GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. This aggressive pricing strategy is a game-changer, making advanced AI more accessible for a wider range of projects and allowing developers to scale their AI applications without prohibitive costs.

The economic advantage extends beyond just token pricing. OpenAI has also improved prompt caching, offering discounts of up to 90% on cached input tokens. This is particularly beneficial for AI agents and long-running conversations where context is frequently reused. By reducing the cost of reprocessing context, agents can work longer, process more information, and undertake more steps before cost becomes a limiting factor. This efficiency gain translates directly into more ambitious and complex AI agent designs, enabling developers to build more robust and capable autonomous systems.

From a performance perspective, GPT-6 Sol and Luna not only offer cost savings but also boast genuine capability improvements. Benchmarks show Sol outperforming GPT-5.6 Sol in coding and computer use, and even matching or exceeding some scores of competitor models like Claude Fable 5.1 on certain tasks. For agent automation, GPT-6 Sol at maximum reasoning effort scored 33.2% on OpenAI's AutomationBench evaluation, compared to 30.3% for GPT-6 Astra at low effort, completing tasks at a fraction of the cost of some competing models. Luna also shows generational improvements, enhancing performance over GPT-5.6 Luna by 5.4 percentage points at a 58% lower cost per task on similar benchmarks.

The new reasoning effort tiers in Sol (none, low, medium, high, xhigh, and max) provide developers with fine-grained control to optimize for specific use cases. For simple, repetitive tasks, setting reasoning to none can significantly reduce latency and cost. For more complex problems, higher reasoning tiers ensure deeper analysis and more accurate results, albeit at a higher computational expense. This flexibility allows developers to dynamically adjust model behavior based on the real-time demands of their applications, ensuring optimal resource utilization and performance.