OpenAI’s GPT-6 Sol and Luna Models Cut Token Costs by Half While Maintaining High Performance

• OpenAI has reduced API prices for GPT-6 Sol and Luna by 50% compared to their GPT-5.6 promotional pricing, leveraging improvements in caching and inference infrastructure.
• On the AutomationBench benchmark, GPT-6 Sol outperforms Claude Opus 5 at just 9% of the cost per task, demonstrating superior cost-intelligence efficiency.
• GPT-6 Sol (max) achieves an output speed of 131 tokens per second and ranks highest in intelligence with a score of 48 on the Intelligence Index.
• Task costs for the Intelligence Index have dropped to US$1.06 for GPT-6 Sol and US$0.07 for GPT-6 Luna, representing significant savings despite increased token output volumes.
• Both new models exhibit reduced hallucination rates on the AA-Omniscience benchmark, enhancing reliability for professional workflows.

OpenAI positions its GPT-6 lineup to provide businesses with greater flexibility in matching specific AI models to distinct operational needs. The company emphasizes that these models lead across the cost-intelligence curve, combining high-tier capabilities with infrastructure designed for efficient scaling. By passing infrastructure savings directly to users through a 50% price reduction, OpenAI aims to make advanced AI practical for daily organizational use.

Matt Weaver, Head of Solutions Engineering at OpenAI, explains that the different models serve specific purposes within an enterprise ecosystem. GPT-6 Astra remains the choice for the most demanding projects where maximum capability is non-negotiable. However, GPT-6 Sol and Luna bring Astra’s advancements in factuality, coding, and computer use to faster, more affordable tiers.

GPT-6 Sol is designed for difficult professional work, ranging from complex business workflows to intricate coding tasks. In contrast, GPT-6 Luna offers a more efficient option for high-volume, clearly defined tasks such as summarizing documents or extracting information. This segmentation allows organizations to optimize their spending by avoiding the use of premium models for routine operations.

The performance of these new models has been evaluated against several industry standards, highlighting significant improvements in both speed and cost-effectiveness.

On AutomationBench, a test measuring business workflows across applications, GPT-6 Sol at xhigh effort outperforms Claude Opus 5 at max effort while costing only 9% of Opus 5’s price per task. Similarly, on Agents’ Last Exam, which evaluates agents on complex professional workflows, GPT-6 Sol at max effort scores 56.4%, surpassing Claude Opus 5’s highest score in the evaluation at a cost that is 61% lower per task.

In the realm of software development, GPT-6 Sol shows substantial improvement over its predecessor, GPT-5.6 Sol. On FrontierCode, which assesses whether coding agents produce changes ready to merge into real codebases, GPT-6 Sol matches Claude Fable 5.1 xhigh at a much lower cost. The evaluation criteria for FrontierCode include not only correctness but also "mergeability," such as test quality, scope discipline, code style, and adherence to codebase standards.

GPT-6 Luna also demonstrates efficiency gains. At high effort, it improves on its predecessor by 5.4 percentage points while reducing the cost per task by 58%. Both GPT-6 Sol and Luna show reduced hallucination rates on the AA-Omniscience benchmark, a key metric for knowledge reliability.

Artificial Analysis’s data reveals that the GPT-6 Sol release comprises six distinct models, each with unique intelligence, performance, and pricing characteristics. The variability in pricing across these models can be up to eight times.

Key metrics identified by Artificial Analysis include:

1. Intelligence: The top model is GPT-6 Sol (max), achieving a score of 48 on the Intelligence Index.
2. Speed: GPT-6 Sol (max) is the fastest, delivering outputs at 131 tokens per second.
3. Latency: GPT-6 Sol (Non-reasoning) offers the lowest time to first answer token at 0.93 seconds.
4. Lowest Cost: GPT-6 Sol (low) provides the most affordable option at US$0.13 per task.

The cost reductions are substantial when compared to previous generations. For the Intelligence Index, GPT-6 Sol (max) costs US$1.06 per task, down approximately 50% from US$1.99 for GPT-5.6 Sol. Notably, this reduction occurs even though output token volume has increased from 29k on GPT-5.6 to 31k. Similarly, GPT-6 Luna (max) costs US$0.07 per task, a drop of approximately 60% from US$0.18 for GPT-5.6 Luna, despite token usage rising from 41k to 51k.

The intensifying cost wars in the AI sector are largely driven by the availability of low-cost Chinese models. These open-source alternatives allow users to customize models and access them at lower prices while retaining most of the capabilities found in US-developed systems. This dynamic has led to a rapid gain in market share among cost-motivated enterprise firms.

OpenAI’s strategy is to remain competitive by capturing a significant portion of the cost-efficiency Pareto frontier—the optimal boundary where performance cannot be increased without a corresponding price hike. By lowering barriers to entry without sacrificing capability, OpenAI aims to address "token fatigue" in enterprises while defending against low-cost market competitors. The goal is to help organizations move faster, make better decisions, and allocate more time to higher-value work by utilizing the right model for each specific type of task.