OpenAI GPT-6 Halves Token Costs Across Benchmarks as Sol and Luna Join

According to Artificial Analysis, GPT-6 Sol (max) ranks as the top model for intelligence, priced at $48. Image credit: Getty Images

After OpenAI released GPT-6 Sol and Luna, Matt Weaver, Head of Solutions Engineering, explained to AI Magazine how these new models provide businesses with enhanced operational flexibility.

OpenAI has recently unveiled GPT-6 Astra, describing it as the “most intelligent and aligned model in the world.” The AI laboratory has subsequently announced GPT-6 Sol and GPT-6 Luna, cost-effective variants trained using methodologies similar to GPT-6 Astra, its flagship model.

Amid growing concerns over rising token usage costs, AI companies are increasingly competing on price, particularly as Chinese open-weight models threaten to capture enterprise clients.

Benchmarking firm Artificial Analysis states that GPT-6 Sol and Luna are “pushing the cost efficiency frontier” by reducing costs by half compared to GPT-5.6 Sol and Luna.

Matt Weaver, Head of Solutions Engineering at OpenAI, told AI Magazine: “With GPT-6 Sol and Luna, we’re giving businesses more flexibility to match the model to the work.”

Matt Weaver, Head of Solutions Engineering at OpenAI. Image credit: Matt Weaver/LinkedIn

Aligning models with workflows

OpenAI claims its GPT-6 models lead the cost-intelligence curve, delivering exceptional capabilities across all tiers through infrastructure designed for efficient, large-scale deployment.

The company notes that advancements in caching and inference enable these models to be served at lower costs, with savings passed directly to users through a 50% reduction in API prices for Sol and Luna compared to GPT-5.6 promotional rates.

Explaining the distinct roles of each AI model, Matt adds: “GPT-6 Astra is intended for the most demanding projects, where maximum capability is critical.

Matt explains: “GPT-6 Sol is built for complex professional tasks, including intricate business workflows and coding, while GPT-6 Luna provides a more efficient solution for high-volume, well-defined tasks such as document summarization or information extraction.

“Sol and Luna bring Astra’s advancements in professional work, factuality, coding, and computer use to faster, more affordable models. As AI transitions from experimentation to core business operations, selecting the right model for each task type is crucial.

“Our goal is to make advanced AI practical across organizations – helping teams move faster, make better decisions, and focus more time on higher-value work.”

Key facts

• OpenAI cut API prices for GPT-6 Sol and Luna by 50% compared to GPT-5.6
• GPT-6 Sol surpasses Claude Opus 5 on AutomationBench at just 9% of the cost
• GPT-6 Sol (max) achieves output speeds of 131 tokens per second
• Intelligence Index task costs dropped to US$1.06 for Sol and US$0.07 for Luna
• Both GPT-6 Sol and Luna demonstrate reduced hallucination on the AA-Omniscience benchmark.

Model performance analysis

OpenAI highlights that on AutomationBench, a test of business workflows across applications, GPT-6 Sol at xhigh effort outperforms Claude Opus 5 at max effort at just 9% of Opus 5’s cost per task.

At high effort, GPT-6 Luna improves on its predecessor by 5.4 percentage points at 58% lower cost per task, the company notes.

OpenAI points out that on Agents’ Last Exam, which evaluates agents on complex professional workflows, GPT-6 Sol at max effort scores 56.4%, exceeding Claude Opus 5’s highest score in the evaluation at 61% lower cost per task.

The company notes that on FrontierCode, which assesses whether coding agents produce changes ready to merge into real codebases, GPT-6 Sol shows substantial improvement over GPT-5.6 Sol and matches Claude Fable 5.1 xhigh at a much lower cost.

Both models exhibit reduced hallucination in AA-Omniscience, Artificial Analysis’s benchmark for knowledge and hallucination.

With GPT-6 Sol and Luna, we’re giving businesses more flexibility to match the model to the work

Matt Weaver, Head of Solutions Engineering at OpenAI

Cost per Intelligence Index Task. Image credit: Artificial Analysis

The benchmarking firm explains that the GPT-6 Sol release includes six models, each with distinct intelligence, performance, and pricing characteristics. Comparing key metrics across these six models, Artificial Analysis found:

• For intelligence, the top model of GPT-6 Sol is GPT-6 Sol (max) at 48.
• For output speed, the fastest model is GPT-6 Sol (max) at 131 t/s.
• For latency, GPT-6 Sol (Non-reasoning) at 0.93s offers the lowest time to first answer token.
• For pricing, GPT-6 Sol (low) at US$0.13 offers the lowest cost per task. Prices vary up to 8x across models.

The firm found that across models, the Intelligence Index and Coding Agent Index scores remain consistent with GPT-5.6, with progress in some evaluations and regressions in others.

GPT-6 Sol is designed for difficult professional work – from complex business workflows to coding – while GPT-6 Luna offers a more efficient option for high-volume, clearly defined tasks like summarising documents or extracting information

Matt Weaver, Head of Solutions Engineering at OpenAI

In FrontierCode 1.1 Main⁠, AI agents write code that’s graded not only on correctness but also “mergeability”: e.g., test quality, scope discipline, code style and adherence to codebase standards. Image credit: OpenAI

The importance of cost efficiency

Competition over AI model pricing is intensifying, often driven by low-cost Chinese models. As Kyle Chan at the Brookings Institution notes in a blog, Chinese models are mostly open-source, allowing users to customize the model and access it at lower prices with 90% of the capabilities of a US model. Consequently, these models are rapidly gaining ground among cost-motivated enterprise firms.

OpenAI is positioning itself to remain competitive; running the Artificial Analysis Intelligence Index is significantly cheaper with the new models, even though both output slightly more tokens per task:

• GPT-6 Sol (max): Costs US$1.06 per task (down ~50% from US$1.99 for GPT-5.6 Sol), even with output token volume rising from 29k on GPT-5.6 to 31k.
• GPT-6 Luna (max): Costs US$0.07 per task (down ~60% from US$0.18 for GPT-5.6 Luna), despite output token usage increasing from 41k on GPT-5.6 to 51k.

Artificial Analysis highlights that these two releases allow OpenAI to capture 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 without sacrificing capability, OpenAI is positioning itself to address enterprise ‘token fatigue’ while defending against its low-cost market competitors.