OpenAI Launches GPT-6 Sol and Luna: Prices Cut in Half, Focused on Cost Per Task

GPT-6 Family Arrives: Two Models for Different Use Cases

OpenAI has introduced two new members of the GPT-6 family — GPT-6 Sol and GPT-6 Luna — targeting the high-end and low-end segments respectively. Sol is designed for complex recurring tasks, while Luna targets high-throughput extraction and classification workloads. Compared to their predecessors, the biggest change isn't a dramatic performance leap but rather a significant price reduction and a rethinking of cost structure.

According to official pricing, Sol comes in at $2 per million input tokens and $10 per million output tokens — exactly half the price of GPT-5.6 Sol. Luna is even more aggressive, at just $0.10 per million input tokens and $0.50 per million output tokens, representing a 50% reduction in input costs and a 58.3% reduction in output costs compared to the previous Luna generation. OpenAI has explicitly stated these are permanent prices, not limited-time promotions — meaning the cost reduction will directly impact enterprises' long-term deployment budgets.

50% cheaper than GPT-5.6 Sol across both dimensions

The Signal Behind Permanent Price Cuts

Previous price cuts in the LLM industry often came in the form of promotions or phased adjustments. By designating these as "permanent prices," OpenAI is sending a clear signal: the continued decline in inference costs is becoming the central battleground in AI competition. As model capabilities converge, whoever can deliver equivalent quality at a lower per-unit cost is better positioned to win customers at scale.

OpenAI says this is a permanent price, not a promotional price

From Token Pricing to the "Cost Per Task" Narrative

What truly stands out about this launch is OpenAI's deliberate emphasis on the concept of cost per task. Comparing price per million tokens alone often fails to reflect a model's real-world economic efficiency — a smarter model may complete a task in fewer steps with a shorter reasoning chain, ultimately winning on total cost even if its per-token rate is higher.

On Automation Bench, an automated task benchmark, Sol achieved a score of 33.2% with a cost per task of $0.27. By comparison, Claude Opus 5 on the same benchmark costs 11 times more per task than Sol. This gap is far larger than the raw token price difference alone, illustrating that OpenAI wants to shift the evaluation metric from "cost per token" to "cost to complete a task."

This narrative is especially compelling for enterprise users. In production environments, decision-makers care less about abstract token billing and more about the real cost of processing a batch of tickets, handling a set of documents, or running an end-to-end automated workflow.

Automation Bench is a benchmark specifically designed to evaluate AI model performance in real-world automated workflows. Unlike traditional academic benchmarks such as MMLU or HumanEval, which focus on knowledge Q&A or code generation, Automation Bench simulates multi-step enterprise task completion — such as cross-system data processing, form filling, and workflow orchestration. It scores models by task completion rate (%) while also tracking actual API costs per task. This design makes it more reflective of production-environment priorities: a model that costs more per token but reasons more efficiently can easily beat a nominally cheaper competitor on cost per task. By citing this benchmark as its primary comparison, OpenAI is deliberately steering the competitive conversation away from pricing tables and toward efficiency metrics.

Factual Accuracy Improvements and Caching Discounts

Beyond pricing, model reliability has also improved. According to official data, Sol's factual error rate is approximately half that of GPT-5.6 Sol. Halving the error rate on complex, accuracy-sensitive tasks means less manual review and rework — itself a hidden component of cost per task.

Sol's errors are roughly half those of GPT-5.6 Sol on factuality

Luna, when using the high-reasoning tier, can match the previous generation's factual accuracy at roughly 1% of the task cost — demonstrating significantly improved price-performance in targeted configurations. Additionally, cached input reads receive a 90% discount. For applications with large amounts of repeated context (e.g., fixed system prompts, repeated reference documents), this discount can dramatically reduce real-world costs.

The Logic Behind Sol and Luna's Division of Labor

The positioning of Sol and Luna is clearly delineated: Sol handles "complex recurring tasks," suited for multi-step reasoning and high accuracy requirements; Luna handles "high-throughput extraction and classification," suited for large-volume, structured tasks that are sensitive to latency and throughput. This high-low pairing lets enterprises allocate models according to task complexity — using the right tool for the job rather than paying premium-model prices across the board.

Sol handles complex recurring tasks

Prompt Caching is an optimization mechanism in LLM APIs: when a request contains prefix content identical to a previous request (such as a fixed system prompt, a long reference document, or repeated context), the server can reuse the already-computed KV cache, skipping redundant inference for those tokens. For users, cache-hit tokens are billed at a lower rate (10% of the standard price in this case); for the provider, it saves GPU compute. This mechanism is especially beneficial for "high-concurrency + fixed-context" applications — for example, processing hundreds of user requests under the same long system prompt. The higher the cache hit rate, the greater the effective discount relative to the listed price.

Ecosystem Availability: Now Live on Amazon Bedrock

Both models are now officially available on Amazon Bedrock. Through the cloud platform's distribution channel, enterprise developers can call these models within their existing AWS workflows without any separate integration, further lowering the adoption barrier. This also reflects OpenAI's continued effort to expand its commercial reach and accelerate model deployment in enterprise production environments.

Amazon Bedrock is AWS's managed foundation model service platform, allowing developers to access models from multiple AI companies — including Anthropic, Meta, Mistral, and others — through a unified API, without managing model deployment or inference infrastructure themselves. For enterprise users, using models on Bedrock means they can directly leverage existing IAM permission systems, VPC network isolation, CloudWatch monitoring, and AWS compliance certifications, significantly reducing the engineering and compliance costs of embedding AI capabilities into internal systems. By bringing new models into Bedrock, OpenAI extends its reach to a large segment of enterprises deeply integrated with the AWS ecosystem — customers who may have previously delayed adopting OpenAI models in production due to integration complexity.

Closing Observations

The GPT-6 Sol and Luna update is less about pushing capability ceilings and more about systematically optimizing cost structure. Permanent price cuts, the cost-per-task narrative, improved factual accuracy, and caching discounts all point toward the same goal: making large language models more economical in real business scenarios. As model capabilities enter a plateau, competition around per-unit cost efficiency may well become the defining theme of the industry in the period ahead.