OpenAI Launches GPT-6 Sol and Luna, Cutting API Inference Costs by 50%

OpenAI expanded its GPT-6 family with the release of GPT-6 Sol and GPT-6 Luna. Following the earlier debut of the flagship GPT-6 Astra, these two new variants are engineered to distill Astra's performance in coding, factual accuracy, computer-use automation, and alignment into faster, more economical form factors. OpenAI announced a 50% price reduction for Sol and Luna API access relative to previous GPT-5.6 tiers ($2 per million input tokens and $10 per million output tokens for Sol; $0.10 input and $0.50 output for Luna). Alongside the model launches, improved prompt caching mechanisms deliver up to 90% discounts on cached token reads, with immediate availability across the OpenAI API, ChatGPT Work, and Codex. For platform engineers and AI architects, the significance lies in the collapsing cost-to-performance curve for reasoning and computer-use tasks. On long-horizon evaluation benchmarks such as OSWorld 2.0, GPT-6 Sol achieves parity with higher-tier frontier alternatives while running at approximately 80% lower cost per completed task. Production deployments of agentic workflows—such as repository-wide automated refactoring, continuous test generation, and multi-turn document parsing—often exhaust token budgets rapidly. Sol and Luna shift these agentic workloads from high-overhead experimental setups into commercially viable, high-volume production operations. This release reflects an ongoing industry pivot in foundation model engineering: labs are no longer solely competing on raw parameter scale or ceiling capability benchmarks, but on inference efficiency and operational margins. As open-weight and proprietary models rapidly close intelligence gaps, API providers are heavily investing in architectural distillation, hardware-aware execution, and server-side KV cache optimization to drive down token unit economics. By bundling native prompt caching with tiered reasoning efforts, OpenAI is standardizing a paradigm where system latency and cost are dynamically modulated per workload requirement. In practice, engineering teams should evaluate routing strategies across their LLM pipelines. Workloads previously gated by GPT-5.6 or Astra-tier price points can be benchmarked against GPT-6 Sol to determine if accuracy thresholds hold at the lower cost profile. Additionally, teams should audit their prompt engineering patterns to maximize cache hit rates under the new 90% cached token discount, ensuring static system prompts, schema definitions, and persistent workspace contexts are structured sequentially to leverage prefix caching effectively.