OpenAI Releases GPT-6 Sol and Luna: API Prices Cut in Half, but the Top Scores Remain with the Previous Generation

On September 22, 2026, OpenAI introduced two new members of the GPT-6 family—the reasoning-heavy model Sol and the high-throughput model Luna. An OpenAI spokesperson confirmed that the pricing is permanent and not a promotion. GPT-6 Sol's API price dropped to $2 per million input tokens and $10 per million output tokens, half the $4/$20 of its predecessor, GPT-5.6 Sol; GPT-6 Luna is priced at $0.10/$0.50, with a discount of up to 90% on cached input tokens.

At almost the same time, Anthropic released Claude Opus 5.5 and announced a 40% price cut versus Opus 5 and a 60% reduction in cache read costs. The two companies cut prices in sync within the same week, which looks more like a collective response by the industry to the structural pressure of open-source models continually closing in on the closed-source frontier.

Luna's Real Surprise: Near-Level Performance at One-Twentieth the Price

GPT-6 Sol and Luna are designed with layered product positioning: Sol targets complex technical tasks such as code review and data analysis, while Luna handles high-frequency, mass-production work such as document summarization and information extraction. But according to analysis by digital application research firm DigitalApplied, on the DeepSWE coding benchmark, Luna scores 66.6%, nearly level with Sol's 68.8%, while Luna's cost per task is about one-fifth that of Sol.

This means that for the vast majority of developers' day-to-day coding assistance work, Luna may be a more rational starting point than Sol. A cost gap of as much as 20x results in less than 3 percentage points of performance loss.

OpenAI strengthened reasoning and caching mechanisms in both models, causing the cost of calls with repeated context to fall sharply. A 90% cached input discount means that for agent workflows carrying large fixed system prompts or conversation history, actual operating costs may be far below what the list price implies.

The Other Side of Performance Progress: The Old Model Is Still Stronger on Key Benchmarks

This release contains a detail that is easy to obscure. According to DigitalApplied's analysis, under the highest-effort setting for DeepSWE coding tasks, GPT-5.6 Sol scores 72.7%, actually higher than the new GPT-6 Sol's 68.8%. On the business process automation benchmark AutomationBench, Sol scores 33.2%, below the flagship GPT-6 Astra's 41.4%.

In other words, the essence of this release is not "stronger models" but "cheaper models." OpenAI's engineers clearly made trade-offs between computational efficiency and raw capability—for enterprise users, this is not a hidden danger but a purchasing-decision framework that requires clear-headed awareness: if your workload is sensitive to peak performance, GPT-6 Astra is the right option; if you are cost-sensitive, Sol and Luna's cost is an acceptable capability concession.

Split Architecture: A Commercial Response Under Open-Source Pressure

The GPT-6 product matrix—Astra (flagship), Sol (mid-tier complex tasks), and Luna (high-frequency mass production)—is a direct response to how open-source models have attacked over the past two years. Open-source models such as Meta Llama series, Mistral, and DeepSeek already compete intensely with closed-source models on low- and medium-difficulty tasks, and at prices the latter cannot offer.

OpenAI's strategy is to use the flagship Astra to defend the performance ceiling while using Sol and Luna to directly block open-source alternatives across the price band. Luna's input price of $0.10 per million tokens is already approaching the self-hosted marginal cost range of some open-source models.

According to TechCrunch, Sol and Luna scored 1.3% and 2.8%, respectively, in enterprise deception rate tests, far below the previous generation. However, DigitalApplied points out that Sol's pass rate in "warning bypass" tests is as high as 64.4%, meaning that in a substantial proportion of cases the model can be made to bypass its built-in reminders by adversarial prompts. This figure should not be obscured by the pricing narrative—for scenarios that require deploying models in uncontrolled environments, it is an engineering hazard that needs serious evaluation.

The Significance of Cache Architecture: Rebuilding the Cost Model for the Agent Era

The deepest technical signal of this price cut is not the list price itself, but the figure of a cache discount as high as 90%. In traditional API calls, the main cost driver is the ratio of input tokens to output tokens; but in multi-turn conversational agents, code assistants with long system prompts, or RAG systems that need to repeatedly reference the same document set, cache hit rate is the core variable determining cost.

A 90% discount means OpenAI has explicitly acknowledged in its pricing model that future high-value customers are not one-off API callers, but enterprises that continuously run agent workflows. This adjustment echoes Anthropic's simultaneous sharp reduction in Claude Opus 5.5 cache read costs (cache reads fell from $0.50 to $0.20 per million tokens, a 60% drop).

The two companies made architecture adjustments in the same direction in the same week, which has a clear reading at the AI engineering level: cache-intensive agent workflows are becoming the standard usage pattern for flagship customers, and pricing models must be calibrated accordingly.

Judging the Nature of the Price War

This round of price cuts is structural, not strategic. OpenAI officially told VentureBeat that Sol and Luna's pricing is permanent, not promotional or introductory pricing. Anthropic's synchronized move is the same.

A more important signal is that the two rounds of price cuts (ranging from 40% to 90%) were achieved without an obvious breakthrough in hardware costs, indicating that model architecture and inference optimization themselves have created marginal cost space—not that they are subsidizing against competitors. This differs from the situation two years ago when Anthropic followed GPT-4 Turbo's price cut; at that time, the cuts had a clear market-share defense flavor. Today's synchronized cuts look more like the inference cost curve collectively reaching the same inflection point.

For enterprise users, the practical operational judgment is: if the current workload runs on GPT-5.6 Sol, switching to GPT-6 Sol can immediately yield about 50% cost savings, at the price of a slight decline on some peak coding benchmarks; if the workload itself does not depend on code generation under extreme effort, Luna offers nearly the same practical performance at a lower price and is worth testing first. If higher requirements are placed on the raw capability ceiling of these two models, GPT-6 Astra and Claude Opus 5.5 are the real performance insurance, but you need to pay a premium of more than an order of magnitude for it.

This release essentially accomplished one thing: it moved down a level the cost band that previously only flagship models could cover. The barrier to AI infrastructure has become lower, but that does not mean choosing the optimal model for every task has become simpler—on the contrary, it requires developers to have a more precise understanding of their own workloads to truly benefit from this price restructuring.