GPT-6 Sol and Luna make model routing the default architecture
The important change is not a new absolute benchmark winner, but a far cheaper lower tier that teams can deploy as a first pass — if they measure completed-task cost rather than token price.
What launched
OpenAI released GPT-6 Sol and GPT-6 Luna through its API on September 22. For requests up to 272,000 input tokens, Sol costs $2 per million input tokens and $10 per million output tokens; Luna costs $0.10 and $0.50. Both have a 1.05-million-token context window and a 128,000-token maximum output.
The benchmark warning
On AutomationBench, which grades the final state of end-to-end business workflows, GPT-6 Sol at xhigh reasoning scored 33.2% at an average cost of $0.27 per task. That is an attractive operating point, but also a warning: a cheaper attempt is not the same thing as reliable execution.
What we would ship
We would make Luna the first route only for bounded, easily verified steps, with automatic escalation of difficult or failed requests to Sol. We would also watch long prompts: above 272,000 input tokens, input and cache rates double while output becomes 50% more expensive, allowing poor context management to consume part of the saving.
Virtual Arc’s view is that GPT-6 Luna is the more important half of this launch. At $0.10 per million input tokens and $0.50 per million output tokens, it is cheap enough to change system architecture: sending every request to one powerful model now makes less sense. But cheap tokens are not the same as cheap completed work. AutomationBench shows that even Sol at xhigh effort completes 33.2% of strictly graded workflows, so this is not a reason to hand unsupervised business processes to an agent. We would immediately shadow-test Luna for classification, extraction, routine code changes and bounded tool calls, while keeping Sol or an established incumbent as the exception path. We would promote it only after measuring cost per completed task, P95 latency, retry volume and final-state correctness. In short: this is worth touching now, but through a model router rather than a wholesale migration.