OpenAI is encouraging developers moving from GPT-5.6 Sol to its newer GPT-6 Astra model to consider using lower reasoning settings, arguing that Astra at low reasoning can outperform Sol at high reasoning on some workloads.
Thibault Sottiaux, who leads OpenAI’s core products including Codex, said in a post on X that GPT-6 Astra at low reasoning performs better than GPT-5.6 Sol at high reasoning. He suggested that developers who were satisfied with high reasoning on Sol consider moving to low or medium reasoning on Astra.
The recommendation is significant because GPT-6 Astra carries higher API prices than GPT-5.6 Sol. OpenAI lists Astra at $10 per million input tokens and $50 per million output tokens, compared with $4 and $20, respectively, for Sol. The pricing means Astra costs 2.5 times as much per token, making the number of tokens used to complete a task an important part of the overall cost calculation.
OpenAI’s argument is that token pricing alone does not determine the final cost of a task. A more capable model may require fewer output tokens or fewer interactions to complete the same job, potentially offsetting its higher per-token price.
The company’s published evaluations provide examples of this. On Terminal-Bench 4.0, GPT-6 Astra scored 57.9%, compared with 37.3% for GPT-5.6 Sol, while OpenAI estimated Astra’s API cost per task to be about 9% lower. On GPQA Diamond, Astra’s lower-cost setting scored 94.9%, compared with 94.6% for Sol’s best result, with an estimated 37% lower API cost.
Independent testing also points to the importance of reasoning settings. In a developer comparison reported in the supplied material, Astra at medium reasoning completed a coding workload in fewer requests and with substantially fewer input tokens than Sol at high reasoning. The reported Astra-medium run took about 51 minutes and cost an estimated $25.67, compared with approximately 75 minutes and $31.79 for the Sol-high run.
However, those results should not be interpreted as evidence that Astra’s medium or low setting will always be cheaper or better. AI performance and cost can vary considerably depending on the workload, model setting, number of interactions and amount of reasoning required.
OpenAI’s own model documentation presents Astra as its flagship model for complex reasoning and coding, while positioning other GPT-5.6 models for different cost and workload requirements. Astra supports multiple reasoning levels, including low, medium, high, xhigh and max, giving developers more control over the amount of reasoning used.
The difference highlights a broader issue for developers evaluating AI models: per-token pricing is only one part of the economics of an AI application. A model with a higher price per token can potentially be less expensive for a particular workload if it needs fewer tokens, fewer requests or less time to reach a satisfactory result.
At the same time, developers cannot assume that lowering reasoning will always produce the best balance. More complex tasks may benefit from higher reasoning effort, while simpler or highly repetitive workloads may be better suited to lower-cost models or settings.
The practical implication for developers considering Astra is therefore not simply to compare the listed token prices. Instead, organizations will need to benchmark models using their own workloads, measuring accuracy, latency, token consumption and total cost.
OpenAI’s recommendation effectively shifts the migration question from whether Astra is more expensive per token to whether it can complete a given task more efficiently. For developers, that makes workload-specific testing increasingly important when deciding which model and reasoning level to use in production.