You don't have to use the most powerful AI model for every task. Vaibhav Srivastav, OpenAI's developer experience and community lead for EMEA, told The New Stack that he doesn't use GPT-6 Astra every day. He usually chooses GPT-6.1 Sol for everyday work and turns to Astra when a difficult or unfamiliar problem needs closer attention.
His approach goes beyond asking which model is smarter. It also involves deciding what result you need, how much time and money you can spend, and how you will check the work afterward.
Why Astra isn't his everyday choice
According to Srivastav, Sol is enough for most of the tasks he encounters. He points to its balance of capabilities and cost: OpenAI aimed to make a model that comes close to Astra's performance while being more accessible.
In the interview, he refers to model benchmarks and to how people use the models inside the company. His estimate of “about 95%” is not a guarantee that Sol will perform nearly as well as Astra on every individual task. The result depends on the task and on how you define success.
One practical takeaway is to try a model that may be sufficient before choosing the most powerful one. For a draft email or a short document summary, additional reasoning capability may not be necessary. If the answer misses important conditions or takes too many corrections, it may be worth trying another model.
When Srivastav turns to Astra
He says he turns to Astra when starting a new project or facing an unfamiliar problem. He asks the model to look for weaknesses and things he may have missed himself.
Another part of his routine is to separate the work from its review. If Sol helped him develop a feature, he may bring in Astra to examine the result critically. He gives the second model fresh context and asks it to look for ways the solution could fail.
Srivastav compares reviewing work with the same model to checking your own work: in his view, it can be easier to overlook mistakes. A different model may offer another perspective, but that does not make it a fully independent expert. It can also miss problems or confidently flag issues that aren't there.
What another OpenAI employee says
There is another way to look at the decision: consider the whole model lineup. In our GPT-6.1 Terra overview, we discussed OpenAI employee Ted Sanders' comment on Hacker News. He suggested using Luna → Terra → Sol → Astra as a general guide: choose Astra when quality matters most, Sol with moderate budget constraints, and Terra or Luna when the budget is tighter.
Sanders also cautioned that this is a general guide, not a ranking for every task. In his assessment, Terra may be particularly useful for technical writing, image understanding and working with interfaces, while falling somewhat behind the larger models on most other tasks. This is an assessment by an employee of the developer, not a guaranteed result or an independent comparison.
The two views complement each other: Srivastav describes how he personally switches between Sol and Astra; Sanders suggests considering other models in the family as well, depending on the task and budget.
How to decide whether to switch
Srivastav recommends evaluating models on individual real-world tasks: compare whether you can get the same outcome for less or a better outcome at an acceptable cost. He also discusses speed: a faster tier may help when an answer is needed in real time, but not every task justifies the extra expense.
Look beyond the cost of one request. If a cheaper model needs several attempts while a more expensive one produces a usable result immediately, compare the whole process: requests, time spent correcting and the quality of the final work. For a straightforward task, the more expensive model may add little value.
Where a second model can help
The review approach can work beyond software development. One model might create a project plan while another checks for missing dependencies and potential delays. For a report, a second model could look for contradictions and claims that need sources.
For code, the first model might write a function and the second flag edge cases and possible errors. For a budget, one model might assemble the calculation while another checks formulas, assumptions and missing expenses.
Give the reviewing model both the original task and the first result; otherwise it cannot check whether the requirements were met. A person still needs to assess its comments. For important documents, financial decisions or security questions, this review can point to issues that need further checking, but it does not replace a specialist.
Sources
- The New Stack, October 27, 2026 - interview with Vaibhav Srivastav about model choices and OpenAI agents.
- Ted Sanders on Hacker News - comment on choosing models and GPT-6.1 Terra's place in OpenAI's lineup.