OpenAI's own developer lead picks Sol over Astra as his daily driver

An OpenAI developer lead told a reporter this week that he does not use the frontier model on most days. Vaibhav Srivastav is OpenAI's developer experience and community lead for EMEA. He sat for an interview with The New Stack after DevDay 2026. He spelled out the rule he uses to pick between GPT-6.1 Sol, GPT-6 Astra and the Ultrafast speed tier.

The headline is in his own words. "I don't use Astra every day. I do use GPT-6.1 Sol as my main go-to."

The rule Srivastav gives

Sol gets about 95 percent of Astra's performance at one-fifth the cost, Srivastav said. That is the ratio OpenAI quoted at launch. The developer lead treats it as a routing default. Sol runs the normal ticket. Astra comes out for two specific jobs.

The first job is adversarial code review on a feature Sol wrote. One model writing and reviewing the same code is like checking your own work, Srivastav said. For an adversarial pass he gives Astra fresh context. He asks it to poke holes in Sol's output.

The second job is a new project or an unfamiliar problem. He turns to Astra to look for gaps he may have missed. Both are high-stakes cases. The Astra bill shows up where the extra accuracy actually changes the answer.

Where Ultrafast fits

Ultrafast is a separate speed tier, not a cheaper tier. Srivastav framed it for real-time use cases: fraud detection, financial investigations and code investigations. More than 300 tokens per second matters there more than another model generation. Astra in Ultrafast costs more than standard Astra, he said.

Sol in Ultrafast costs 1.2 times standard Astra, by his figure. A team paying for Ultrafast on Sol gets more than 300 tokens per second. The price is 20 percent above Astra's standard rate.

What else is new for a developer

Auto-review puts a Luna-level model between an agent and its action. It became free for Codex and ChatGPT subscribers last week. Srivastav said that reaches more than 5 million weekly active users. The policy you write is the gate the agent cannot bypass.

The deprecation window on the API is three months. OpenAI does not retire a model until the next generation is cheaper and more accurate across the board, Srivastav said. GPT-5.5 is still served to API users even though 6.1 Sol is out.

The pattern is a routing argument, not a benchmark. One developer's workflow at one company is a sample size of one. The ratio of adversarial reviews to normal tickets will differ for every team. Srivastav's split sends the hard tickets to the expensive model. The rest go to the cheap one. A team can match the pattern against its own ticket log before the next invoice.

Source

“I don’t use Astra every day”: OpenAI developer lead tells TNS why, The New Stack.