OpenAI’s developer lead says GPT-6.1 Sol delivers about 95% of GPT-6 Astra’s performance at one-fifth of the price, making the cheaper model his everyday choice. His interview also offers practical details on faster inference, automated policy checks and a new classification API.
OpenAI Watch analysis
What happened
In an interview published by The New Stack on 27 October, OpenAI developer experience lead Vaibhav Srivastav described how his team thinks developers should choose among the company’s models and tools. He said Sol is close to Astra on several benchmarks and is his usual model, while Astra is his choice for unfamiliar problems or a fresh, adversarial review.
Srivastav said GPT-6.1 Sol is also available in Ultrafast mode, which produces more than 300 tokens per second. He put its price at 1.2 times standard Astra, or 20% more than Astra’s standard rate. He also described Auto-review, where a smaller model checks an action against a user-defined policy before it happens. OpenAI made the feature free for Codex and ChatGPT subscribers last week, he said.
The interview includes two further concrete details: OpenAI’s Decisions API classifies text and images into structured outputs, which Srivastav said is ten times faster at the same accuracy; and costs $0.10 per million uncached input tokens. He also described OpenAI dots, persistent agents that can review connected work tools and surface tasks for attention.
Why it matters
This is a useful decision rule rather than another model leaderboard: use a cheaper option for routine work, reserve the pricier model for unfamiliar or adversarial tasks, and pay extra for speed when the job genuinely needs it. The figures and performance comparisons are OpenAI’s claims, as reported in the interview, not independent benchmark findings.
Auto-review and the Decisions API point to a second shift: companies are packaging models not just to generate answers, but to check actions against rules and route information into predictable formats. That can make deployment easier, though the interview does not establish how well these tools perform across customers’ real workloads.
Our read
The most useful advice here is also refreshingly unglamorous: test a model on an isolated task and measure whether it delivers the same result for less, or a better result for the same spend. “Use the biggest model for everything” is a remarkably expensive habit to acquire.
What to watch
- Whether OpenAI publishes comparable independent results for Sol and Astra across real developer workloads.
- How teams use Auto-review, and what actions they choose to block or require approval for.
- Whether the Decisions API’s claimed speed advantage holds up in customer applications.
Discussion spark: Would you trust a cheaper model for everyday work and save the stronger one for review, or is using the same model family for both too much of a blind spot?
Sources and evidence
- “I don’t use Astra every day”: OpenAI developer lead tells TNS why – The New Stack (27 October 2026, 14:30 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.