OpenAI lead explains when to use Sol, Astra, and persistent agents

At its 2026 DevDay in San Francisco, OpenAI unveiled more than 20 updates spanning new models, developer tools, and autonomous agents. Vaibhav Srivastav, the company’s EMEA developer experience lead, clarified how engineers should navigate these options, emphasizing cost-effective model selection and the practical utility of persistent background agents.

What happened

OpenAI introduced GPT-6.1 Sol and an Ultrafast tier for GPT-6 Astra during its fourth annual DevDay on September 29. The company followed up with Ultrafast support for Sol on October 8. These releases aim to provide developers with granular control over performance, speed, and cost. Srivastav explained that while Astra represents the frontier of capability, Sol offers near-equivalent performance for most daily tasks at a significantly lower price point.

The announcements also included OpenAI dots, which are persistent agents capable of running continuously in the background. These agents connect to thousands of applications via OpenAI’s plugin ecosystem to monitor workflows, draft communications, and manage pull requests. Additionally, OpenAI launched the Decisions API for rapid structured output classification and expanded computer use capabilities within the Agents API, allowing models to interact directly with software interfaces.

Srivastav noted that internal adoption patterns at OpenAI mirror what external developers might experience. Many employees have switched to GPT-6.1 Sol as their primary daily driver because it balances intelligence with affordability. The company encourages teams to test specific use cases against different models to determine the optimal balance of cost and outcome, rather than defaulting to the most expensive option for every task.

How it works

GPT-6.1 Sol achieves approximately 95% of GPT-6 Astra’s performance on benchmarks like SWE-bench and Terminal-Bench 4.0, but costs one-fifth as much per token. For tasks requiring adversarial review or deep security analysis, Srivastav recommends using Astra to identify loopholes that Sol might miss. The Ultrafast tier delivers over 300 tokens per second for both models, catering to real-time needs like fraud detection or live code investigation, though at a premium price.

OpenAI dots function as always-on agents that aggregate context from connected tools such as Slack, GitHub, Gmail, and calendars. They prioritize tasks based on user-defined rules, such as flagging messages from managers or identifying urgent product launches. A lightweight Luna-level model performs auto-reviews against safety policies before any action is taken, ensuring that agents do not send emails or share sensitive data without explicit approval. This guardrail system is now free for Codex and ChatGPT subscribers.

Key details

• GPT-6.1 Sol provides 95% of Astra’s performance at 20% of the standard cost.
• Ultrafast mode supports both Sol and Astra, delivering over 300 tokens per second.
• OpenAI dots are persistent agents that connect to over 4,000 apps via plugins.
• The Decisions API classifies text and images ten times faster than previous methods.
• Auto-review uses a lightweight model to check agent actions against user policies.
• OpenAI provides a three-month deprecation notice for older API models.

Why it matters

For engineering leads and founders, the distinction between Sol and Astra offers a clear path to optimize AI spending. Using the highest-tier model for every task is unnecessary and expensive. By reserving Astra for complex, high-stakes reviews and using Sol for routine coding and drafting, teams can maintain high output quality while controlling costs. This tiered approach allows organizations to scale AI usage without proportionally scaling their budget.

The introduction of persistent agents like dots shifts the paradigm from reactive prompting to proactive assistance. Instead of manually sifting through thousands of notifications, developers can delegate monitoring and routine administrative tasks to agents. This reduces context-switching fatigue and frees up time for high-value engineering work and team collaboration. However, it requires careful configuration of safety guardrails to prevent unauthorized actions.

What you can do

• Audit your current AI usage to identify tasks that can shift from Astra to Sol.
• Implement GPT-6.1 Sol for daily coding, drafting, and routine code reviews.
• Reserve GPT-6 Astra for adversarial testing, security audits, and complex problem-solving.
• Configure OpenAI dots to monitor specific channels and repositories for urgent items.
• Define strict auto-review policies to prevent agents from sending data or merging code without approval.
• Test the Decisions API for high-volume classification tasks to reduce latency and cost.