OpenAI's EMEA Lead on Choosing Between Sol and Astra: When Speed and Cost Matter Most

OpenAI held its fourth annual DevDay on September 29 near San Francisco, unveiling more than 20 major announcements spanning ChatGPT, Codex, models, and developer APIs. The releases included GPT-6.1 Sol and a faster Ultrafast tier for GPT-6 Astra, with Ultrafast support for Sol arriving on October 8. The company also introduced OpenAI dots—persistent agents running continuously in the background, powered by GPT-6 Astra and capable of connecting to over 4,000 apps through OpenAI's plugin ecosystem to handle ongoing responsibilities.

These background agents research and monitor information, then propose actions such as completing pull requests, drafting communications, creating documentation, or scheduling meetings. Depending on user-defined rules and the nature of the task, a dot may execute independently, request approval, or defer sensitive work to human review. Additional updates included the Decisions API, which uses predefined answer choices to classify inputs or route requests, and computer use in the Agents API, enabling agents to interact with software interfaces.

Ahead of OpenAI's Dev Day Exchange on November 3 in London, Vaibhav "VB" Srivastav, OpenAI's developer experience and community lead for EMEA, discussed the cascade of features with The Silicon Ledger. Srivastav, based in London, reflected on how the new releases affect the day-to-day work of engineers and the decisions developers face when choosing among competing tools.

Sol Versus Astra: The Intelligence-to-Cost Calculation

Among the four features Srivastav highlighted as particularly important for developers, GPT-6.1 Sol stands out. The model achieves roughly 95 percent of Astra's performance while costing one-fifth as much. "When we shipped GPT-6 Astra, everyone was like, 'Hey, this is a very good model. We love it,' but we couldn't bring it to a price point everyone could afford for this level of intelligence," Srivastav explained. "I don't use Astra every day. I do use GPT-6.1 Sol as my main go-to."

OpenAI's assessment of Sol's capabilities rests on traditional benchmarks including Terminal-Bench 4.0, GDPval, GDP.pdf, and SWE-bench. On both Terminal Bench and SWE-Bench—which measure performance on difficult and long-running tasks—GPT-6.1 Sol performs empirically very close to GPT-6 Astra. The company also dogfoods models internally before release. When Sol entered testing, engineers across product, engineering, marketing, and human resources rapidly adopted it as their primary model, despite having no internal restrictions on which model they could choose.

Srivastav outlined a practical decision rule: Sol handles about 95 percent of most tasks. For new projects or unfamiliar problems, he turns to Astra to identify potential gaps. "If I use GPT-6.1 Sol to work on that feature and then use the same model for code review, it's like checking my own work. I'd be biased toward saying it's good," he said. "For an adversarial review, I give GPT-6 Astra fresh context and ask it to poke holes, take the work apart, and help me make sure it's safe, secure, and gets me the outcome I want."

Speed Tier and Real-Time Demands

The Ultrafast tier addresses time-sensitive applications such as fraud detection, financial investigations, and code investigations. Developers requested something extremely fast for these cases, and Ultrafast mode delivers over 300 tokens per second. GPT-6 Astra supports this tier at a higher price than standard, though still affordable. When Sol was added to Ultrafast, it costs 1.2 times standard Astra pricing, delivering near-Astra performance at over 300 tokens per second for 20 percent more than Astra's standard cost.

Safety Guardrails and Auto-Review

OpenAI defaults to opt-in for any feature that could increase costs for enterprise or business users. On the safety front, as models become more capable and faster, alignment with user intent becomes critical. Auto-review, one of Srivastav's favored features, deploys a lightweight, ultra-efficient model to check actions against a user-defined policy before execution. A developer or organization can specify restrictions—such as preventing the model from submitting work on their behalf, sending emails, or sharing personally identifiable information—allowing these models to run safely and autonomously for extended periods.

Last week, OpenAI made Auto-review free for Codex and ChatGPT subscribers, extending the benefit to over 5 million weekly active users. "We think it's important to let these models run safely," Srivastav noted.

The Decisions API: Speed and Structure

The Decisions API came together four or five days before DevDay. A team of four or five people shipped a production-ready API in preview for the event, then moved it to beta a couple of days later. The API is multimodal, classifying both text and images and returning structured output rapidly. Compared to the traditional approach of providing an LLM with a schema and requesting structured output, the Decisions API runs 10 times faster at the same accuracy, costing $0.10 per million uncached input tokens.

How OpenAI Dots Operate

A dot is a persistent, always-on agent that monitors tools, plugins, and other surfaces a user grants it access to. It identifies what requires attention and orchestrates work to address it. Srivastav's dot connects to Slack, where much of OpenAI's work occurs, as well as to his GitHub, Gmail, and calendar. Periodically, the dot reviews Slack messages, direct messages, mentions, email, and calendar entries, surfacing items he has not yet addressed or tasks it can handle.

With his team distributed globally and work happening around the clock, the dot alerts him to pull requests awaiting review or information requests from colleagues. Previously, Srivastav spent mornings sifting through thousands of Slack mentions and threads. Now he instructs the dot to prioritize certain team members and flag specific categories of work. The dot learns his preferences—for instance, that messages from his manager should appear at the top of his list, and that product launches happening tomorrow or the next day warrant automatic high priority.

The dot also monitors his engineering work, including pull requests. It might inform him that a PR is already approved and ready to merge, or offer to merge it itself. Srivastav can approve the action on his phone, and the dot will merge the PR, verify that tests pass, and confirm the change looks correct in production.

When asked whether AI agents might leave developers feeling isolated, missing informal team interactions like meme sharing, Srivastav offered a contrasting view. He joined OpenAI in December of the previous year, when models were capable but could not reliably act autonomously. Now, with improved models that gather context from multiple sources, he spends less time verifying model behavior and more time with his team. "Every Friday, it checks my calendar and blocks 30 minutes with someone I haven't spoken to in the past three or four days. If I did it manually, I would never stick to it," he said.

Measuring Success Across Multiple Models

With so many recent updates, different teams will pilot different features. Srivastav recommended testing each new model on an isolated use case—a migration, new product, or monitoring task—and measuring the lift. The lift can mean achieving the same outcome at lower cost or delivering a better result. For organizations, the question becomes how much they are willing to pay for a given outcome. OpenAI's model family—GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, and others in development—allows teams to choose based on their cost-performance trade-offs.

The same logic applies to features like Ultrafast. Not every problem requires Ultrafast, but monitoring a major user-impacting migration might. Srivastav recommended running a few problems through this test to determine whether a model or feature helps. OpenAI conducts A/B tests internally and examines product-market fit for each feature and model.

Model Deprecation and Migration Paths

OpenAI has accelerated its model release cycle, raising questions about stability and migration paths for developers building production systems. Srivastav pointed to status.openai.com, where the company publishes availability metrics showing four or five nines of uptime, indicating serious commitment to reliability. The way OpenAI ships models has changed significantly over the past year. Better models enable faster development of the next generation, improving productivity for developers and researchers.

In November of the previous year, OpenAI offered GPT-5.1. The company is now at GPT-6 Astra and GPT-6.1 Sol. "Because we ship faster, we don't deprecate a model until we're certain the next generation is cheaper and more accurate across the board," Srivastav said. For the API, OpenAI provides three months' deprecation notice and a migration path so workflows can become cheaper and better with the next model. The company still serves GPT-5.5 to API users, despite already releasing Sol at version 6.1.

Managing continuous change and migration requires rethinking traditional agile software development. Internally, OpenAI observes that teams build something, then use the next model to improve it—a constant trade-off. If building software to solve a problem within a team, frontier performance may not be necessary. But for handling transactions or features users interact with directly, the best performance and current state of the art become important.

Before AI, teams developed software in sprints toward a launch, then paused to fix issues before repeating the cycle over months or quarters. The same pattern now applies to models. A new generation arrives; teams sprint to use it, then pause. When another model arrives, they use it to improve security and build better features. "I see it as an extension of what developers were already doing," Srivastav concluded.