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 alongside 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.
These agents monitor and research autonomously, then suggest or execute actions such as completing pull requests, drafting communications, or scheduling meetings, with human review built into sensitive decisions. Additional updates included the Decisions API for classifying inputs and routing requests using predefined choices, and computer use capabilities within 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 wave of features and their practical implications for engineering teams.
Sol versus Astra: performance and cost
Srivastav highlighted GPT-6.1 Sol as a cornerstone release. The model delivers up to 95% of Astra's performance at one-fifth the cost. "I don't use Astra every day. I do use GPT-6.1 Sol as my main go-to," Srivastav said. When GPT-6 Astra launched, adoption was strong but pricing limited accessibility. Sol was designed to bring comparable intelligence to a price point accessible across all plans.
OpenAI's evaluation of Sol's "near-Astra intelligence" rests on traditional benchmarks including Terminal-Bench 4.0, GDPval, GDP.pdf, and SWE-bench. Both Terminal Bench and SWE-Bench serve as strong indicators of model performance on difficult, long-running tasks, with GPT-6.1 Sol testing empirically very close to GPT-6 Astra. Internal dogfooding before release proved decisive: when OpenAI staff could choose any model without restrictions, they rapidly adopted Sol as their daily driver across engineering, product, marketing, human resources, and adjacent teams.
When to choose each model
For most tasks, Sol reaches approximately 95% capability, Srivastav explained. On new projects or unfamiliar problems, he turns to Astra to catch potential gaps. The decision hinges on use case and risk tolerance. "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," Srivastav noted. For adversarial review—catching security flaws or design weaknesses—he uses Astra with fresh context to stress-test the work.
Ultrafast mode targets time-sensitive scenarios: fraud detection, financial investigations, and code analysis. Developers requested extremely fast outputs for these cases. Ultrafast delivers over 300 tokens per second with both GPT-6 Astra and GPT-6.1 Sol, priced at 1.2 times standard Astra cost. This positions Sol in Ultrafast at near-Astra performance exceeding 300 tokens per second for 20% more than standard Astra pricing.
Safety guardrails and Auto-review
Any feature adding costs for enterprise or business users defaults to off, requiring explicit opt-in. On safety, Srivastav emphasized Auto-review, a lightweight model-level check that validates agent actions against user-defined policies before execution. Organizations and individual developers can specify rules—such as preventing model submissions on their behalf, email sending, or personally identifiable information sharing—allowing models to run safely and autonomously for extended periods.
OpenAI made Auto-review free for Codex and ChatGPT subscribers last week, benefiting over 5 million weekly active users. The company views this as essential for enabling safe autonomous operation at scale.
The Decisions API
The Decisions API came together four to five days before DevDay, with a team of four or five engineers shipping a production-ready preview API that moved to beta within days. The multimodal API classifies text and images, returning structured output 10 times faster than traditional LLM approaches at equivalent accuracy, costing $0.10 per million uncached input tokens. Traditionally, developers would provide an LLM with a schema and request structured output; Decisions API streamlines this workflow significantly.
How OpenAI dots function
A dot is a persistent, always-on agent with access to tools, plugins, and surfaces a user designates. It scans these resources to identify what requires attention and orchestrates work to completion. Srivastav's dot connects to Slack, GitHub, Gmail, and his calendar. Periodically, it reviews Slack messages, direct messages, mentions, email, and calendar events, surfacing unaddressed items or tasks it can handle.
With his globally distributed team, someone works around the clock. His dot alerts him when teammates tag him for pull request review or when they await information about upcoming model launches. Previously, Srivastav spent mornings sifting through thousands of Slack mentions and threads. Now he instructs his dot on team members and their roles, requesting notifications when specific people flag particular items. The dot learns context: messages from his manager rank higher; product launches tomorrow automatically become high priority.
The dot also monitors engineering work, identifying approved pull requests ready to merge or offering to merge them directly. Srivastav can approve via phone, and the dot executes the merge, verifies tests pass, and confirms production status.
AI and team connection
When asked whether AI creates isolation and diminishes informal team bonding—like Slack meme sharing—Srivastav reported the opposite experience. He joined OpenAI in December of the previous year when models were capable but unreliable for autonomous work. Now models retrieve context from multiple sources and operate independently. Previously, he spent significant time ensuring models performed correctly, leaving little time for team interaction. Today, staying on top of work requires less time sifting through email and Slack, freeing capacity for team engagement.
Srivastav asked his dot to facilitate weekly catch-ups with teammates he hasn't spoken to recently. "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. This systematic approach to connection would not happen through manual effort.
Measuring success with new models
With multiple teams piloting different features, measurement becomes critical. Every new model should demonstrate overall improvement based on metrics and internal observations, though performance gains vary by use case. Srivastav recommends testing on isolated use cases—a migration, new product, or monitoring task—and measuring the lift: either the same outcome at lower cost or better results at similar cost.
Organizations must determine how much they will pay for a given outcome. This is where the model family matters: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, and others in development each serve different cost-performance points. The same logic applies to features like Ultrafast—not every problem requires it, but critical user-impacting migrations may justify the expense. Testing a few problems through this framework reveals whether a model or feature delivers value. OpenAI runs A/B tests internally and evaluates product-market fit for each feature and model.
Model deprecation and migration paths
OpenAI has accelerated its model release cycle. Stability and end-user experience remain priorities, with status.openai.com showing availability at four or five nines. The company's shipping approach has shifted significantly over the past year. Better models accelerate development of the next generation, improving productivity for developers and researchers.
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.
Vaibhav Srivastav
In November of the previous year, GPT-5.1 was current; now GPT-6 Astra and GPT-6.1 Sol are available. The API provides three months' deprecation notice and a migration path enabling workflows to become cheaper and better with the next model. OpenAI still serves GPT-5.5 to API users despite having released 6.1 Sol.
Managing continuous change
Developers managing constant model updates and migrations face pressure to adapt traditional agile development cycles. Internally, OpenAI observes that teams build something, then use the next model to improve it—always a trade-off. Not every capability requires frontier performance. Software solving internal team problems may not need state-of-the-art capability. But systems handling user transactions or directly serving products demand current best performance.
Before AI, teams developed software in sprints toward launch, then paused new features to fix issues, repeating the cycle over months or quarters. The same pattern applies to models: a new generation arrives, teams sprint to adopt it, then pause. When another model arrives, teams use it to improve security and build better features. "I see it as an extension of what developers were already doing," Srivastav said.