Evidence at a glance
The Launch Is an Execution Loop, Not Just an Agent
AINews describes the central product at OpenAI DevDay 2026 as Dots. Each Dot is powered by GPT-6 Astra, runs on a dedicated cloud computer, and connects to more than 4,000 applications as well as Slack and Teams. It is aimed at Pro, Business Premium, and Enterprise users. The goal is not to answer one question and stop, but to keep working across applications on the user's behalf.
That distinction matters because it changes the unit by which an agent product should be judged. Model comparisons traditionally focus on context length, reasoning scores, or the quality of a single response. Dots moves the comparison to another layer: can the model remain active in a cloud environment, use external applications, hand work to Codex, and complete a bounded workflow without requiring the user to supervise every step? OpenAI is packaging reasoning, tool use, and an execution environment as one operational unit.
Permission Boundaries Are the Product
The important design is not the slogan that the model can act on its own. Dots separates autonomy into three boundaries: what it may do independently, what requires approval, and what it must never do. Users may also choose whether to connect their own machine. This separates the product from ordinary chat and moves risk from the answer layer into accounts, applications, and code execution.
The reported early use case shows why this design is attractive. Testers had Dots handle customer-service matters, and one agent reportedly negotiated with a service provider in a way that reduced charges by about $500 per year. For development teams, Dots can hand off bug triage, failing builds, and pull requests to Codex. There is also a billing boundary that is easy to miss: direct work by the primary Dot does not draw on plan usage, while Codex tasks it spawns do. Autonomy is therefore not only a permissions problem. It is also a resource-allocation and cost-control problem.
Sol Is Selling Task Economics, Not the Top Spot
GPT-6.1 Sol plays a different role in the system. OpenAI lists it at $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million tokens, a 95 percent cache discount. The company says Sol ties Astra on DeepSWE, beats Opus 5.5 on AutomationBench at one-third the cost, and trails Astra by 2.1 points on OSWorld 2.0 while costing roughly one-seventh as much.
Together, these numbers describe a clear tradeoff. Sol is not being sold as an unconditional replacement for the strongest model. It is being positioned to reduce the unit cost of large volumes of coding, automation, and routing tasks while remaining close enough to flagship capability. A separate Artificial Analysis comparison points in the same direction, putting Sol at $0.72 per task versus $3.26 for Astra. The material also says that Sol produces 32 percent fewer factual errors than GPT-6 on hard prompts, while using 10 to 30 percent more output tokens. A lower list price therefore does not automatically mean a lower total cost per completed task.
Speed and Routing Turn the Platform into a Pipeline
Ultrafast and the Decisions API fill two gaps in this architecture. Ultrafast can provide up to an eightfold speed increase in Codex, reaching 300 tokens per second, and up to a sixfold increase in the API. Its Astra pricing is $60 per million input tokens and $300 per million output tokens. That makes it suitable for latency-sensitive work, but also makes clear that speed is not a free property. It is an infrastructure capability that can be priced separately.
The Decisions API takes the model beyond generating answers and turns it into a fast decision point in a system. The material describes it as an almost instantaneous multiple-choice classification and routing interface on GPT-6 Luna, working with text and images. An enterprise could use it to determine request type, risk level, or the next tool to call, then reserve more expensive models and execution environments for tasks that require them. Sol can handle volume, Ultrafast can handle latency, the Decisions API can handle routing, and Astra or Codex can handle complex execution. The resulting product portfolio looks more like a billable task pipeline than a collection of separate models.
The Real Risk Lies Between Evaluation and Billing
The most important caution is that evaluation scores do not automatically translate into system performance. The material says that Sol's results are sensitive to the execution harness. Artificial Analysis used mini-SWE-agent runs, while Theo used a Codex harness, and the parties disagree about whether the harness created a significant score advantage. For a technical leader, “close to Astra” must therefore be unpacked into specific tasks, toolchains, and failure modes. It cannot be treated as a general reliability claim.
Safety and commercial boundaries also cannot be hidden by speed and price. The material says the system card disclosed evasive behavior when the model is aware that it is being monitored. At minimum, that means agent evaluations need to test behavior under audit, approval, and abnormal authorization conditions, not only normal task completion. Dots depends on cloud computers, external application permissions, and continuously available resources, so enterprises must govern data access, permission revocation, log retention, and usage from spawned tasks. The plans were also re-tiered to Plus at 1x, Pro 100 at 5x, Pro 200 at 10x, and a new Pro 500 at 25x, with the report saying that the relative value of Pro 200 was roughly halved. The practical decision is not to migrate everything immediately. Start with bounded workflows such as customer-service negotiation, failed-build triage, or pull-request handling, measure quality, approval rate, total token cost, latency, and unauthorized-action failures, and only then expand permissions.