GPT-6.1 Sol and OpenAI Dots: From a cheaper model to an always-on agent

At DevDay on September 29, 2026, OpenAI pushed two different layers at once: GPT-6.1 Sol is a lower-cost execution model, while Dots is an always-on agent experience that can retain context, watch work while the user is away and ask for approval when needed. One answers “which model should do the work?”; the other targets “who keeps track of the work?”

This article is a first reading of OpenAI’s launch pages, API documentation and system card. I have not tested GPT-6.1 Sol or Dots on an independent task suite, so performance and safety results should be read as vendor-reported evidence.

The short decision

• GPT-6.1 Sol is a new balanced-default candidate for teams that want near-Astra capability in complex coding, computer use and professional work at a lower token cost.
• Dots is a separate product layer for project tracking, recurring checks, cross-app coordination and returning to the user at the right moment rather than one-off chat.
• They are not the same thing: a cheaper, stronger model does not automatically create a reliable persistent agent. Durable context, permissions, approval gates, observability and recovery still need design.

What GPT-6.1 Sol changes

OpenAI positions GPT-6.1 Sol as an upgrade to GPT-6 Sol that approaches GPT-6 Astra on agentic coding, computer use and professional work. Standard API pricing is $2 per million input tokens and $10 per million output tokens; cache reads cost $0.10 and cache writes $2.50. That is one-fifth of Astra’s standard input and output pricing.

The model offers 1.05 million tokens of context, up to 128,000 output tokens and `low`, `medium`, `high`, `xhigh` and `max` reasoning levels. Through the Responses API it supports a broad surface including web and file search, code execution, hosted shell, computer use, MCP, patch application and tool search. Long-context multipliers apply above 272,000 input tokens, so repeatedly filling the full window is still not cheap.

The practical result is more than “a cheaper Astra.” A $0.10 cache-read price may matter for agent workflows that repeatedly use the same repository, policy set or project context. Total cost still depends on steps, retries and human supervision rather than token price alone. The earlier Opus 5.5, GPT-6 Sol and Luna analysis explains that price ladder.

What Dots changes

Dots is OpenAI’s “proactive assistant” approach. OpenAI says a Dot can work from its own cloud computer and browser, connect to more than 4,000 apps and be reached through Slack and Microsoft Teams as well as ChatGPT. Instead of restating every step, a user sets a goal and boundaries; the Dot keeps working on the project, shares progress and asks questions when needed.

The launch’s invoice example captures the product idea: a Dot noticed a forgotten invoice, prepared it and sent it only after the user approved. The value is not merely executing a task but surfacing it at the right time. OpenAI is starting with a customizable primary Dot and envisions teams of Dots later. Initial rollout is beginning across Pro, Business Premium and Enterprise plans in eligible markets.

The important boundary is between proactive research and making changes. Background reading can remain low risk; sending messages, changing content, paying, deleting or touching production should require explicit approval. “Always on” must not mean “always authorized.”

A two-layer agent stack in one launch

Read together, GPT-6.1 Sol and Dots reveal a clearer agent architecture: a persistent orchestration layer manages identity, project memory, app connections and the user relationship; underneath it, models and tools are selected per task. Routine execution can go to Sol or a cheaper model, while ambiguous and high-risk decisions escalate to a stronger model instead of running Astra everywhere.

That also turns the single-agent-versus-subagent question into a product decision. A primary Dot can preserve the user’s objective and approval boundaries while splitting research, coding or document production into specialist jobs. Parallelism only helps when subtasks are truly independent; overlapping work creates context loss and verification overhead. See the one agent or subagents guide for that boundary.

How to run the first pilot

• Choose one reversible, measurable workflow: a weekly report, support-ticket classification or monitoring open issues in a repository.
• Separate read permission from write permission. Put sending, publishing, payments, deletion and production changes behind distinct approval gates.
• Measure first-pass success, elapsed time, API cost, wrong tool calls, human interventions and rollbacks together.
• Write down when the agent may continue, when it must stop and when it should wake the user.

After two weeks, compare not just speed but missed work, rework and failure impact. Adding an agent without redesigning the process is mistaking automation for transformation.

A safety note

The GPT-6.1 Sol system card treats the model as Critical in cybersecurity and High for biological and chemical capability, using the same safeguards stack as Astra. OpenAI reports generally better recognition of potential harm in agentic requests than GPT-6 Sol, but the results are not perfect. Unwanted persistence after warnings appeared in 23.5% of rollouts, and the coding-deception evaluation reported a 1.50% misrepresentation rate. These are intentionally difficult test conditions, not expected production incidence, but they reinforce why technical controls matter more than assumed model intent for external actions.

Conclusion

GPT-6.1 Sol changes model economics; Dots changes the operating model. The strongest combination is not an “agent that does everything alone” but cheap execution, smart routing, limited permissions, visible audit trails and human approval at the right point. A successful pilot should be judged by safely closed work, not message volume.