For most of the generative AI era, working with a language model meant a simple loop: you prompt, it responds, and the session ends. OpenAI GPT-6 is the clearest sign yet that this loop is breaking down. With a three-tier model family, cheaper mid-range options, and a new class of always-on agents called Dots, OpenAI is betting that the future of AI is less like a chatbot and more like a colleague who keeps working after you log off. This article breaks down the GPT-6 AI models, the features that matter to technical teams, and the practical trade-offs to weigh before adopting them.
GPT-6 is not a single model but a lineup. Understanding which tier does what is the first step toward using it well, so let's start with the release timeline and positioning.
OpenAI began rolling out GPT-6 Astra on September 3, positioning it as its most powerful model and claiming it can operate an entire computer autonomously, with a mechanism to shut it down if it strays beyond its intended use. OpenAI says Astra came from its largest-ever training run, using more than 100,000 GPUs at its Stargate site in Texas. Access was staged: it went first to business users in OpenAI's Daybreak program, with a version carrying added cybersecurity guardrails planned for paid ChatGPT users.
On reasoning benchmarks, the gains are notable. On ARC-AGI-3, Astra scored 62.7% with the standard harness, compared with 30.2% for Claude Opus 5 and 7.8% for GPT-5.6 Sol. That capability comes at a premium, however, since Astra is priced at $10 per million input tokens and $50 per million output tokens, matching Anthropic's pricing for Claude Fable 5.1.
Recognizing that few teams can run every workload on a flagship, OpenAI followed up quickly. GPT-6 Sol and GPT-6 Luna launched on September 22, 2026, at roughly half the price of their GPT-5.6 equivalents, with Sol at $2 input and $10 output per million tokens and Luna at $0.10 and $0.50. OpenAI described these as permanent prices rather than promotional or introductory pricing.
The division of labor is straightforward. Sol is aimed at coding, agent tasks, and business workflow automation, while Luna targets high-frequency jobs such as extracting fields from documents or summarizing tickets. Both models also offer substantial headroom, as they support a 1.05M-token context window and reasoning effort settings from none to max.
One week later came an iteration. GPT-6.1 Sol is an update aimed at coding, computer use, and professional tasks, and OpenAI says it approaches Astra on several evaluations at one-fifth of Astra's standard token prices, with cached input at $0.10 per million tokens. Notably, it is available in ChatGPT Work and Codex but not yet in regular Chat. With the model lineup clear, the more interesting story is what OpenAI built on top of it.
If GPT-6 is the engine, Dots are the vehicle. They represent OpenAI's most ambitious move into AI agent technology to date.
Announced on September 29, 2026, at OpenAI DevDay, Dots are always-on agents that live inside ChatGPT and run on GPT-6 Astra. Each Dot gets its own cloud computer and browser, connects to more than 4,000 apps through OpenAI's plugin ecosystem, and keeps working between conversations. A useful analogy is the difference between hiring a consultant for a single meeting and bringing on a staff member with their own desk, laptop, and standing responsibilities.
The technical shift is subtle but important. Most agent frameworks treat work as bounded runs, where a request starts, tools execute, and the run ends, whereas Dots introduce a long-lived agent object whose conversations act as entry points into the same workspace. Users can reach that workspace from several surfaces, since Dots communicate through ChatGPT, Slack, and Microsoft Teams, and OpenAI says they will adapt to a person's preferences and working style over time.
Governance is built into the loop as well. OpenAI says Dots use built-in rules to decide when they can act independently and when they need approval. OpenAI is also looking beyond personal assistants, as specialized Dots in company trials would have their own organizational identities, credentials, system access, and defined job responsibilities. These capabilities become clearer through concrete scenarios.
The strongest early use cases cluster around continuous, multi-step work that previously required a human to babysit the process.
Continuous bug triage. One example OpenAI gives is pointing a Dot at a stream of customer feedback so it can work through issues on its own schedule, writing fixes and building requested features.
Pull request workflows. Through Codex integration, Dots can triage bugs, scope builds, and ship pull requests for developers.
High-volume data processing. A support operation can route ticket classification and extraction through Luna, reserving Sol for anything requiring multi-step reasoning.
Legacy system automation. Because Dots operate through browsers and connected apps, companies may not need to rebuild existing software as AI-native applications to use them.
Codex and the Developer Toolchain
For teams focused on AI coding tools, DevDay delivered meaningful upgrades. Codex can now run in cloud environments as well as on local machines, the Codex CLI gained voice input and an /agents interface for delegating and monitoring multiple tasks, and Codex Security Cloud can scan repositories, investigate findings, remove duplicates, and prepare fixes.
Builders who want this capability inside their own products get a direct route too. The Agents API brings OpenAI's managed Codex harness to developers in public beta, and that same foundation powers Codex, Work, and Dots. It adds computer use for driving software through graphical interfaces, along with tool search, tool calling, and context compaction. Taken together, these pieces explain why autonomous AI workflows are the headline benefit of this release.
The advantages of GPT-6 and its surrounding ecosystem fall into a few clear categories.
Tiered cost control. Teams can match each workload to the right price point, from Luna for bulk tasks to Astra for frontier reasoning.
Reduced supervision overhead. Persistent agents can carry multi-day tasks without constant re-prompting.
Broad integration. Thousands of connected apps mean Dots can work inside existing toolchains rather than replacing them.
Improved reliability. According to one independent analysis, Sol makes roughly half as many factual errors as GPT-5.6 Sol.
Limitations and Important Considerations
No release this significant is without caveats, and expert users should evaluate several before committing.
Security risk is the most serious. GPT-6 is OpenAI's first model to reach the highest risk tier on its own cybersecurity evaluation scale, and automated monitoring can interrupt user tasks, including work unrelated to cybersecurity. That second point matters for workflow reliability, not just safety.
Benchmark gains are also uneven. At max effort, GPT-6 Sol scores below GPT-5.6 Sol on DeepSWE (68.8% versus 72.7%) and on OSWorld offline (64.4% versus 66.2%). Teams pinned to the older model for hard coding tasks should rerun internal evals before migrating.
Dots, meanwhile, launched with constraints. At launch, a Dot cannot initiate calls or have its own standalone email address, and OpenAI does not treat "always on" as an uptime guarantee. Rollout is starting on Pro, Business Premium, and Enterprise plans in eligible markets. Developer sentiment is mixed as well, since some questioned whether the releases were a significant advance over capabilities already available on existing agent platforms.
Competition is the final consideration. Anthropic shipped Claude Opus 5.5 shortly before the Sol release, and Sol lands at half of Opus 5.5's price. In persistent assistants, Dots enter a growing market that includes Meta's Muse. Price, ecosystem fit, and your own evaluation results should drive the decision rather than headline benchmarks alone.
Engineering teams running large codebases will get the most immediate value from GPT-6.1 Sol and cloud Codex. Operations and support teams with high-volume, well-defined tasks should test Luna first. Organizations exploring persistent automation across many SaaS tools are the natural audience for Dots, provided they can invest in permission design and oversight.
OpenAI GPT-6 marks a shift from models you query to agents you delegate to. The three-tier lineup of Astra, Sol, and Luna gives teams real control over the cost-capability trade-off, while Dots and the Agents API push AI toward persistent, cross-application work. The key takeaways are to match model tier to workload, run your own evals instead of trusting launch charts, and treat agent permissions as a first-class engineering concern. As a next step, pilot one bounded, recurring workflow with clear approval gates, measure the results against your current process, and expand only once the agent has earned your trust.
Final word count: approximately 1,320 words
I can also put this into a file if you'd like a copy outside the chat.