OpenAI DevDay 2026: the platform takes shape

OpenAI held DevDay in San Francisco on September 29, 2026; the opening keynote fell on September 30 in Beijing. Its official recap lists more than 20 announcements. Some are new products, others extend existing tools, and several remain previews or promises of later availability. Treating all of them as fully launched would miss an important distinction. [1]

Our reading is that OpenAI is joining model intelligence to execution, shared work and distribution. A model answers; an agent acts; a workspace retains the result; a plugin brings in another application; a subscription or marketplace helps pay for it. The strategic question is how much of that chain developers and customers will choose to place inside one platform.

GPT-6.1 Sol is the new model upgrade. Standard short-context API pricing is $2 per million input tokens, $0.10 for cached input and $10 for output; long-context rates differ. It is available through the API, ChatGPT Work and Codex, but was not yet in ordinary Chat at publication. OpenAI reports performance approaching Astra on several evaluations; that is a provider-reported result under stated test conditions, not proof of equal performance on every workflow. [2] [22]

OpenAI reports that Sol matches Astra on DeepSWE 1.1 at roughly one-fifth the task cost, exceeding GPT-6 Sol’s best score by 6.4 percentage points at lower reasoning effort. On OSWorld 2.0’s offline set, at maximum effort, it comes within 2.1 points of Astra at roughly one-seventh the task cost. Astra still leads the tested scientific-research models. This is a significant capability–cost shift, rather than evidence that Sol wins every task. [2]

These are provider-reported results from specified evaluations. Research/API settings can differ from production ChatGPT, and competitor results were drawn from public reports. For coding and agents, our inference is that stronger routine workers could become affordable enough to leave budget for review and retries, with difficult work escalated separately. The useful question is where Sol preserves quality in your workflow, not whether one benchmark permits replacing every model. [2]

Ultrafast adds a premium speed tier. Astra Ultrafast is available, while Sol Ultrafast is coming later. Private Intelligence adds privacy-related infrastructure: Private Safety Processing supports automated safety review with protected customer content, while Private Inference is announced for a fall preview. These are different capabilities, not a blanket claim that every OpenAI workload already runs confidentially. [1] [7]

For developers, the useful metric is cost per accepted result. Lower token prices may permit more checking and retries; faster generation may shorten iteration. Neither removes tool latency, incorrect actions or review work. A cheaper model deserves a trial against the actual workload before it becomes the default.

dots introduces persistent agents that remember context and take on ongoing work through connected tools and computers. Rollout is limited by eligibility and workspace controls. Codex Cloud moves coding tasks into reusable cloud environments, where work can continue while a local computer sleeps. Cloud execution is distinct from remotely controlling a local machine that still needs to be online. [3] [4]

The refreshed Codex CLI adds voice interaction and an /agents view. Code Review brings review workflows into the desktop app, with cloud automatic reviews; GitLab support is still preview. Codex Security Cloud scans connected GitHub repositories and monitors commits. Finding a potential issue or preparing a patch is not the same as proving a repository secure. [1] [17] [18] [19]

Decisions API focuses Luna on user-defined questions with finite answers; it is in limited preview. The existing Agents API gains computer use, including a hosted browser. Bedrock Managed Agents is a separate AWS-native preview. These offers cover different jobs: making a constrained choice, running a broader agent, or operating within AWS governance. [1] [5] [6]

ChatGPT Space and Pages provide shared context and editable documents for people and agents. At launch, Space supports creation and editing on web and desktop; mobile supports finding, reading and sharing. Collaborative slides, including concurrent editing and export, are still coming soon. Do not infer that a demonstration means every collaboration feature is already available. [10]

Teams and Team Tasks organize colleagues and recurring work. @ChatGPT in Slack and Microsoft Teams brings assistance into conversations. The Meetings plugin turns meeting notes into personalized summaries and action items; it is beta on macOS for Pro and Business, with broader support still pending. Shareable profiles make creations easier to discover. [1] [11] [12]

A lasting workspace changes the acceptance problem. A useful answer can become shared material that others act on later. Teams need to distinguish source material, an agent proposal, a reviewed decision and an action already taken. Real-time editing alone does not supply that distinction.

Plugin extensions add surfaces such as sidebar panels and file viewers. OpenAI also announced improvements to plugin creation, submission and discovery. Sites can host supported plugins, and MCP Events can trigger automations from connected applications. Availability differs by surface: extension support on the web for Free and Go is still coming soon, and composer mentions are desktop-only. [1] [8] [9] [21]

Sign in with ChatGPT combines identity with optional plan usage in participating tools. Eligible Plus and Pro requests consume the existing Work and Codex allowance, rather than creating an unlimited new pool. A third-party service may still charge its own fees. This does not make ordinary chat quota a general-purpose API key for any VPS program. [13]

The new Pro 500 tier expands usage and includes Ultrafast. OpenAI Marketplace lets eligible enterprise customers direct part of an existing commitment toward approved partner software. Marketplace participation has commercial eligibility and terms; it is not a promise of free software. Together, these changes make distribution and procurement relevant alongside model access. [20] [14]

TypeSafe AI introduced Jev on September 15, before DevDay. Its System One approach takes unstructured state and produces predefined typed decisions with probabilities. Decisions API targets a visibly related need: classification, routing and choosing a next action from finite answers. This is a meaningful overlap in the kind of interface software consumes. [15] [1]

The difference matters too. TypeSafe describes a specialized architecture, parallel sampling and Reinforcement Learning for Calibrated Decisions. OpenAI describes Decisions as applying Luna to bounded questions. The public material reviewed here does not establish that OpenAI reproduced TypeSafe’s architecture, training method or all output semantics. Earlier release plus functional similarity is insufficient evidence of copying. [15] [1]

For a buyer, compare error rates on ambiguous cases, calibration, abstention, latency and total cost using the same workload. A response can perfectly match its schema and still select the wrong answer. We have not reproduced either provider’s benchmarks or run a head-to-head evaluation.

Meta introduced Muse on September 8. Persistent work, a personal cloud computer and browser, remembered goals, proactive assistance and approval before sensitive actions make Muse a clear product comparison for dots. The resemblance is substantial enough to discuss without turning it into an allegation. [16] [3]

Their distribution differs. Muse centers on its app and WhatsApp and began a US rollout with free use and subscription options. dots connects to ChatGPT’s work environment and channels such as Slack and Microsoft Teams, with paid-plan and enterprise rollout conditions. Those differences affect how a user delegates work and how an organization governs it. [16] [3]

The sources examined provide no verified evidence that dots copied Muse’s code, protected design or internal system. Nor does similar product positioning prove independently invented implementations. What we can support is a later entrant competing on a closely related experience. Claims of copying would need evidence beyond a launch calendar and a feature checklist.

First, narrow AI features face stronger competition from platform bundles. A standalone tool that only generates a summary or wraps a model may find its feature inside a workspace. Domain knowledge, trustworthy integration, reviewed outputs and a complete business process become more important possible advantages. This is our strategic inference, not a forecast that every specialist will disappear.

Second, distribution may become a more important choice for developers. Native plugins and shared spending can reduce adoption friction, while increasing dependence on discovery rules, permissions and platform economics. A team should measure whether the platform brings new customers, rather than confuse inclusion in a directory with demand.

Third, cloud agents make ongoing operation easier to provide, but concentrate more operational responsibility in the execution platform. Enterprises will still need verifiable permissions, evidence, exportable work and a way to stop or move tasks. Better models expand the opportunity; they do not determine who owns the workflow or whether the result is reliable.

Choose one bounded pilot. For an engineering team, compare Sol with the existing model on representative bugs and code changes. For an operations team, compare a decision service on labeled routing cases, including low-confidence examples. For a content team, test whether shared documents preserve the source, review decision and final version across a real handoff.

Record availability, authorization, cost and acceptance criteria separately. Check the account’s actual feature access; keep a baseline; measure end-to-end completion and human corrections. Decide which data and actions belong in the platform and which remain in existing systems. Keep an exit path before transferring a recurring responsibility.

The most consequential part of DevDay may be the connection between the announcements. OpenAI is offering more of the place where work starts, runs, is stored and is purchased. Jev and Muse show that important ideas also emerge outside that platform. The useful response is to evaluate both the products and the dependency being created, with the same care.