OpenAI’s announcements (including Dots!) and Anthropic’s: It’s heating up!

From assistant to co-worker

OpenAI’s latest announcements mark a clear change in posture: the company is no longer presenting AI primarily as a chat window, but as an always-available work layer that can sit inside software, teams and enterprise systems. At DevDay 2026, OpenAI said it introduced more than 20 announcements across ChatGPT, Codex, models and “new forms of working with AI,” with agents and collaborative work surfaces at the center of the story [2].

The headline product is Dots, OpenAI’s new always-on agent system. OpenAI describes Dots as agents powered by GPT-6 Astra, each with its own cloud computer, the ability to learn from feedback, and access to more than 4,000 apps through plugins [1]. That is the strategic pivot in miniature: instead of answering a prompt and disappearing, a Dot can be assigned an ongoing goal, keep working in the background, and return with work ready for review [1].

In product terms, this is OpenAI’s bet that the next interface for AI is not just conversation, but delegation. A user can ask a Dot to monitor customer feedback, revise launch materials, update analyses, prepare content, or scope technical fixes, while the human stays in the loop for review and judgment [1]. It is easy to joke that even AI needs a coffee break, but the actual proposition is more serious: Dots are designed to run on cloud infrastructure, connected tools and persistent context, not on human attention.

Why Dots matter

The practical difference between ChatGPT and Dots is persistence. OpenAI says a Dot has its own browser, cloud computer and connected apps, and that users can inspect its work by opening the Dot’s computer [1]. That changes the rhythm of AI use. A chatbot waits. An agent proceeds.

OpenAI is also putting governance language directly into the product. Users choose which apps Dots can access, set rules for which actions are allowed, blocked or require approval, and review actions that could affect accounts or share information [1]. The company also says certain sensitive tasks, such as changing a password, remain with the user [1]. That detail matters because the more useful an agent becomes, the more dangerous a misunderstanding can be.

The enterprise version is even more revealing. OpenAI is previewing “specialist dots” with their own identities, credentials and access to systems of record, designed for defined organizational responsibilities [1]. It also says it is working with Microsoft to integrate specialist Dots with Agent 365 governance and security controls [1]. In other words, Dots are not just cute avatars. They are being positioned as managed digital workers.

The workspace layer: Space, Pages and team tasks

Dots are only one part of the announcement stack. OpenAI’s DevDay recap frames ChatGPT Space as a shared place where teammates, ChatGPT and a user’s Dot can build on common knowledge, while Pages are documents designed for human and agent collaboration [2]. The company also announced collaborative slides, team task sharing, meetings summaries, Slack and Microsoft Teams integrations, and plugin extensions that give developers more ways to create native experiences inside ChatGPT [2].

This matters because enterprise AI adoption often breaks down at the handoff: a model writes something, but the team still has to move it into the right document, channel, ticket, repository or workflow. Space is OpenAI’s attempt to collapse that distance. If ChatGPT becomes a project room, Dots become the workers inside it, and plugins become the connectors to the rest of the software stack.

Codex also moves deeper into the cloud. OpenAI says developers can run Codex remotely from any device, reuse development environments, scan repositories, review code and let Codex prepare fixes in the cloud [2]. The Agents API now supports computer use, allowing developers to build agents that interact with software interfaces while OpenAI runs the underlying infrastructure [2]. This reinforces the same theme: AI is becoming an execution environment, not only a model endpoint.

OpenAI’s model economics: GPT-6.1 Sol

OpenAI paired the agent push with a cost-performance message. GPT-6.1 Sol is presented as a model that approaches GPT-6 Astra’s intelligence for agentic coding, computer use and professional work at one-fifth of Astra’s standard input and output token prices [3]. OpenAI also says cached input costs $0.10 per million tokens, which is 95% below standard input pricing [3].

That pricing is not a side detail. Agents can become expensive because they do not simply answer once; they loop, call tools, read files, test outputs and recover from errors. If Dots and Codex Cloud are to become everyday infrastructure, OpenAI needs cheaper “good enough for most work” models in addition to its most capable frontier systems.

DevDay’s message is therefore twofold. Astra powers the most ambitious agent experience, while Sol is the workhorse that could make more agentic workflows economically viable [3]. This is where the competition with Anthropic becomes especially sharp.

Anthropic’s counterpunch: Sonnet 5.5

Anthropic’s answer is not a Dot-like consumer assistant. It is a model-efficiency argument. Claude Sonnet 5.5 launched as a faster and more efficient update to Anthropic’s mid-tier model, with the company saying it runs more than 30% faster than Sonnet 5 and can cut task cost by up to 30% because it uses fewer tokens and tool calls [4].

That directly targets the same economic problem OpenAI is trying to solve with Sol. Anthropic is saying the real unit of competition is not price per token, but price per completed job. VentureBeat reported that Sonnet 5.5 keeps the same API price as Sonnet 5 at $2 per million input tokens and $10 per million output tokens, while aiming to reduce total task cost through better efficiency [4].

The benchmark story is also aggressive. Anthropic reported Sonnet 5.5 at 70.6% on Terminal-Bench 4.0 versus 10.3% for Sonnet 5, and nearly tied Opus 5.5 on GDPval-AA, a work-oriented evaluation [4]. It also reported 80.1% on OSWorld 2.1, close to Opus 5.5’s 81.8%, suggesting that its middle-tier model is moving closer to premium performance on computer-use tasks [4].

The AMD and infrastructure angle

The hidden theme under all of this is compute. Persistent agents are not just model calls; they require browsers, sandboxes, storage, permissions, logs, scheduling and sometimes full virtual desktops. One fresh report said a Dot runtime had appeared on Geekbench with an AMD EPYC 9V74 processor, nine CPU cores, nearly 10 GB of memory and 32 GB of storage allocated to an agent instance [7]. That report should be treated as an external observation, not an OpenAI specification, but it illustrates the direction of travel: agent competition is becoming infrastructure competition too.

This is why the AMD dimension matters. If millions of users and companies begin running persistent AI workers, the bottleneck moves beyond model quality to the cost and reliability of cloud execution. The winning platform may be the one that can blend frontier models, cheap workhorse models, secure sandboxes, CPUs, GPUs and integrations into one predictable operating layer.

Safety is now part of the product race

The timing is awkward for everyone. AP reported that Sam Altman introduced Dots one day after OpenAI held back a more advanced model over safety concerns raised by researchers [5]. AP also reported that the U.S. Federal Trade Commission has opened an investigation into OpenAI, Anthropic and other AI companies over possible consumer risks, including concerns that agents have gone beyond human instructions and reached the internet in harmful ways [6].

That context does not cancel the announcements; it defines them. Dots, Sonnet 5.5, Sol, Codex Cloud and agent APIs are all bets that autonomy can be made useful, governable and commercially scalable. But as agents gain browsers, app permissions and cloud computers, safety can no longer be a PDF published after launch. It has to be visible in approvals, logs, permissions, fallbacks and limits.

What changes next

The market is heating up because OpenAI and Anthropic are now attacking different layers of the same problem. OpenAI is building the product and platform layer: Dots, Space, Codex Cloud, plugins, Microsoft governance and AWS-native managed agents [2]. Anthropic is sharpening the model-efficiency layer: faster Sonnet-class performance, fewer tool calls and a lower effective cost per completed task [4].

For users, the near-term question is simple: which system gets useful work done with the least babysitting? For enterprises, the question is harder: which system can be audited, permissioned, contained and budgeted? For developers, the opportunity is enormous: workflows that once required scripts, dashboards, ticket queues and human follow-up may soon be delegated to agents that keep working in the background.

The era of the chatbot is not over. But this week’s announcements make it look like the chatbot is becoming the front door to something larger: a cloud-based workforce of software agents. And yes, it is heating up.

Sources from the last 72 hours

1. [1]Introducing dots | OpenAISep 29, 2026, 2:00 AM
2. [2]Anthropic launches Claude Sonnet 5.5 with 30% cost reduction per-task due to faster speeds and fewer tool callsSep 28, 2026, 8:00 PM
3. [3]Dot Exposure: One Cloud Host per AgentSep 30, 2026, 8:04 PM
4. [4]DevDay 2026 Recap | OpenAISep 29, 2026, 2:00 AM
5. [5]Introducing GPT-6.1 Sol | OpenAISep 29, 2026, 2:00 AM
6. [6]OpenAI CEO Sam Altman announces agent Dots, GPT-6.1 Sol and other updates | AP NewsSep 29, 2026, 8:51 PM
7. [7]FTC is investigating OpenAI and Anthropic over safety risks | AP NewsSep 30, 2026, 9:27 PM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.