The latest in AI.

OpenAI: DevDay turns agents into ongoing collaborators

OpenAI announced more than 20 updates at DevDay, with the strongest theme being collaborative work across people, agents and connected applications.

• Dots are always-on agents powered by GPT-6 Astra, with their own cloud computer. They can work across connected apps and continue projects between conversations. Rollout is starting on eligible paid plans.
• GPT-6.1 Sol is positioned by OpenAI as near-Astra performance at one-fifth of Astra's standard token prices, particularly for coding, computer use and professional work. That is the company's claim, rather than an independent verdict.
• ChatGPT Space and Pages provide shared workspaces and documents where teams and agents can collaborate. Collaborative slides are expected in the next few weeks.
• Developer upgrades include cloud Codex, a refreshed CLI, expanded code review and security tools, computer use in the Agents API, and a Decisions API for constrained business decisions.

There is also an important counterpoint: OpenAI withheld GPT-6.1 Astra after safety concerns. That is separate from the GPT-6.1 Sol launch and the existing GPT-6 Astra powering Dots.

UK availability: Dots currently excludes UK users on personal Pro plans. OpenAI says Business Premium is available across supported ChatGPT regions; Enterprise access requires an administrator to enable the beta.

Meta: Muse brings the agent closer to small business

Meta launched Muse for Small Business on 29 September, following its consumer agent announcements and work around smart glasses and the Muse Charm device.

This is the most direct competitive pressure on Dots: personal agents expanding into business and hardware. It also points to a distribution advantage. Agents may arrive through the communication, advertising and devices a business already uses, rather than through a new standalone application.

Google: Gemini 4 is one to watch

DeepMind's chief says Gemini 4 is in refinement, targeting release well before the end of the year. There is no firm date announced.

For businesses, this is a major upcoming model launch rather than an available release. It is a reason to avoid hard-coding a long-term provider decision around today's benchmark or pricing snapshot.

Anthropic: capability with controlled access

Mythos 5.1, announced earlier this month, remains restricted to vetted organisations. Reuters also reports an IPO prospectus targeting a $2 trillion valuation, alongside substantial losses and cloud commitments.

The combination matters: advanced capability is arriving alongside controlled access and a huge financing push. Availability, procurement and governance may be as important to a business decision as raw model quality.

What this means for business

AI is moving from helping employees produce work to taking responsibility for parts of a workflow. The opportunity is greater output from the same team, but the business needs to define what AI can do, what it can access and who checks the result.

The biggest advantage will come from redesigning a few processes well. Buying licences alone will not deliver much if employees still move information manually between disconnected systems. Clean data, clear responsibilities and access to the right tools become more valuable.

The business questionCan an agent complete a meaningful part of the task through the systems you already use, with a level of authority you can accept?

Start with one reversible workflow

Pilot one frequent, measurable workflow with reversible actions. Preparing a weekly sales report or triaging support requests is a better starting point than giving an agent unrestricted authority over payments, contracts or external commitments.

1. Record the current time, error rate, review effort and cost per completed task.
2. Give the agent only the systems and permissions needed for that workflow.
3. Keep approval for payments, contractual commitments and sensitive external communications.
4. Review the results with the people who own the process, then expand only where the evidence supports it.

OpenAI's fundraising target is also worth separating from product news: it is reportedly seeking another $30 billion at a $1.4 trillion valuation. That is a reported fundraising target, not a completed deal.

Treat these announcements as a shortlist of things to test. They do not yet establish which supplier will deliver the best results in your business. The competitive focus is shifting towards who can reliably complete ongoing work across your apps.