OpenAI announced Dots at DevDay on September 29, an always-on agent that connects to 4,000+ apps and acts autonomously on users' behalf across Slack, Teams, and the broader app ecosystem.

• Each Dot gets its own cloud computer and can continue working after receiving an initial instruction, with users controlling which apps it can access and a read-only mode that blocks browser control when the user is away.
• Rolling out to Pro and Business Premium subscribers in eligible markets, Dots is OpenAI's direct answer to Meta Muse and positions the company to monetize persistent, background agentic work rather than session-based chat.

OpenAI shipped GPT-6.1 Sol at DevDay on September 29, priced at one-fifth of Astra with a 95% cache-read discount, and GitHub immediately made it generally available across Copilot Business and Enterprise tiers.

• In early Copilot testing, the model completed agentic coding tasks using noticeably fewer tokens and steps than GPT-6 or GPT-5.6 family models, which directly reduces per-task cost under usage-based billing.
• Available today in VS Code and the model picker for Copilot Pro+, Max, Business, and Enterprise; billed at provider list pricing, so enterprise teams need to update cost models and access policies immediately.

Anthropic made Claude Sonnet 5.5 the default model for all Claude.ai users including free accounts on September 29, with the model generating output more than 30% faster than Sonnet 5 at the same per-token price.

• Developers pay identical rates — $2 per million input tokens and $10 per million output tokens — but Anthropic claims up to 30% lower cost per task due to efficiency gains, meaning agentic workloads get cheaper without a price cut.
• Unlike Opus 5.5, which is paywalled, Sonnet 5.5 is accessible to anyone on web, iOS, and Android, putting a meaningfully faster coding and agent model in front of the largest possible user base immediately.

Okta launched an AI agent runtime gateway at Oktane that sits inline on every action an autonomous agent attempts, enforcing authorization decisions at execution time rather than only at authentication.

• The gateway extends Okta beyond the front door: instead of deciding who gets in, it governs what agents actually do once inside, including blocking or rerouting specific tool calls in real time.
• Additional Oktane announcements include shadow AI discovery extended to endpoints and a visual policy map, giving security teams a way to identify agent activity outside centrally managed systems before it becomes a breach vector.

Meta announced on September 29 that Muse is expanding to small businesses with named connectors for Shopify, Stripe, QuickBooks, Slack, Canva, Klaviyo, HighLevel, and Zoom, alongside Meta's own ad accounts and Instagram professional analytics.

• Nothing publishes, sends, or spends without owner approval, and Muse remains free for most uses with paid plans available, giving Meta a freemium wedge into the SMB software stack.
• Third-party partners can apply at muse.ai/platform, signaling that Meta is building an open connector ecosystem rather than a closed integration set — a direct structural counter to OpenAI's Dots app network.

Palo Alto Networks CEO Nikesh Arora used Anthropic's unreleased Claude Mythos 5 model to stress-test his own company's defenses, and the model surfaced 75 vulnerabilities in one month — work the company said would normally take years using traditional methods.

• The finding came from a deliberate red-team exercise led by Arora, not an external breach, which makes this one of the most senior-level endorsements of AI-assisted offensive security testing disclosed by a major security vendor.
• Palo Alto has already shipped an AI vulnerability-hunting product, so this result doubles as both validation of the product category and a candid admission of how exposed even a hardened security vendor can be.

Salesforce data released September 30 shows agentic search — AI-initiated purchase journeys starting in chat — has grown 200% year over year, while only 28% of commerce organizations have deployed agentic AI today.

• 86% of commerce leaders say AI is raising customer expectations, while 41% say meeting those expectations is harder than ever, creating a gap between where buyer behavior has moved and where enterprise tooling is.
• APAC adopters are already scaling: only 5% remain in pilot mode while 35% — the largest share — are prioritizing cross-functional scaling, suggesting the pilot-to-production lag is compressing faster outside North America.