OpenAI shipped GPT-6 Sol and GPT-6 Luna on September 22, adding cost-optimized tiers to the GPT-6 lineup already anchored by Astra.

• Available now in the API, ChatGPT Work, Codex, and GitHub Copilot, with a 1.05M token context window, 128K output limit, and a reasoning effort ladder from none to max.
• Sam Altman flagged this DevDay as the first time in OpenAI history he said 'this is too much stuff to launch,' signaling the pipeline behind Sol and Luna is deeper than two models.

Anthropic released Claude Opus 5.5 on September 22, its first model since publicly calling for a pacing of frontier AI development.

• Priced at $4 input / $20 output per million tokens, Opus 5.5 matches Fable 5.1 on most tasks while running 40% cheaper than Opus 5, and scored highest of any Anthropic model on its automated behavioral audit.
• GitHub Copilot integration went live immediately, with early internal testing showing Opus 5.5 resolved agentic tasks in fewer steps and recovered faster from errors in multistep workflows than its predecessor.

Palo Alto Networks launched Unit 42 Continuous Frontier AI Defense on September 22, upgrading its point-in-time exposure service to a continuous, agent-driven model.

• The service deploys Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6-Cyber alongside open-weight models to continuously test web applications, APIs, and internal systems for exploitable paths.
• Zscaler's CEO disclosed publicly that his red team found critical flaws in nearly every enterprise AI system tested, with a median breach time of 16 minutes, framing continuous defense as a necessity rather than a premium.

Slack shipped Code Channels, a feature that puts AI coding agents directly inside team channels, announced at Dreamforce 2026.

• Coding agents join channels alongside developers, receiving tasks, posting diffs, and participating in code review as conversational turns rather than requiring a separate IDE or terminal context switch.
• Slack GM Rob Seaman frames it as 'multiplayer AI,' arguing that siloed agent sessions lose context that lives in team conversation, and that a shared channel is a better development environment than a solo terminal.

Clay published results from an experiment in which its Claygent AI filled out demo and contact forms at 6,346 companies to measure how long any human responded.

• Most companies never wrote back, with the experiment quantifying that speed-to-lead failure is systemic across B2B software, not an edge case, giving Clay a data-backed argument for why AI-driven follow-up agents are necessary infrastructure.
• The self-referential design of using Clay's own agent to audit the GTM responsiveness of the broader market doubles as a public benchmark that positions Claygent against manual SDR workflows.

Appfigures data reported by The Elec shows Meta Muse reached approximately 1.1 million installs by September 18, roughly 10 days after its September 8 launch, ranking first on both the U.S. App Store and Google Play.

• Downloads and daily active users exceeded ChatGPT's levels in the same post-launch window, according to multiple analytics sources, making Muse the fastest-growing AI assistant launch on record by install velocity.
• The agent handles email, travel booking, payments, and multi-agent handoffs, and Fortune reports it is already drawing backlash from Amazon over its shopping and payment capabilities, which is the clearest signal yet that Muse is competing for commerce transaction volume, not just attention.