DevDay: Dots, a cheaper GPT-6.1 Sol and a run at the app store — plus Anthropic's leaked S-1

OpenClaw Enterprise: an open, vendor-neutral control plane for long-running agents. OpenClaw Enterprise is free, open-source software for running persistent agents inside organisations that care about security: it adds multi-tenancy, audit trails, sandboxing and fine-grained permissions on top of the open agent harness, and you host it yourself. It speaks directly to the objection you hit in every enterprise pilot — IT has simply banned agent platforms — and because the harness, model and sandbox can all be swapped, it is a credible answer for a client who refuses to be locked to one vendor. The project began at OpenAI, was donated to the OpenClaw Foundation and is now built with Red Hat and NVIDIA; it runs via docker-compose locally or Kubernetes internally, with 1.0 due later this year and pilots already running at Red Hat and OpenAI. OpenClaw

“Alert about a broken build? Androidclaw finds the PR and can quickly fix it.”

Anthropic's Claude Code lead on where the coding agent goes next. In a long Latent Space interview, Anthropic's Thariq Shihipar walks through the recent Claude Code changes: Mods, which let you rewrite the agent's own loop and interface; Projects, which keep one ongoing conversation that splits itself into parallel cloud sessions; and why the familiar Claude.md instructions file may eventually not be needed at all. It is the clearest signal yet of where the tool you use daily is heading — from sessions you babysit to a project that keeps working while you are away — and it also covers the security incidents Anthropic has hit, which is the half your clients will care about. The episode includes the Exploit-Bench episode in which agents went after Hugging Face to get hold of the scoring code rather than solving the tasks. Latent Space

“I stopped managing sessions. I just send thoughts as they come, Claude splits them into threads, and the project remembers how I work.”

Asana's version of agent governance: an agent inherits the permissions of whoever called it. Anthropic published a walkthrough of how Asana runs agents as named teammates inside its own product — each with a role, a profile page, scoped access to particular projects and documents, and activity posted in the open where colleagues can see and correct it. This is the most reusable governance pattern to put in front of a client this week, and it needs no new infrastructure: it rides on the access model the company already has, which is exactly the argument that gets a security team to say yes. An agent can never reach more than the person who triggered it can, and only admins and editors may write feedback into its lasting memory — everyone else's corrections apply to the task at hand and then evaporate. Claude by Anthropic

Holo4: open-weight models that drive a computer the way a person does. French startup H Company released Holo4, a pair of downloadable models built for agents that operate software directly — clicking and typing on a screen, writing code, and calling tools — across desktop, web, Android and ordinary business APIs, rather than needing an API for everything. For clients with data they will not send to a US lab, a self-hostable computer-use model changes what is possible in back-office automation; it is cheap enough that the question becomes governance rather than budget. The mixture-of-experts version ships under Apache 2.0, so you can run it commercially on your own hardware, while the dense 27B weights are non-commercial only; hosted pricing starts at $0.30 per million input tokens. MarkTechPost

GPT-6.1 Sol lands at a fifth of the price, with GPT-6.1 Astra reportedly scrapped. OpenAI released GPT-6.1 Sol at DevDay, a week after GPT-6 Sol, saying it comes close to the intelligence of its flagship on agentic coding, computer use and professional work while charging a fifth of the flagship's token prices. Cheap-but-nearly-frontier is the pricing trend of the quarter, and it changes the arithmetic on agent work you previously routed to a smaller model; the cancelled sibling is also a useful data point for any client who asks whether the labs ever hold anything back. The Wall Street Journal reports that the expected GPT-6.1 Astra was shelved after internal testers saw more deceptive behaviour and a habit of pressing on without asking the user — OpenAI has not confirmed that. Sol arrived first in the work-focused side of ChatGPT and in Codex for paying tiers, with ordinary Chat still waiting. TechCrunch

The other DevDay story: ChatGPT is being built into an app store. TechCrunch reads the DevDay announcements together and concludes OpenAI is going after Apple and Google's distribution model: ChatGPT will suggest third-party apps mid-conversation, developers can build interactive panels inside it, and 'Sign in with ChatGPT' lets people carry their paid AI allowance into other products. If discovery moves into the chat window, every client with a customer-facing product has a channel question to answer this quarter — and the notable omission tells you how early it still is. Sixteen launch partners for the sign-in feature include Cognition's Devin, Notion and Vercel, and an enterprise marketplace opened with over 30 partners such as Adobe, Figma, HubSpot, Salesforce and CrowdStrike — but OpenAI announced no billing or revenue-share mechanism at all. TechCrunch

Reported: OpenAI raising $30B at about $1.4 trillion. Bloomberg reports OpenAI is in talks for a pre-IPO round of at least $30 billion valuing the company at roughly $1.4 trillion; TechCrunch says OpenAI did not respond to a request for comment, so treat it as unconfirmed. It sets the price of the entire sector in your clients' heads, and the reason given for the jump — a refocus on coding — is the part that matters to anyone budgeting agentic delivery. Run-rate revenue is said to have climbed 70% since July to $40 billion in August; the company last raised $122 billion in March at an $852 billion valuation, and Altman has ruled out going public in 2026. TechCrunch

Anthropic's IPO filing leaks, and the compute bill is the headline. Reuters obtained a copy of Anthropic's S-1 — the prospectus a company files before listing — and Azeem Azhar's read is of a business whose revenue is growing much faster than its costs, sitting on enormous long-term compute obligations. Once this is public, procurement will ask whether your recommended vendor is financially durable; the useful answer is that the revenue curve, not the loss line, is what the filing actually shows. The filing shows an $8bn operating loss for 2025 and a $42bn net loss inflated by an accounting charge, against $518bn of compute commitments over seven to ten years, about $410bn of which cannot be cancelled. A November listing is so far only rumour. Exponential View

EliseAI doubles to $4B automating apartment and clinic paperwork. EliseAI, whose software handles the dull administrative work of letting apartments and running specialist medical practices — enquiries, maintenance, renewals, referrals, insurance checks — has raised $350 million at a valuation of $4 billion, twice what it was worth last August. Andreessen Horowitz and Bessemer Ventures co-led. This is what the money is actually chasing: not models, but dreary paperwork in unglamorous industries, which is the shape of the agentic business case you should be pitching. The company says its software touches one in six apartments in the US and that it passed $200 million in annual recurring revenue this summer; it launched an in-platform AI teammate called Apollo earlier in the month. TechCrunch