The one-screen version: today's three stories all sit below the demos — how you run agents, what inference costs, and whether the valuations hold. (1) GitHub's Oct 28 Universe changelog gave Copilot's cloud agents session continuity — start on your phone, steer from VS Code, check from the web, one session running in the cloud throughout — but shipped none of the retention or billing terms. (2) DeepSeek's V4.1-Flash line keeps dropping cheap, MIT-licensed, ~1M-context updates, keeping the inference price war structural. (3) Reuters, via secondary outlets, reports Anthropic's November share-sale valuation being talked down from ~$2T to ~$1.5T. Here's what each one asks of you.

1. Your coding agent now follows you across devices — mind the meter#

GitHub's October 28 Universe changelog for Copilot in VS Code shipped the feature that makes cloud agents feel ambient instead of tab-bound: session continuity. Kick off a cloud agent session from your phone during standup, open VS Code at your desk and keep steering the same run, then glance at it from github.com over lunch — the agent works on GitHub-hosted compute the whole time, and you move between clients without losing the thread. The update also makes the Copilot harness, built on the Microsoft Copilot SDK, the default for new sessions in both the editor and the Agents window.

The capability is real and genuinely useful — the hand-off is the kind of friction removal that changes how a small team actually works. What's not in the post is the part a founder has to care about: it discloses no pricing, no quota, no data-retention policy, and no cost-accounting detail. An always-on cloud agent is a metered thing by nature, and "it just keeps running" is a billing sentence as much as a productivity one.

What it changes for you: adopt the workflow, but instrument it before you lean on it. Put the cloud-agent usage behind a dashboard you check, set an org policy for who can spawn unattended runs, and treat "the agent kept working while I was away" as a line item, not a free lunch — the same discipline we flagged for agentic computer-use costs.

2. DeepSeek keeps the price war lit#

While the labs chase trillion-dollar valuations, the open-weight camp keeps doing the unglamorous thing: shipping cheaper. DeepSeek's V4.1-Flash line added an October 17 "1017" checkpoint — open weights under the MIT license on Hugging Face and ModelScope, an ~1M-token context window, and API serving at prices that secondary trackers put around $0.15 per million input and $0.60 per million output off-peak, roughly doubling at peak. The aggregators disagree on the exact parameter count and even on which date counts as "the" release, so this is a benchmark-it-yourself story, not a spec sheet to quote — verify against DeepSeek's own API docs before you wire it in.

The point isn't this one checkpoint. It's that the downward pressure we traced through yesterday's chip story — OpenAI lining up three silicon vendors, Nvidia's pricing monopoly cracking — has a twin on the model side: a steady drip of capable, long-context, openly licensed models you can run or rent for cents. Cheap inference is becoming structural, not promotional.

What it changes for you: keep your model layer swappable. The teams that win the margin game in 2027 are the ones who can re-route to the cheapest capable model the week it ships, not the ones who hard-wired one provider's SDK into their product. If you've been meaning to put an abstraction between your app and your model, the price curve just made it worth the afternoon — see how to route every request to the cheapest capable model.

3. The first crack in the IPO story#

Four days after Anthropic filed its S-1 on October 26 to list on Nasdaq as ANT — disclosing $16.3 billion in first-half revenue — the number everyone was watching moved the wrong way. Reuters, as summarized by secondary outlets, reports that investors at Anthropic's investor day were discussing a valuation of roughly $1.5 trillion for the November share sale — down from the ~$2 trillion previously floated — amid talk of a third-quarter growth slowdown.

Read the caveats before you read the headline. The $1.5T figure is investor-room discussion, not a priced deal; "weak Q3" is reported framing, not released financials; and a company that raised a $65B round at a $965B valuation in May stepping toward a trillion-and-a-half is still a staggering number by any pre-2025 standard. Nobody is calling this a down round.

But the direction is the signal. For eighteen months the AI-boom narrative ran one way: usage up, so valuation up, so raise more. The first time one of the flagship names gets marked toward reality — even informally, even by a few hundred billion — is the moment the questions in your own fundraise change.

What it changes for you: don't over-read one number, but do update your deck. The reflexive "we're in a hot category" pitch is aging out; the durable-revenue pitch — retention, gross margin, net dollar retention, the unit economics under the usage — is aging in. If the most valuable lab on paper is being asked whether its growth justifies its price, your Series A investor is going to ask you the same thing, and "the category is huge" will not be the answer that lands.

The one line to take into the week#

All three stories rhyme: the interesting action is under the floor. Build on agents you can run anywhere but whose cost you can see, on models cheap enough and portable enough to swap, and on revenue durable enough to survive the market pricing growth honestly again. Last edition: OpenAI moved to buy Google's chips, and the model learned to talk.