|
|
|
Plus: a court makes saying no to the Pentagon costly, Copilot's agents land on a meter, and LangChain agents remember you outside Slack.
|
|
Cursor shipped Rollouts and a faster Security Review on September 23: one watches your deployments, the other scans PRs in 3.8 minutes flat, and engineers now act on 60 to 70 percent of its findings, per Cursor's blog. Trial credits run through about October 3.
|
|
|
|
|
Google put a price on a talking, lip-synced support agent.
Gemini 3.8 Live Avatar went GA in Gemini Enterprise on US and EU endpoints on September 24, per Google's announcement. It generates speech and a lip-synced face in one pass, switches among 97 languages unprompted, and watermarks output with SynthID. The rate card runs $0.75 per million text-input tokens, $3.00 per million audio-input, $12.00 per million audio-output, and $1.00 per million video-output, billed only while the avatar speaks. Per DataNorth's calculation, that comes to about $0.39 per speaking minute and roughly $4 for a ten-minute session. You can cost out a support or onboarding pilot today. You cannot plan for responsiveness yet: Google has published no latency figure, so test against your own call flow before promising anyone a natural conversation. See the full rate card.
|
Microsoft moved Copilot's agent work onto a meter.
Microsoft rebuilt Copilot on September 25 around Home, Code, and Autopilot, a tenant-resident agent with its own identity and memory that keeps working after you log off, per Microsoft's announcement. The flat seat still covers Chat and the Office apps. Cowork, Code, Autopilot, and frontier-model calls (GPT-6 Astra, Claude Fable 5.1) now bill by usage at provider list price. Microsoft counted 30 million paid Copilot seats as of July against more than 450 million commercial Microsoft 365 seats, roughly 7 percent of that base, per GeekWire's reporting. For IT and finance, agent work is now a variable line item. Set a usage budget and alerts before anyone provisions an Autopilot agent; it enters private preview at the end of September. Read the billing split.
|
LangChain agents can now remember each user without leaking it in group chats.
Managed Deep Agents 0.8 shipped September 24, per LangChain's blog. User-scoped memory at /memories/user/ stays private by default in Slack group channels and HTTP calls. A JSON-webhook channel lets agents run outside Slack, Slack file transfer landed, and a Parallel web-search tool comes built in with no separate API key while MDA is in beta. The webhook channel matters most: a managed agent can now sit behind a support portal or a customer web app. Plan for two gaps. MDA requires a LangSmith Plus or Startup plan, per Superpower Daily, and Okta or Entra ID auth at the webhook boundary is yours to build. See what 0.8 changes.
|
An appeals court let the Pentagon penalize Anthropic for keeping its safety limits.
The D.C. Circuit ruled 2-1 on September 25 that the Pentagon had grounds to keep its "supply chain risk" label on Anthropic, which refused to drop contract restrictions on lethal autonomous weapons and mass surveillance, per Defense One and ABC News. The decision conflicts with an August ruling from a California district judge on a related designation, and Anthropic says it is weighing further review. If you sell models or model-backed software into federal or defense-adjacent work, a federal appeals court has now accepted the argument that a vendor's own usage restrictions can count as a security risk. Have counsel read your acceptable-use terms against this opinion before your next renewal. Read the ruling breakdown.
|
|
•
|
Alibaba showed a chip-to-agent stack built outside US clouds. At Apsara on September 23 it announced AgentCore, a Zhenwu V900 chip claimed at 3x its predecessor's performance (mass production Q1 2027), and a Qwen roadmap targeting 5 to 10 trillion parameters, per the press release carried by HPCwire; worth tracking only if your team cannot route inference through US-hosted APIs. Details →
|
|
|
|
|
Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.
|
|
|
Affiliate disclosure
·
Unsubscribe
·
Manage preferences
Pondero earns commissions on some links. This does not affect our editorial picks.
|
|