AI News Briefing — Google's Gemini 4 Carbon launches only to cyber defenders

Model releases

[2026-10-28] Google announced Gemini 4 Carbon, two weeks after Argon, and launches it to a subset of cyber defenders first, widening to trusted partners through its Fairwind Program. Developers, businesses and consumers get it “as soon as possible”, with no date. Google reports 59 on the Artificial Analysis Intelligence Index v5.0, ranked first, and 88.2% on SEC-bench Pro vulnerability finding. Pricing matches Argon: $2/$10 per million tokens introductory, then $4/$20, with no end date set. (official, source)

For now only vetted defenders can call it. Teams comparing it with Argon get the same price card and no access.
[2026-10-28] NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model with 30B parameters and about 4B active per token. It transcribes 40 languages, translates 31 to or from English, and takes natural-language prompts for things like speaker labels or the spelling of domain terms. NVIDIA reports 5.33% average WER on the English Open ASR leaderboard. The Granary v2 training manifest covers 463,062 hours. (official)

Taking the spelling of domain terms as a prompt could stand in for the custom vocabulary lists many transcription setups keep by hand.

[2026-10-28] Copilot Fusion is generally available on all Copilot plans in VS Code, the Copilot app, Copilot CLI and JetBrains. Instead of a model, you pick an effort level and Copilot routes each prompt: one model, a cheap draft escalated when a check fails, or a draft reviewed by a second model. Each model used bills at its own rate, with 30% off for a limited time. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, two points under Claude Code with Opus 5.5 Max, at about half the API cost per task. (official)

Choosing an effort level hands the model choice to GitHub’s router, so the monthly bill now depends on how often drafts escalate or get a second review.

For Software Developers: A routine rename across a few files can run at a low effort level on one model, leaving the draft-plus-review path for changes where a second model reading the diff is worth paying for twice.

[2026-10-28] Google opened a preview of its Connected Apps Platform, letting outside developers put apps into the Gemini app, Google Search, or both. Apps integrate through a cloud-hosted MCP server, or on Android through the AppFunctions APIs, and are registered, tested and submitted for review in a Google Connected Console. Access opens in phases, US developers first per TechCrunch. Google’s post does not say which user plans get these apps or how actions are approved. (official, source)

A cloud-hosted MCP server that already exists covers most of the integration work. Plans and action approval are still unanswered.

[2026-10-28] Microsoft put the Copilot harness into Foundry Agent Service as a public preview: Microsoft runs the agent loop and sandbox, and the developer supplies skills, Foundry or MCP tools with configured approvals, and a source-controlled agent definition. The harness is free until December 1, when sandbox compute starts billing at $0.0994 per vCPU-hour plus $0.0118 per GiB-hour of memory, on top of model inference. (official)

Free until December 1 leaves about a month to measure real sandbox hours before they appear on an invoice.

[2026-10-28] GitHub Agentic Workflows is generally available on Free, Team and Enterprise. Workflows are written in Markdown and compile to Actions YAML. Agents get read-only permissions by default, run in a sandboxed container behind a firewall, and a threat-detection job scans proposed changes before they apply. GA adds micro-VM isolation, self-hosted ARC runners and gh aw forecast for credit spend. The default Copilot CLI engine needs a Copilot license. (official)

Teams that wired an agent into Actions by hand with a write token now have a default that starts read-only.

For Platform / DevOps Engineers: Run gh aw forecast on a workflow before rolling it out across repositories, so the credit estimate is in hand before the first scheduled run spends anything.

[2026-10-28] Claude Enterprise gets smart reports, in-app surveys and pooled group budgets, all generally available. A smart report samples a team’s transcripts and groups the work into workstreams with spend, deliverables and friction points for each. A pooled budget gives a group one monthly limit while member limits still apply, and alerts admins at 50%, 75%, 95% and 100%. Anthropic says the reports are not for judging individual performance. (official, source)

Reports built from sampled transcripts read what people actually typed, which is worth telling a team before they are switched on.

For Engineering Managers / Tech Leads: With alerts at 50% and 75%, a group budget set at last month’s spend shows by mid-month whether usage is running ahead, well before the limit is reached.
[2026-10-28] GitHub Enterprise Cloud customers with an account team can now get Copilot for every licensed user, with no included allowance: usage meters against the enterprise budget from the first request. Completions keep working when the budget runs out, while Chat, code review and the cloud agent stop. Self-serve access comes later, and Copilot Standalone costs $10 per user per month. (official)

No included allowance absorbs a first month here. Whatever budget is set is also where Chat and the cloud agent stop.

[2026-10-28] LangChain’s “loop engineering” post stacks four loops around one agent: the tool-calling loop, a grader that sends failures back with feedback, triggers such as cron or Slack messages, and a trace-reviewing agent that proposes prompt and tool changes. Its docs agent checks that links resolve, CI passes and the diff stays in scope. The post reports no measured results. (official)

Nothing here is measured. Its docs agent’s three checks, links, CI and diff scope, are cheap to copy into any grader.

[2026-10-27] FeatureArc asks ten coding agents to grow a feature over 2 to 10 stages in one session, with no test feedback or rollback, across 73 tasks in four languages. Success fell from 62.7% at the first stage to 28.5% on the final code. Inherited defects appeared in 72.9% of failed later stages, and agents missed the earlier error about 95% of the time. (paper)

Agents rarely went back to repair what an earlier stage broke, so long sessions need a test run between stages that comes from outside the agent.
[2026-10-27] A study of agent scaffold evolution finds sparse scoring during search plus targeted evaluation of finalists beat full evaluation of every candidate in most settings, using about half the evaluation runs, across four benchmarks and two evolution frameworks. (paper)

Mistral Large 4 weights: October 31 per the Hugging Face page; license unpublished.
DeepSeek Search & Fetch billing: free harness calls end October 31.
Kimi K3.1 Pro weights: promised by November 4.
Gemini 4 Carbon general access: limited to cyber defenders and Fairwind partners; Google gives no date for developers or the end of introductory pricing.

A developer release date, or an end date for the $2/$10 rate, would decide whether Carbon belongs in an evaluation this quarter.