NPDev Daily AI Briefing - November 11, 2026
Google put Gemini 4 Carbon in every paying developer's hands and previewed /swarm, a coordinator-plus-verifiers agent harness, while Anthropic turned Claude into a workspace with Sheets, Whiteboards and in-chat partner apps. The open-weight side moved too: Meituan published LongCat-2.5 at 1.6T parameters and Moonshot's Kimi K3.1 Pro reached Amazon Bedrock.
Written by NPDev Labs with AI assistance and reviewed before publication. How we work.
Google ships Gemini 4 Carbon to all paying developers and previews /swarm in Antigravity
Gemini 4 Carbon moved from a Fairwind cyber-defender preview to general availability on November 10: Gemini API in AI Studio and Gemini for business for paying developers and enterprises, plus Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity.
Google's post is explicit that Carbon is for tasks that test a model's limits - multi-step coding and debugging, open-ended research - and that Argon remains the sensible default for most software engineering and knowledge work. Carbon "uses more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops." Google claims first place on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next best external model; those are vendor-reported. Pricing is Argon's introductory rate, $2 per million input tokens and $10 per million output tokens, until the introductory period ends on November 30, 2026, after which $4 and $20 apply. The model card was published the same day. Carbon first went out on October 28 to cyber defenders in the Fairwind Program; this is the broad release, and Google has not said when free or lower-tier users get it.
Why it matters: The frontier model and the orchestration harness shipped the same day, and the introductory price doubles on November 30 - the second number is the one to put in a budget.
Alongside the model, Google previewed /swarm in Antigravity: a coordinator that spins up specialized workers in isolated git worktrees so edits never collide, adversarial verifiers that run tests and challenge diffs, a shared append-only message board where agents park at zero token cost when idle, and per-role model pinning - Carbon for architecture, coordination and auditing, lighter models for parallel file edits.
Sources:
Anthropic turns Claude into a workspace: Sheets, Whiteboards, Data and apps inside the chat
Claude Sheets and Claude Whiteboards entered beta on all paid plans on November 10, Claude Dashboards became generally available as Claude Data, and Interactive Connectors let partner apps render inside Claude and be operated side by side with the model.
Sheets generates workbooks with tabs, formulas, formatting and dropdowns from a prompt and annotates each cell edit so changes stay traceable; Anthropic says functionality is deliberately simple for now, with .xlsx or CSV export and Claude for Excel for the rest. Whiteboards turns a spoken plan into a board with owners and dates, groups retro notes by theme, and exports to a doc, deck, Lucid, Miro or Vani. Claude Data adds data stories, certified queries in beta and live sharing over Amazon Redshift, BigQuery and ClickHouse. The same day, Anthropic shipped Interactive Connectors, which let partners embed their apps directly in Claude so people and the model work in the same tool on the same screen.
Why it matters: The integration surface is shifting from "your app calls the model" to "the model hosts your app" - Interactive Connectors is the MCP-shaped distribution channel, so shipping a tool now includes a placement decision inside Claude.
This is a seat-grab aimed at Excel and Miro, announced eight days before Anthropic's expected Nasdaq debut and one day after Fable 5.2 and Mythos 5.2 shipped. For builders the interesting piece is not the spreadsheet: it is that Claude is becoming an app runtime with its own distribution rules.
Sources:
Meituan publishes LongCat-2.5, a 1.6T-parameter multimodal agent model
LongCat-2.5 landed on Hugging Face as a native multimodal mixture-of-experts with 1.6 trillion total parameters and roughly 48 billion activated per token, trained for long-horizon agentic work: complex codebase development, full-project builds, design and 3D tasks.
The model card reports Terminal Bench v2.1 at 88.4 (from 70.8 in LongCat-2.0), Terminal Bench v4.0 at 38.5, DeepSWE v1.1 at 64.6, PaperBench at 82.1 and SWE Atlas codebase Q&A at 58.3. Scores marked with an asterisk are cited from other labs' reports or public leaderboards; unmarked scores are measured in-house and described as indicative only. The comparison columns put LongCat-2.5 against Claude Opus 5, GPT-5.6-Sol, GLM-5.3, Kimi K3 and DeepSeek-V4.1-Flash rather than against its own predecessor alone. Training notes describe multi-teacher distillation across agent, reasoning and interaction teachers, agentic RL built on real engineering tasks and reproducible agent failures, visual features converted to discrete tokens in a shared semantic space, and 86% higher per-device throughput than a DDP baseline on long-sequence SFT, with training and inference aligned to bitwise-identical token log probabilities.
Why it matters: A ~48B-active open model posting frontier-adjacent terminal and codebase numbers changes the build-versus-buy math for coding agents, and it is the third credible open-weight release in ten days.
Read the two terminal scores together: 88.4 on v2.1 and 38.5 on v4.0 is the same pattern the closed labs show, and it is where harness difficulty stops flattering a model. Set the card beside Kimi K3.1 Pro and GLM-5.3 and the open tier is a real option for agent traffic, with the usual caveats about self-hosting a trillion-parameter mixture-of-experts.
Sources:
Kimi K3.1 Pro reaches Amazon Bedrock with a 2-3x faster Highspeed variant
Moonshot's Kimi K3.1 Pro arrived on Amazon Bedrock on November 10 as Kimi K3.1 Pro Highspeed, a post-trained K3.1 with the same 1M-token context and open-weight 3T-class architecture, which Moonshot says generates 2-3x faster than K3.1 Pro.
AWS says the model uses the same endpoint, APIs and features as Kimi K3, supports tool calling, structured output, configurable reasoning, image input and streaming, and is reachable through global, US and India cross-Region inference profiles with zero data retention and no training on customer data. Moonshot describes extra post-training compute aimed at long-horizon agentic work and fewer factual errors, plus gains in financial modeling, office documents, and coding for game engines, 3D scenes and CAD scripts. A parallel listing through Fireworks on Microsoft Foundry gives 2.8 trillion total parameters with 104 billion active, a 1M-token window and $2.05 input / $10.35 output per million tokens, with vendor-reported Terminal-Bench 4.0 at 65.7 and BenchCAD with tools at 92.7.
Why it matters: Open-weight frontier models increasingly ship through clouds first, where residency and retention terms decide adoption more than benchmark deltas. If you already call Kimi K3 on Bedrock, the fast tier is a model-ID change.
The speed improvement is a post-training change on the same weights family, not a new generation - the kind of increment that quietly resets latency budgets for agent loops without a migration project.
Sources:
OpenAI opens ChatGPT to businesses: claimed profiles, product feeds and Business Agents
Businesses in ChatGPT entered beta on November 10: companies claim a verified profile in Business Hub, publish self-service product feeds with live pricing and availability, see aggregated insights, and - if eligible - publish a Business Agent customers can talk to from their ChatGPT profile.
OpenAI cites 1.2 billion weekly ChatGPT users. Claimed profiles carry a blue check and can be updated once verified; Business Hub shows aggregated mentions, website clicks and common topics in conversations where a business is cited, and OpenAI states it does not share people's conversations or Business Agent chats with businesses. Business Agents launch for eligible US ecommerce and online service businesses in a clearly labelled conversation kept separate from ChatGPT's own answers, and businesses can sponsor their agent as a ChatGPT Ads destination. Eligibility initially requires a website and a supported category; OpenAI says more markets and categories follow.
Why it matters: ChatGPT is becoming a discovery surface with an ad model attached, and the agent is the storefront. "Does our product appear correctly when ChatGPT answers a buying question" is now a growth channel with a feed, an owner and analytics.
Two surfaces now matter for distribution: what the model says about you unprompted, and what your agent says when someone clicks through. The second one you control, which is precisely why the first one stays the harder problem.
Sources:
OpenAI sets a December 11 end date for Advanced and Standard voice in ChatGPT
OpenAI began notifying users on November 10 that Advanced and Standard voice leave ChatGPT Voice on December 11, 2026, with everyone defaulting to GPT-Live, the voice model introduced in July.
OpenAI says voice usage has nearly doubled since GPT-Live launched and that virtually all voice users are already on it, which is the stated reason for the retirement. On December 11 the older options disappear from Settings > Voice and anyone who selected them is moved to Live. Live video and screen-sharing continue on Advanced voice until GPT-Live supports them, and voice with custom GPTs also runs on Advanced voice until the same date. Some users have argued publicly that GPT-Live interrupts more and returns shallower answers than Standard voice; that criticism is reported, and OpenAI's post only acknowledges that some people prefer the older response style while promising work on depth, continuity and personalization.
Why it matters: Voice is being consolidated onto one model with an explicit capability gap left open - live video and screen-sharing still run on Advanced voice - so anything built on that path has no GPT-Live replacement yet.
Deprecations are where integrations break quietly. Any product that shells into ChatGPT voice, or teaches users to use live video and screen-share for support, needs a written answer for December 12 - and the same date retires voice for custom GPTs, which compounds the custom-GPT-to-plugins migration already underway.
Sources:
IBM and Microsoft bundle 6,000 agents into Imagine Enterprise
IBM and Microsoft announced Imagine Enterprise on November 10, pairing IBM Consulting's 6,000-plus pre-built agents and assets with Microsoft Foundry, Copilot Studio, Fabric, Entra, Purview and Defender on Azure, sold with forward-deployed engineers from both companies.
Initial industry frameworks cover financial services, insurance and retail, with manufacturing, energy, public sector and mobility named as future additions. IBM Consulting Advantage's Process Studio, Context Studio and Agentic Core handle design, deployment and governance; Microsoft supplies the model, data, identity and security layer, including Foundry, GitHub Copilot, Copilot Studio, Fabric, Entra, Purview, Defender and Sentinel. Customers engage IBM Consulting to deploy and operate the platform, co-innovate on enterprise-scale implementations, or self-manage. SEI is named as an early example of the transformation pattern.
Why it matters: This is an incumbent's answer to build-versus-buy, with governance and identity as the product. The differentiator being sold is control around the agents, which is the part most in-house platforms underinvest in.
"6,000 agents" is a marketing number; the real claim is that agents need the same onboarding, permissions, observability and offboarding as human workers. If you are building an internal agent platform, that checklist is a fair summary of what auditors will ask for - and it is cheaper to copy than to license.
Sources:
Model releases & progress
State of the labs, as of November 11, 2026, 15:00 UTC.
Google: Gemini 4 Carbon general availability on November 10 is a new availability surface for a model previewed October 28; the model card is published and /swarm is in preview. No checkpoint beyond Carbon.
Anthropic: No new checkpoint in this window - Fable 5.2 and Mythos 5.2 from November 9 remain current. November 10 was product surfaces (Sheets, Whiteboards, Data, Interactive Connectors) plus the launch of its IPO roadshow.
OpenAI: No new checkpoint announced. GPT-6.2 chatter continued on user reports of Codex and ChatGPT traffic being routed to an unreleased model; as of November 10-11 it is absent from OpenAI's model catalog and changelog, so treat it as unconfirmed. Product news dominated: Businesses in ChatGPT and the voice retirement.
Meta: Muse Ember (November 9) remains the current release - hosted only, weights private, $0.10/$0.50 per million input and output tokens.
Moonshot: Kimi K3.1 Pro Highspeed on Amazon Bedrock (November 10) is a serving and speed update on the K3.1 family, not a new base model.
Meituan: LongCat-2.5 open weights (November 9-10) are the largest open multimodal release of the window at 1.6T total parameters.
Practical implications for engineers
What to do about it this week.
1. Price the task, not the token. Carbon ships at Argon's introductory $2/$10 rate but Google says it spends more tokens per task, and the rate becomes $4/$20 on November 30. Run your own eval set on cost per resolved task before making it a default, and put the pricing date in config next to the model ID.
2. Treat /swarm as a pattern, not a product. Isolated git worktrees per worker, adversarial verifiers that only run tests and challenge diffs, an append-only shared board, and per-role model pinning are all portable. Most teams can approximate this in CI with existing job runners this quarter.
3. Re-run the open-weight serving math. LongCat-2.5's ~48B active parameters and Kimi K3.1 Pro's 1M-token window with a 2-3x fast tier change what a single accelerator node can absorb. Measure time-to-first-token and decode throughput at your real context lengths before renewing a frontier API commitment.
4. Claim the two new distribution surfaces. A verified ChatGPT business profile with a live product feed, and an Interactive Connector inside Claude if you ship a tool. Both are opt-in, both are cheap now, and both need an owner before a competitor takes the slot.
5. Put the deprecation calendar in one ticket. Gemini 4 Carbon introductory pricing ends November 30, Advanced and Standard voice plus voice for custom GPTs end December 11, and custom GPTs retire the same day. Three teams' problems on one date.
Bottom line
The frontier labs spent this window shipping surfaces, not checkpoints: Google's Carbon plus /swarm, Anthropic's sheets and in-chat apps, OpenAI's business profiles and agents, IBM and Microsoft's governed agent fleet. The model progress that matters is on the open side, where LongCat-2.5 at 1.6T and Kimi K3.1 Pro on Bedrock put a ~48B-active open model within reach of closed frontier coding agents. Every release here changes a price, a deprecation date or a distribution surface - and all three are cheaper to handle this week than next.
As of
November 11, 2026, 15:00 UTC. Vendor-reported benchmarks and pricing are labelled. Rumored releases are marked unconfirmed. Every story links to the first-party announcement or documentation used.