💸 Gemini 4 Carbon opens to the public, with a short-lived introductory price


Claude gains collaborative workspaces, Grok Build exposes sensitive apps, and New York opens frontier-AI registration.
The Merpati Post
Daily AI Briefing

Issue · November 11, 2026

Gemini 4 Carbon reaches the public as AI moves deeper into business discovery and shared workspaces, while exposed apps and new oversight underline the cost of weak controls.

A dark faceted monolith stands beyond open terracotta gates in a sunlit courtyard, with a small gray pigeon perched on the foreground parapet.

AI in general

Frontier models, research and policy

3 stories

Score · 99 / 100

Gemini 4 Carbon opens to the public, with a short-lived introductory price

Source: Google — November 10, 2026

FRONTIER MODELS · AGENTS · PRICING

Google is expanding Gemini 4 Carbon beyond its initial cybersecurity-defender release to paying developers, enterprises, and Google AI Ultra subscribers. Carbon targets difficult, extended work rather than routine chat: complex debugging, document-heavy analysis, and scientific research, with a preview of coordinated agent teams in Antigravity. The immediate decision is whether its additional reasoning improves your hardest tasks enough to justify the token usage—and a scheduled price increase.

What matters

  • API pricing starts at $2 per million input tokens and $10 per million output tokens; after November 30, it rises to $4/$20.
  • Antigravity’s preview /swarm feature coordinates agents in isolated workspaces, pairing deeper reasoning with lighter models and adversarial verification.
  • Google still recommends Gemini 4 Argon for most engineering and knowledge work. Carbon deliberately spends more tokens exploring and checking difficult problems.
  • Benchmark leadership and cost-per-task advantages are claims presented in Google’s announcement, not guarantees for your workload.

VerdictREAD FULL** — Worth checking availability, pricing, and examples before adding it to your model-routing mix.

Score · 97 / 100

ChatGPT gives businesses verified profiles, discovery analytics, and branded agents

Source: OpenAI — November 10, 2026

COMMERCE · PRODUCT DISCOVERY · BUSINESS AGENTS

OpenAI’s new Businesses in ChatGPT beta lets eligible companies claim their organization, maintain business details and product feeds, and see aggregated discovery metrics. Eligible US ecommerce and online-service businesses can also launch Business Agents that answer questions using their business information. This makes ChatGPT a more explicit customer-acquisition channel, while introducing a boundary users need to recognize between independent answers and a brand’s own representative.

What matters

  • Business Hub includes mentions, website clicks, and discovery topics; OpenAI says businesses do not receive users’ conversations or agent chats.
  • Claiming a business is available more broadly than launching an agent, which initially has US and category restrictions.
  • Business Agent conversations are labeled and separate from the original ChatGPT conversation. Sponsored agent placements are also meant to be clearly distinguished.

VerdictSKIM** — Useful for anyone responsible for product discovery, ecommerce, or keeping company information accurate across AI channels.

Score · 95 / 100

New York opens registration ahead of frontier-AI safety requirements

Source: TechCrunch — November 10, 2026

POLICY · AI SAFETY · COMPLIANCE

New York has opened the registration portal for large frontier-AI developers ahead of the RAISE Act’s January 1, 2027 start. The important distinction is timing: registration preparations began Tuesday, but the principal requirements take effect at the start of next year. The law turns safety frameworks and incident reporting into concrete compliance obligations for covered developers.

What matters

  • The report identifies covered large frontier developers as companies with more than $500 million in annual revenue.
  • Requirements include public safety frameworks, model transparency reports, and reporting critical safety incidents within 72 hours.
  • Penalties can reach $1 million for a violation and $3 million for subsequent violations; the new DIGIT office will oversee implementation.

VerdictSKIM** — Read for the enforcement timeline and scope, especially if your organization procures frontier models or tracks vendor governance.

Software engineering

Coding agents, developer tools and infrastructure

3 stories

Score · 99 / 100

Grok Build exposes sensitive apps through discoverable sharing addresses

Source: Ars Technica — November 10, 2026

SECURITY · AI APP BUILDERS · ACCESS CONTROL

Security firm Claroc found more than 35,000 publicly accessible Grok Build apps, including apps containing confidential business material and personal information. The central failure is architectural: supposedly obscure sharing addresses put their random component in a subdomain, where passive DNS services can reveal it. A hard-to-guess link is not a dependable privacy boundary when the infrastructure makes that link discoverable.

What matters

  • Exposed material included internal business plans, employee information, and medical exam results; 24 apps belonged to SpaceXSI itself.
  • Ars reports that the 24 company apps became unavailable after its inquiry, while almost all other exposed data remained accessible at publication.
  • Claroc’s tests and a company representative disagree about default privacy settings. Check the actual setting rather than relying on an assumption.
  • For sensitive prototypes, require authenticated access, review existing share links, and treat accidentally published secrets as potentially compromised.

VerdictREAD FULL** — A concrete failure mode worth checking across every AI app builder your team uses.

Score · 97 / 100

Claude Fable 5.2 and Mythos 5.2 target harder technical work at lower prices

Source: Anthropic — November 9, 2026

CODING AGENTS · MODEL RELEASES · REASONING

Anthropic’s latest Fable and Mythos share the same underlying model, with different safeguards and access arrangements. The company reports stronger coding, reasoning, and vision, alongside faster output and lower pricing than their predecessors. Its practical recommendation is more restrained than the demos: start most everyday coding work with Opus 5.5, and escalate to Fable for ambitious, open-ended problems.

What matters

  • Standard pricing is $8 per million input tokens and $40 per million output tokens; optional fast mode costs $16/$80.
  • Anthropic reports up to 30% faster output than the previous versions, and up to 2.5-times token-generation speed with fast mode.
  • Fable is broadly available; Mythos access for cybersecurity and biomedical applications goes through verification programs.
  • Anthropic reports strong results from migrations and large agent teams, but its benchmark scores depend on effort settings and evaluation conditions.

VerdictREAD FULL** — Focus on effort-versus-cost comparisons and model-selection guidance, not just headline benchmark scores.

Score · 96 / 100

NVIDIA Snapshot reuses initialized GPU workloads instead of repeating cold starts

Source: NVIDIA — November 10, 2026

INFERENCE · INFRASTRUCTURE · KUBERNETES

NVIDIA’s open-source Snapshot captures an initialized workload’s CPU and GPU memory, then restores it into new Kubernetes pods. This skips repeated weight loading, compilation, and CUDA graph capture when scaling inference or recovering a failed replica. The useful distinction is that it accelerates workload initialization—not the entire process of provisioning a machine and pulling its container image.

What matters

  • Restore targets need the same GPU model and driver; portability is conditional, not arbitrary.
  • Workloads need a small pause/resume handshake, with examples available for vLLM, SGLang, and TensorRT LLM.
  • Partner results include Crusoe reducing GLM-5.3 pod readiness from roughly eight minutes to 31 seconds on HGX B300.
  • Multi-GPU support covers compatible workloads within a single node; validate storage and runtime compatibility before extrapolating results.

VerdictREAD FULL** — Valuable for inference-platform teams dealing with bursty demand, slow scaling, or expensive warm capacity.

Design & creative

Creative workflows and user experience

2 stories

Only 2 strong recent items found.

Score · 99 / 100

Claude adds shared whiteboards, spreadsheets, and governed data workspaces

Source: Anthropic — November 10, 2026

COLLABORATION · UX · PRODUCT WORKFLOWS

Claude Whiteboards and Sheets are launching in beta, while Claude Dashboards becomes generally available as Claude Data. The meaningful shift is from generating an artifact you copy elsewhere to keeping a shared, editable workspace alongside the assistant. For product and design teams, that could shorten the path from a messy discussion to a structured plan, working budget, or reusable data story.

What matters

  • Whiteboards supports real-time collaboration, thematic grouping, timelines, and conversion into documents or decks, with continuation in tools such as Miro and Lucid.
  • Sheets supports tabs, formulas, formatting, and dropdowns, but Anthropic explicitly describes current spreadsheet functionality as simple; export to Excel or CSV remains available.
  • Claude Data adds capabilities including certified metric queries, live sharing, design-system support, caching, and tracked changes; certified queries and live sharing are in beta.
  • All are on by default for Team plans. Enterprise Whiteboards and Sheets default off; Claude Data is scheduled to turn on by default November 19 unless admins change settings.

VerdictREAD FULL** — The capability limits, sharing model, and admin defaults matter more than the office-suite framing.

Score · 97 / 100

Adobe brings generative removal deeper into photo and video editing

Source: The Verge — November 10, 2026

CREATIVE TOOLS · VIDEO EDITING · GENERATIVE AI

Adobe’s MAX announcements focus heavily on reducing cleanup work inside existing creative applications. Premiere gains a beta Generative Remove tool for unwanted objects directly on the timeline, while Photoshop improves removal and adds lighting and comparison tools. Firefly is also expanding into repeatable creative workflows and app prototyping, though those newer surfaces have mixed availability.

What matters

  • Premiere lets editors mask and track an object before generating replacement detail, and adds tools for separating dialogue, music, ambience, and effects.
  • Photoshop’s updated Remove Tool handles associated elements such as reflections and shadows; Quick Compare and Relight support review and lighting adjustments.
  • Firefly Quick Actions is generally available, while Firefly Design and Graph are beta experiences and Firefly Builder is in private beta.

VerdictSKIM** — Check which features are actually available in your applications, then test cleanup quality on representative footage or images.

Open-source watch

Projects gaining meaningful traction

3 stories

These are trending projects, not necessarily newly launched releases. Traction figures below are from today’s source inventory.

Project · 01

Agent Skills — addyosmani/agent-skills

AI · DEVTOOL · AGENT

A collection of engineering skills intended for AI coding agents, described by its maintainer as production-grade. It is interesting as a way to package repeatable engineering guidance rather than restating your expectations in every prompt.

What matters

  • GitHub Trending recorded 465 stars today.
  • Best suited to teams experimenting with reusable instructions and engineering workflows.
  • Review the actual instructions before adoption; popularity and a “production-grade” description do not establish correctness or compatibility with your agent.

VerdictTRY** — Start with one relevant skill in a low-risk repository and inspect whether it improves the agent’s changes.

Project · 02

OpenSandbox — opensandbox-group/OpenSandbox

AI · AGENT · INFRA

OpenSandbox provides a sandbox runtime for AI agents, positioning itself around security, speed, and extensibility. It is relevant when an agent needs to execute code or interact with an environment without inheriting unrestricted access to the host.

What matters

  • GitHub Trending recorded 308 stars today.
  • A useful candidate for teams building agent execution infrastructure rather than only consuming a hosted assistant.
  • Check isolation boundaries, network policy, credential handling, and cleanup behavior yourself; the project description is not a security audit.

VerdictWATCH** — Worth evaluating against your threat model before entrusting it with sensitive workloads.

Project · 03

Plannotator — backnotprop/plannotator

AGENT · DEVTOOL · PRODUCTIVITY

Plannotator is a workspace for reviewing agent plans, annotating work, reviewing code, and sending feedback back to agents. Its appeal is the human-control layer: making it easier to steer work before a long agent run compounds an incorrect assumption.

What matters

  • GitHub Trending recorded 48 stars today—smaller traction than the other picks, but a directly relevant workflow.
  • The project description emphasizes planning, decisions, annotations, code review, and team sharing.
  • Check compatibility with your agents and how shared plans or code are stored before using it on private projects.

VerdictTRY** — A focused experiment for anyone who finds chat alone awkward for reviewing plans and giving precise feedback.

Editor’s note

Fresh, non-repeated developments with concrete consequences for model choice, customer discovery, access control, infrastructure costs, and creative collaboration—not announcements selected merely to fill slots.

30-second feedback

How useful was today’s digest?

★★★ Very useful ★★ Somewhat useful ★ Not useful

Your choice opens the short form with your rating filled in. Or share a quick note.


Prefer reading on Telegram? Subscribe to The Merpati Post →

The Merpati Post · Daily AI Briefing

background

Subscribe to The Merpati Post