Your AI briefing.
Clear. Practical. Sourced.

OpenAI announced Businesses in ChatGPT on November 10. After verification in Business Hub, eligible companies can update their details and supply product feeds containing pricing and availability. Shopify merchants can connect through Business Hub or Shopify admin to streamline verification. The broader claim-your-business feature is offered to eligible businesses where ChatGPT is available, while Business Agents initially target eligible ecommerce and online service businesses in the United States. An agent conversation is labeled and separate from ChatGPT’s independent answers; sponsored placements are also labeled. Business Hub reports aggregated mentions, website clicks and discovery topics when interaction thresholds are met. OpenAI says businesses do not receive users’ original conversations or Business Agent chats. The release gives operators a more direct way to maintain source information, rather than establishing a guaranteed position in AI recommendations. Availability depends on category and market, and the announcement describes a beta that will expand over time.

Google made Gemini 4 Carbon available on November 10 through the Gemini API in Google AI Studio and Gemini for business for paying developers and enterprises. It is also rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search. Google positions Carbon for difficult coding, debugging, research and complex knowledge work, while recommending Argon for most everyday engineering and knowledge tasks. Carbon can use more tokens to explore alternatives and run extended verification loops. The announced introductory API price is $2 per million input tokens and $10 per million output tokens; introductory pricing should be checked again before budgeting a deployment. Antigravity’s accompanying /swarm preview coordinates agents in isolated workspaces, with Carbon usable for roles such as architecture and auditing. Google reports strong benchmark and prerelease results, but those claims do not establish performance on a reader’s repository or scientific problem. Cybersecurity access through Fairwind retains its program context.

Google DeepMind introduced AlphaProof Code on November 10, describing research on key security properties of Signal’s cryptographic implementation. Human experts reviewed formal specifications derived from the security goals, after which Gemini Carbon-powered agents generated Lean proofs. Lean checked the proofs against those specifications. DeepMind reports that the work reached compiled implementation code, extending beyond protocol-design analysis, and took four weeks in this particular research effort. As a countercheck, researchers restored two previously fixed bugs in an isolated version and obtained machine-checked proofs of insecurity. The team is releasing formal-verification artifacts and a technical paper for community inspection. This is a distinct research result rather than a general claim that AI can certify any application. A valid proof still depends on the chosen properties, modeling assumptions and trusted components. It also does not establish the security of every interface, device or operational practice around a messaging system.

DeepMind announced Gemini Robotics 2.5 on November 10 and is rolling it out to robotics partners. The vision-language-action model was evaluated across more than 200 tasks, including tool use, delicate-object handling and manipulation of flexible objects. A proof-of-concept paired Apptronik’s Apollo 2 humanoid with Mercedes-Benz production-line kitting, using an embodied reasoning model to track work and Robotics 2.5 to control hands and arms. DeepMind reports a 96% success rate for that pilot; the number describes that setting rather than a universal reliability guarantee. The company also describes inference-time guidance around physical hazards and introduces the ASIMOV-Jailbreak benchmark for adversarial robotics requests. Gemini 4 Argon supported the research through simulation and analysis of failure footage. This launch concerns a partner robotics stack, not a consumer download that turns arbitrary hardware into a safe autonomous worker. Real deployments still require equipment integration, workspace controls and task-specific validation.

Anthropic introduced Claude Whiteboards and Claude Sheets in beta on November 10, alongside the general availability of Claude Data, formerly Claude Dashboards. The products are offered on paid plans across web, desktop and mobile. Whiteboards supports shared brainstorming and turning discussion into timelines, owners and action items. Sheets builds and edits spreadsheets with tabs, formulas, formatting, dropdowns and tracked changes, with export to XLSX or CSV when fuller spreadsheet functionality is required. Claude Data connects to sources such as warehouses or CRMs. Certified queries and live sharing remain beta features: a shared viewer sees approved query results without gaining arbitrary use of the connector owner’s access. Data also adds caching, refresh settings and version history. Team plans enable the three products by default; Enterprise Whiteboards and Sheets start disabled, while Data is scheduled to turn on by default on November 19 unless administrators change settings. These controls warrant review before sensitive collaboration begins.

MiniMax’s November 11 announcement introduces H3-Image, an open-weight model for text-to-image generation, reference-based generation and editing. The company says the model can work with up to 16 reference images and produce 4096-by-4096 outputs in under ten seconds. Those are provider claims for its reported setup, not measured guarantees for arbitrary hardware or hosted request queues. MiniMax attributes the result to changes across training, sampling and deployment rather than a single optimization. The weights are linked from the announcement, and hosted access is offered through the MiniMax API Platform, MiniMax Code and Hailuo AI. Open weights should not be equated with unrestricted reuse: developers still need to inspect the actual model license, deployment requirements and applicable terms. The launch also advertises relative pricing and a temporary promotion, but a production budget should use the current explicit price sheet for the selected resolution and mode.

NVIDIA’s November 10 technical article describes Alpamayo 2 Edge, a vision-language-action model designed for DRIVE AGX Thor. A two-billion-parameter Nemotron-based reasoning backbone is paired with a diffusion action expert. Surround-camera observations, temporal context and navigation inputs are used to produce driving trajectories and Chain-of-Causation reasoning when needed. The article explains post-training in AlpaGym and deployment recipes using TensorRT and DFlash2. NVIDIA reports 160–190 milliseconds of pixel-to-trajectory latency on the stated Thor configuration; this does not establish latency on other devices or road safety certification. The accompanying AlpaJudge workflows assess reasoning against a reference or visual evidence, support reviewing training examples and provide a possible reasoning-reward component. NVIDIA specifically advises examining low-scoring or safety-critical clips rather than automatically discarding them. The practical release is a research and integration foundation for qualified vehicle teams, requiring aligned clips, timestamps, trajectory evidence and held-out evaluation.

Cursor announced a multiplayer upgrade to Projects on November 10, rolling out in beta to all Cursor users. Teams can invite colleagues into a shared project, see tasks and associated pull requests, and maintain common requirements, diagrams and technical documents. Each teammate keeps a one-to-one conversation and local context with their own coordinator and worker agents, while a shared layer makes project decisions available across the team. Cursor presents on-call handoff and overlapping feature work as examples: agents can use shared context to understand unfinished work and coordinate ordering. The announcement’s examples describe intended collaboration benefits, not proof that merge conflicts or misunderstandings disappear. A team still needs ownership rules, review expectations and appropriate permissions for connected resources. The main change is that agent coordination moves beyond an individual’s session to a shared working environment, where outdated assumptions and incorrect shared decisions can also propagate more broadly.

ElevenLabs released Scribe v3 for recordings and Scribe v3 Live for realtime audio on November 10. Both are available through ElevenAPI, and the live model is also offered in ElevenAgents. The provider reports lower error rates than the preceding models across its evaluations; recorded word error and live semantic word error use different measures and should not be compared as one accuracy score. Live transcription gains diarization and turn detection that considers conversational context alongside audio, aiming to reduce interruptions during pauses. The recorded model supports keyterm prompting and output-format controls, while the live model adds multilingual switching and a larger keyterm-library option. US processing is the default. Isolated regional environments and Zero Retention Mode are described as deployment options, not a blanket assurance that every request uses those settings. Changing the model ID may be straightforward, but downstream summaries and voice-agent actions still need evaluation on real accents, numbers and product vocabulary.

Alibaba Cloud’s November 11 documentation introduces qwen4-max-preview through Token Plan only; pay-as-you-go access is not supported for this preview. The stated window runs from November 11 at 10:00 to November 25 at 24:00, UTC+8, after which the model will be retired. It supports text generation, image and video understanding, tools and structured outputs, with OpenAI Chat Completions, Responses and Anthropic Messages compatibility. Thinking is enabled by default. An important integration detail is that max_tokens and max_completion_tokens both limit the combined reasoning and final-answer budget for this model, so a setting previously sized only for an answer may truncate output. The documentation also specifies handling for preserved thinking in multi-turn Chat Completions. The preview is suitable for controlled comparisons, but its planned retirement means an integration needs a fallback rather than treating the identifier as a permanent production dependency.