GenAI Daily - November 7, 2026: Thinking Machines' Inkling 2 Small, DeepMind's RSA Findings, Perplexity Teammates
Top Stories
Thinking Machines Ships Inkling 2 Small, 594B MoE With Apache 2.0 Weights Promised Next Week
Thinking Machines Lab released Inkling 2 Small on November 5, a 594B-parameter Mixture-of-Experts model with 17B active parameters.
It scores 31 on the Artificial Analysis Intelligence Index. If the weights ship as planned, it would be the highest-scoring open-weights model developed outside China.
It accepts image inputs and supports variable thinking effort.
Today it is closed.
Tinker lists $0.30 per million prefill tokens, $0.03 cached, $1.00 sampled and $1.50 for training at the 64K tier.
Serverless inference is labelled beta, and production use is not self-serve. Teams must email Tinker to join a waitlist with their volume and latency requirements.
Artificial Analysis lists the Apache 2.0 release date as unconfirmed.
The company says it will release the weights after external safety testing and red-teaming.
On the lab's own card, DeepSeek V4.1 Flash leads on most coding rows. Evaluate it for knowledge-work and agent tasks rather than as a coding upgrade.
Artificial Analysis | VentureBeat
Why it matters: If the Apache 2.0 weights land, teams get a self-hostable US-origin model near the open-weights frontier. Memory planning for a 594B checkpoint should start now.
DeepMind Says AI Agents Found Two Weaknesses in Legacy 1024-Bit RSA
Google DeepMind says its AI agents, working under human supervision, found two weaknesses in legacy 1024-bit RSA.
One is a padding result for signature forgery, with a theoretical 1,600x gain in one setting and a 3.5x CPU-time saving reproduced on an already-solved key. The other is a batch-factoring method about 30x cheaper per additional key.
The work was done in a matter of weeks.
The scope is narrow.
The essay says modern protocols using 2048-bit or larger keys remain secure.
The 1,600x figure is a theoretical estimate. The practical task is an inventory.
Legacy 1024-bit keys still exist in embedded devices, old certificate chains, and long-lived infrastructure nobody remembers provisioning.
Why it matters: Security teams should audit for 1024-bit RSA in signing and certificate paths now, because AI-assisted analysis lowers the cost of probing the old keys that remain.
Perplexity Launches Teammates, Always-On Work Agents, Into a Crowded Field
Teammates lets Max subscribers create persistent, named AI agents for recurring work.
They inherit existing Projects, skills, memory and tool permissions from Perplexity Computer. They collaborate with other teammates and message human coworkers directly in Slack.
Perplexity says they suit deep research, data analysis, QA and ticket triage.
It puts Perplexity in the same race as OpenAI's Dots and Anthropic's Claude Duet.
Teammates is available now on the web for subscribers of the $200-a-month Max plan.
Perplexity cites internal usage: about 10,000 teammates created in the last 30 days, with about 1,000 active in one day. Those are vendor-reported figures.
Open questions include whether agents retain accurate context, recover from failed tools, and avoid compounding earlier mistakes.
Another is how downstream agents respond when another teammate supplies incomplete or incorrect work.
AlphaSignal | Free Press Journal
Why it matters: Persistent, handoff-capable agents are becoming a standard product tier across vendors, and buyers should test long-running reliability and handoff failures rather than demo-level tasks.
Key Developments
Anthropic Adds Security Center and GA Inference Hooks to Claude Enterprise
Anthropic's November 5 update adds an in-product Security Center. It lines up a customer's Claude settings against Anthropic's recommended configuration and flags the gaps.
The page covers chat, Cowork, and Claude Code together.
Prompt and tool-call checks are generally available, while the settings center and company-wide console are in beta.
A separate console at admin.claude.com handles parent-organization management across Claude Enterprise and Claude Platform orgs. It covers single sign-on, SCIM directory sync, domain verification, and Compliance API keys.
Security teams can also sign a member out of Claude with a single Admin API call.
Claude Desktop can be locked down on managed computers to block personal and other outside Claude accounts.
Anthropic says the beta security center does not certify compliance. One limit to plan around: sensors run in detection mode and report activity to security teams without blocking it.
Impact: Teams can point existing DLP and EDR tooling at Claude traffic, which removes one common blocker for moving pilots into production. Block-mode enforcement is still absent.
Five9 Buys Level AI for $325M as Contact Center AI Consolidation Continues
Five9 announced the $325 million all-cash deal alongside its third-quarter results.
Level AI has about $20 million in standalone ARR, growing about 50% year over year. The deal closes in Q4.
Level AI serves Gusto, Chime, Empyrean and VistaPrint. Its platform runs on its own Latitude family of models, which the company says deliver lower costs and latencies than frontier models.
Five9 says Level AI gives it a way to reach enterprises that want AI today but aren't ready to move their contact center to the cloud.
The market did not reward it.
Five9 shares fell about 21% Friday after it cut its 2026 adjusted EPS guidance.
AI revenue grew 91% year over year to about $45 million, 17% of subscription revenue. The deal follows another consolidation move.
On November 4, Dialpad announced it is combining with Ada.
The all-share deal reportedly values Ada at about C$600 million, against $1.2 billion USD in 2021.
Impact: Buyers of standalone CX AI tools should check vendor roadmaps and contract terms, since point-solution startups are being absorbed into platform vendors.
HubSpot's Q3 Shows Agent Adoption Growing, but Buyers Want Proof
HubSpot reported Q3 revenue of $947.6 million, up 17%.
CEO Yamini Rangan said 20% of Pro+ customers now use HubSpot's agents, up 11 points since the start of the year, and 61% use Breeze Assistant.
Customer Agent resolved over 2 million support issues in Q3.
The outlook was softer.
HubSpot expects fourth-quarter growth of 12%, and CFO Kate Bueker said net new ARR would decline year-over-year in fiscal 2026.
Net revenue retention slipped to 101%, which she attributed to a step-up in budget sensitivity. Agent usage is rising, but that is not yet translating into faster commitments.
Impact: Expect vendors to push usage-based or outcome-based pricing, and bring your own resolution and cost-per-ticket data to negotiations.
Product Launches
Zendesk Employee Service AI Agents
Zendesk's employee service AI agents are built to support IT, HR, Legal, Finance, and other internal service teams.
Admins can connect them to workflows built in advanced action flows, so agents can execute multi-system tasks end to end.
From November 3, Growth and Professional plans include 10 automated resolutions per agent per month, and Enterprise tiers include 15.
The launch puts Zendesk in a more direct fight with ServiceNow, Microsoft, and Atlassian for employee service budgets.
Why it matters: Employee service is becoming a contested market for agent vendors, and the included resolution allowances give buyers a concrete basis for comparing costs across platforms.
Microsoft Copilot Studio October Updates
A new MCP server lets developers access Copilot Studio through tools like GitHub Copilot, Claude Code, and OpenAI's Codex.
Connecting agents to other agents using the agent-to-agent (A2A) protocol is generally available.
Agent accounts let users delegate tasks to AI agents through Teams, Outlook, and Office files.
An agent usage estimator forecasts Copilot credit consumption before scaling.
Magnetism Solutions | Microsoft Learn
Why it matters: Generally available agent-to-agent connections and a credit usage estimator help teams build multi-agent workflows and forecast costs before scaling them.
Claude Personal Finance (Beta)
The feature is available on Claude's Pro, Max and Ultra plans in the US. Users can link bank and card accounts via Plaid and ask about balances, spending and bills.
Access is read-only, and conversations in the hub are not used to train models. It is a consumer feature. For enterprise teams, the read-only, no-training design is a useful reference pattern for financial data connectors.
Why it matters: It shows how a major AI vendor is approaching sensitive financial data, with read-only access and no training on user conversations.
Funding & Deals
Sand AI Raises $200M Series A+
Sand AI is a China-based video-generation model startup.
It will use the funds for model research, computing infrastructure, product upgrades, and overseas expansion.
The funding brings total capital raised over the past 10 months to $300 million. Led by Cornerstone Capital and Wangsu Science & Technology.
Why it matters: The $300 million raised in 10 months signals strong investor appetite for video-generation model developers, including those planning overseas expansion.
Ocient Raises $100M Series C
Chicago-based Ocient provides an agentic AI solutions platform designed for dynamic, petabyte-scale data.
It serves enterprise, service provider, defense, and national security clients in high-volume, data-intensive industries.
The raise brought total funding to $260M. Co-led by BOKA Capital and OCA Ventures.
Why it matters: Continued funding for petabyte-scale data platforms shows demand for infrastructure that supports agentic AI in data-intensive and regulated industries.
Five9 to Acquire Level AI for $325M
The Mountain View contact center AI startup is being bought in cash.
Management expects about one point of adjusted EBITDA margin dilution in 2027, with Level AI reaching EBITDA breakeven about six months after closing. Acquirer is Five9; details are in Key Developments above.
Why it matters: The margin dilution and breakeven timeline show the near-term financial cost Five9 is accepting to add Level AI's technology.
Tomorrow's Watch List
- Inkling 2 Small Apache 2.0 weights: Thinking Machines says next week, but no date is confirmed. Watch for a Hugging Face release.
- Conductor's cloud service shuts down November 20 after Meta's acquisition, so existing users need to migrate. Conductor is the coding-agent startup Meta bought this week.
- Zendesk's employee service AI agent rollout continues through December 1.
Related reading: Check out this week's [Deep Insights analysis] for strategic context on these developments.
GenAI Daily is curated by rAInvent. For in-depth analysis, see the rAInvent articles.